clairvoyance

Each entry: what the direction looks at, the questions that open it, the tell that selects it, what it characteristically yields, and the specific way it fails. The failure line is the important one — every direction here has a degenerate form that feels exactly like the real thing from the inside.

Read the direction the tells selected. Reading all eight is the campaign-sized sweep and it is rarely the right size. Seven of these look at the problem; the last looks at the field the problem sits in — the people and forces that will decide what happens to any answer — and it is the one a view from above is mostly made of.

---

Behind — the goal the request is already a solution to

Looks at: the problem the requester solved in their head before writing to you. A request is almost never a problem statement. It is a proposed remedy, with the problem compiled out of it.

Questions:

• What would be true if this were done? And what would that make possible?
• What happens if it is not done — what is the actual cost, to whom, when?
• Why now? What changed? A request that could have been made a year ago and was not is a request with a recent cause that is not in the sentence.
• Is there a cheaper thing that produces the same outcome rather than the same mechanism?

Tell: the request names a mechanism rather than an outcome. "Add caching", "add a retry", "add a flag", "make it async" — each names a remedy, and each is one member of a family the requester picked from without saying so.

Yields: the highest-value findings in the whole set, and the ones that save the most work, because a request satisfied at the goal level is often an order of magnitude smaller than the mechanism it arrived as.

Fails as: condescension, every time, and it is the failure that gets this whole skill switched off. A recovered goal is a hypothesis about someone else's intent. You have noticed their words admit two readings — a fact about the words, not privileged access to their mind. Ask "is this about X or about Y?". Never say "what you actually want is". And when they are available to be asked, asking beats inferring, always; inference here is what you do when presence is not paired.

---

Under — the assumptions the vocabulary smuggles in

Looks at: the commitments you accepted by answering in their words. Every noun asserts a thing exists and is one thing. Every verb asserts a remedy.

Questions:

• Which nouns here are assumed to be singular? ("the uploader", "the user", "the config")
• Which are assumed to exist at all?
• What is the request treating as given that is actually a choice someone made?
• What would have to be true for this framing to be correct? Is it?

Tell: the third fix did not hold. Or the bug moved rather than dying. Or a parameter appears to have no effect. All three are the signature of a correct fix applied inside a wrong model — diagnosing-before-fixing owns the loop, and this is what its "escalate to questioning the architecture" step actually does when you get there.

Yields: the finding that explains a whole run of failures at once. Under is the direction that pays in Diagnose specifically.

Fails as: infinite regress. Every assumption rests on another one, so the questioning never terminates on its own. Bound it: list the assumptions, mark each load-bearing or not by asking whether the work changes if it is false, and examine only the load-bearing ones. Three is usually all there are.

---

Above — the class this is one instance of

Looks at: the general case. This request has a shape, and the shape has probably been through here before.

Questions:

• What is this an instance of?
• Has something of this shape come up before — in this repo, this month, this conversation?
• Is the general case cheaper than the specific one? Sometimes it is: one mechanism replacing four special cases is less code, not more.
• And the inverse, which matters more often: is the specific case genuinely all that is needed, and is the general one a way of avoiding a decision?

Tell: this is the second or third time something of this shape has come up. Repetition is the only honest evidence that a general case exists; automating-repetition owns the judgment about when a repeated shape becomes machinery, and this is the sight that feeds it.

Yields: either a much smaller total, or a clear decision to stay specific that will not be second-guessed later.

Fails as: premature abstraction, which is the most expensive failure in this file. One instance is not a class. Two is a coincidence. The generalization built on one instance is always the wrong generalization, because the features you cannot see are exactly the ones belonging to the single case you have. judging-duplication is the counterweight and it wins by default.

---

Beside — what stands next to this once it exists

Looks at: the neighbourhood. Not consequences over time (that is ahead) but adjacency now — what this touches, enables, blocks, or duplicates.

Questions:

• What becomes trivial once this exists, that is currently hard?
• What becomes impossible, or much harder?
• What already does most of this? A capability that exists under another name is the single most common thing a sweep finds.
• Who else is writing to this, and what do they not know is coming?

Tell: the change touches a boundary something else depends on. Or the tree has another writer in it, in which case beside is not optional — see the envelope note in the main file.

Yields: the duplicate-capability find, and the collision nobody had connected to this work.

Fails as: the adjacent-possible trap. Beside reliably produces genuinely good ideas that are genuinely not this task, and they are seductive precisely because they are cheap right now while you have the context loaded. That cheapness is real and it is not a route. Third row of the routing table: into the record, offered at close, built if they say yes.

---

Ahead — the second-order consequence

Looks at: the consequence of the consequence, and the person who inherits this after everyone who built it has moved on.

Questions:

• Who maintains this in six months, and what do they need to know that is currently only in this conversation?
• What does this make hard to change later? Every mechanism is also a commitment; which one is being made here?
• If this succeeds completely, what breaks? Success is a load profile too.
• What is the reversal cost — and is anyone treating this as reversible when it is not? deciding-reversibility owns the answer; this is the question.

Tell: the work is easy now and permanent afterwards. Schemas, public interfaces, file formats, anything with a migration on the other side of it, and anything at all when the blast radius is live.

Yields: the one-way door nobody had labelled as one.

Fails as: speculative futureproofing — building for a load, a scale, or a requirement nobody has. The discipline is that ahead produces an observation and a reversal cost, not a feature. "This is a one-way door and here is what it costs to walk back" is the finding. "So we should build the general version now" is above failing, wearing this direction's clothes.

---

Against — the inversion

Looks at: the negation of the request. The direction nobody takes, which is why it is the highest-variance one here.

Questions:

• What if the answer is to delete something rather than add?
• What if the answer is to do nothing — what actually happens then?
• What if this were done by hand, once, to find out whether it recurs?
• What if the failure were made impossible rather than handled? An unrepresentable bad state needs no check.
• What if the constraint everyone is designing around were simply removed?

Tell: the design is accumulating special cases. Every special case is the frame charging interest, and a fourth one is the signal to invert.

Yields: the largest simplifications available anywhere. deleting-code and modeling-state-machines are where the results usually land.

Fails as: contrarianism — inverting to be interesting rather than to be right. The check is cheap: an inversion has to be argued for on the same terms as the original, and if the argument is "it is unexplored", it has not been made. Novelty is a reason to look, never a reason to choose.

---

Outside — what another discipline calls this

Looks at: the same problem, in a field that has already solved it and named it something else.

Questions:

• What would someone who has never seen this codebase, this team, or this conversation ask first?
• Which discipline has this problem as a solved standard — distributed systems, control theory, accounting, logistics, typesetting, aviation, law?
• What is this problem called there, and what does that field consider the obvious answer?
• What does the stranger's question reveal that the expert's does not? Expertise is frame-acceptance made efficient, which is its value and its cost.

Tell: the problem feels novel. It almost never is, and that feeling is mostly a report about your own vocabulary rather than about the problem.

Yields: an existing solved answer, an existing name to search under, and occasionally a whole missing concept. This direction is routinely dismissed as decorative and is not — on this skill's own creation it was the direction that paid best, and self-application.md records exactly what it produced and what it cost.

Fails as: analogy mistaken for argument. An import from another field is a hypothesis with a good pedigree, and pedigree is not evidence. It gets tested against this problem's actual constraints like anything else, and the constraints that made it work there are the first thing to check for here.

---

Around — the field of people and forces

Looks at: everything that is not the problem and will still decide what happens to it. A request lands in a field: people who have a stake in the answer, incentives that pull the design one way whatever its merits, a deadline or a budget or a rule that was never written into the sentence, and a surrounding world whose motion is the reason this arrived this week. The seven directions above are about the thing being asked for. This one is about the room it was asked in.

Questions:

• Who else has a stake in this — who pays for it, who is measured on it, who has to live with it, who could veto it, and who does not know it is coming?
• What incentive does each of them answer to, and where does that incentive point that the request does not?
• What is fixed here that no engineering choice can move — a date, a contract, a regulation, a headcount — and is the design already respecting a constraint of that kind without naming it?
• What is happening around this that made it arrive now? A request is a response to something, and the something is rarely in the request.
• What organization builds and runs this, and will the boundary in the design match a boundary in that organization? A boundary nobody owns has no owner.

Tell: the work sits in a system with more than one owner, or the request carries a date, an audience, or a budget it did not mention. Also: a design that is technically sound and keeps being argued with — the argument is coming from the field, not from the design.

Yields: the stakeholder or force that will actually settle the design, found before it settles it; the incentive that explains a requirement nobody could justify on the merits; the deadline that makes the elegant option the wrong one; and the seat that was empty when the request was written, which is the finding seat exists to produce on demand.

Fails as: speculation about people. Motive is the one thing in this list you cannot observe, and an "around" finding that names what someone wants rather than what they are measured on or have said is a story. Stay with what is observable — the metric, the deadline, the org chart, the message — and phrase the rest as the question it is. The second failure is politics as design input: the field is a constraint to design inside, never an argument for a worse design. threat-modeling owns the one class of actor that must be enumerated adversarially, and it stays there.

A direction says where to look from. A lens is a specific instrument to look with — a question some other field turned into a standard technique because its practitioners kept being blind in the same place. Each one here is filed under the blindness it cures, because that is how you pick one: name what you are probably not seeing, then take the lens built for it.

The catalogue is not a checklist. Running every lens is furniture, and the tell for each is what selects it. Most sweeps use one; an orbit uses one per pane; lens <name> runs a single one by hand when you already know which.

Each entry: what it asks, the tell that selects it, what it characteristically yields, how it fails, and — where this library already has the technique in full — the skill that owns it. When an owner is named, the lens is the reason to open that skill, not a substitute for it.

---

Vocabulary lock-in — you inherited the words

Ideal final result (TRIZ). Asks: what if the outcome simply occurred, with no mechanism added at all — nothing built, nothing run, nothing maintained? Then: what is the least that has to exist for that? Tell: the request names a mechanism. Yields: the smallest design available, and often a deletion. Fails as: a wish written down as a plan; the lens ends at the least mechanism, never at zero.

The job it is hired for (product research). Asks: who "hires" this, in what situation, to make what progress — and what did they fire to hire it? Tell: the request describes a feature and never a moment in someone's day. Yields: the outcome-level statement of the request, which is what an acceptance line should be made of. Fails as: inventing the person; ask them when presence=paired. Owner: scoping-before-building aims its questions at purpose for this reason.

Commander's intent (military planning). Asks: if the situation changed after the order was given, what would the person who gave it still want? State the request one level of purpose up, so it survives a changed situation. Tell: a run that will outlast the assumptions in the request. Yields: the stop condition an unattended run should carry. Owner: bounding-autonomous-work for the stop conditions; finishing-what-you-started for the ledger line.

Anchoring — the first thing seen became the model

Key assumptions check (intelligence tradecraft). Asks: list every assumption the current reading rests on; for each, would the work change if it were false, and how would you know? Examine only the load-bearing ones. Tell: a third fix did not hold, or a conclusion has hardened without new evidence. Yields: the one assumption that explains a run of failures. Fails as: infinite regress — three load-bearing assumptions is usually all there are. Owner: the under direction is this lens applied to a request; diagnosing-before-fixing applies it to a reproduction.

Quality of information check (intelligence tradecraft). Asks: for each piece of evidence — where did it come from, how was it obtained, could the source be wrong or partial, and is anything else corroborating it? Tell: a conclusion resting on one report, one search, one tool result. Yields: the finding that the decisive evidence was never corroborated. Owner: calibrating-confidence marks the tiers; this is the pass that asks.

The stranger's first question. Asks: what would someone with none of this context — not the codebase, not the thread, not the field — ask first? Tell: expertise everywhere in the room and no dissent. Yields: the question the experts stopped asking years ago. Owner: red-teaming-your-own-work calls it distance obtained cheaply and has the mechanics.

Scale — the size was assumed

Ten times, one tenth. Asks: what changes if the load, the count, the budget, the team, or the deadline were ten times larger — or ten times smaller? Tell: a design with numbers in it nobody chose. Yields: the mechanism that only works at the assumed size, in either direction — the elaborate thing built for a scale of one is the commoner find. Fails as: futureproofing; this produces an observation and a threshold, never a feature.

Population one. Asks: what does this mechanism do when the population it needs — reviewers, corroborators, tenants, users — is one, or zero? Tell: any design built on independent agreement. Yields: the safeguard that is inert until the day it matters. Owner: scoping-before-building, sizing a trust circle.

Altitude ladder. Asks: say the problem at five altitudes — the line, the module, the system, the product, the world it operates in — one sentence each. At which altitude is it cheapest to solve? Tell: a fix that keeps growing, or a request stated at a level nothing else in the design lives at. Yields: the level the problem should be solved at, which is rarely the level it was stated at. Owner: deciding-reversibility decides at the right altitude; explaining-technical-work chooses the altitude of a report. This is also the sixth move in the main file, because it works on anything.

Class — the instance was taken as unique

The reference class by shape. Asks: what is this an instance of — named by its shape, never by its domain — and what did the last several members of that class cost, take, or turn into? Tell: the second or third time this shape has come up, or an estimate about to be built from decomposition alone. Yields: the number the inside view cannot produce, and the general case that four special cases were hiding. Fails as: premature abstraction — one instance is not a class; judging-duplication wins by default. Owner: estimating-effort for the number; automating-repetition for when the repetition becomes machinery. This is the above direction's instrument.

Time — the present was assumed permanent

Why now. Asks: what changed to make this arrive this week? A request that could have been made a year ago and was not has a recent cause outside the sentence. Tell: any request. Yields: the trigger, which is often the real requirement. This is behind's third question and earns its own entry because it is the cheapest lens here.

Pace layers (Stewart Brand). Asks: which layer does this change live in — the ones that churn (interface, feature) or the ones that hold (schema, data model, contract, organization)? Is it being built at the pace of a different layer? Tell: a fast change to something slow, or a permanent change made for a passing reason. Yields: the interface frozen for a UI whim, the schema migrated for a quarter's experiment. Owner: deciding-reversibility for the door it turns out to be.

Pre-mortem (Gary Klein). Asks: it is a year later and this has failed badly. Write the first sentence of the postmortem. Then: what would have to be true today for that sentence to get written? Tell: a plan with no named risks, or a room where nobody is objecting. Yields: the failure nobody would raise as a criticism but everyone will state as a prediction. Fails as: a list of generic risks; the sentence has to be specific enough to be checked. Owner: red-teaming-your-own-work attacks the finished thing; writing-postmortems is where the real one goes if the lens is ignored.

Indicators (intelligence tradecraft). Asks: if the current reading is wrong, what would you see first — and can it be watched? Tell: a decision about to be relied on for a long time. Yields: a tripwire instead of a belief. Owner: bounding-autonomous-work for stop conditions; revalidating-decisions for the reopen condition on a recorded decision.

The history in the tree. Asks: what did this look like a year ago, what was tried and reverted, and what has this project already refused? Tell: a proposal that feels obvious and has no trace of ever being made. Yields: the closed pull request that already answered the question. Owner: orienting-in-unfamiliar-code (the decisions that are not in the code) and code-archaeology (the reason a line exists).

Actors — one seat was assumed

The empty chairs. Asks: which seat was nobody sitting in when this was written — the operator at 3 a.m., the new hire, the auditor, the person who inherits it, the integrator, the second tenant, the customer's customer, the translator, the user on the slowest machine? What does each notice first? Tell: a request written from one seat, which is every request. Yields: the reader or user the design silently excludes. seat <chair> runs one of these on demand; the roster is at the end of this file. Owner: explaining-technical-work for the reader's seat; formidable for the user's.

Attacker tiers. Asks: who could want this to fail, with what capability? Tell: a new trust boundary, a new class of data, a new integration. Owner: threat-modeling — entirely. This entry exists so the actor pass does not forget the adversarial chairs, and it hands off the moment it finds one.

Show me the incentives. Asks: who benefits, who pays, who is measured on what, and where does each measurement point that the request does not? Tell: a requirement nobody can justify on the merits, or a sound design that keeps being argued with. Yields: the force that will shape the outcome regardless of the design. Fails as: cynicism, and speculation about motive — stay with what is observable: the metric, the deadline, the message. Motive is a question you ask, never a finding you write.

The mirror of the organization (Conway). Asks: which team builds and runs each part of this, and does every boundary in the design match a boundary in who maintains it? Tell: a component two teams both half-own, or one nobody does. Yields: the boundary that will drift because nobody owns the property across it. Owner: drawing-boundaries.

System — a part was mistaken for the whole

The purpose of a system is what it does (Stafford Beer). Asks: describe what this system actually does — to its users, its operators, its data — ignoring every statement of what it is for. Where do the two diverge? Tell: a charter, a README, or a mission that nobody has compared to behaviour lately. Yields: the gap between what a project says and what it is, which is the first pane of an orbit.

Leverage points (Donella Meadows). Asks: at what level is this request intervening — a parameter, a buffer, a structure, a delay, a feedback loop, an information flow, a rule, a goal, the paradigm the goal comes from? Is there a level up where the same outcome costs less? Tell: the same class of fix has been applied repeatedly at the parameter level. Yields: the information flow or rule whose change makes the parameter-tuning unnecessary. Fails as: the grand redesign; anything found above "rule" routes to the record, never to the gate.

Flows across the boundary. Asks: draw the boundary of the thing under discussion; what crosses it — money, data, trust, attention, dependencies, obligations — in each direction, and which of those has no owner on this side? Tell: a property that is sensible inside every component and wrong in composition. Owner: drawing-boundaries (the property no single component can see) and mapping-dependencies (the three graphs).

Feedback and delay. Asks: what loop does this sit in, and how long between an action here and its visible effect? What will someone do in the meantime? Tell: a control that is adjusted before its last adjustment has taken effect. Yields: the oscillation nobody modelled. Owner: rate-limiting-and-backpressure for the mechanics.

Negative space — what is absent was invisible

The seen and the unseen (Bastiat). Asks: for every effect you can point at, name the one you cannot — the thing not built because this was, the user who left without a ticket, the request never made because it was known to be refused. Tell: a benefit stated with no cost beside it. Yields: the cost that was paid silently. This is the absence section in the main file, under the name the discipline gave it.

Deviation guidewords (HAZOP, process safety). Asks: for each parameter of the thing — every input, output, timing, quantity, actor — apply each guideword and describe the case it produces: no, more, less, as well as, part of, reverse, other than, early, late, before, after. Tell: a design reviewed only against the cases its author imagined. Yields: cases, and ideas, that no unguided brainstorm produces, because the guideword does the generating and the parameter does the grounding. This is the seventh move in the main file. Owner: auditing-new-input-categories and property-based-testing turn the cases into checks.

The unasked question. Asks: what do people in the requester's position usually ask that this request does not? Tell: a request unusually short for its consequences. Yields: the requirement that was compiled out because it was obvious to them.

Ask the fluent expert. Asks: what would a native of this category — this script, this currency, this protocol, this regulation — flag first as the thing outsiders always get wrong? Tell: extending a system to a category it has never handled. Owner: auditing-new-input-categories.

Measurement — the instrument was trusted

Goodhart's law. Asks: if this number becomes the target, what does it stop measuring, and what will people do to move it? Tell: a metric about to become an acceptance criterion or an incentive. Yields: the second metric needed to keep the first honest.

The ruler measures the hand (Wittgenstein's ruler). Asks: a measurement is as much a statement about the instrument as about the object — what does this reading say about the thing that produced it? Tell: a number that is implausibly good, byte-identical across a change, or zero. Yields: the instrument that never ran, the proxy wearing the name of the thing it stands in for. Owner: diagnosing-before-fixing and calibrating-confidence each carry a case of this.

The base rate and the control. Asks: how often does this shape occur anyway? What happens if nothing is done? What would the measurement read with the change absent? Tell: an effect claimed with no comparison. Yields: the improvement that was regression to the mean. Owner: red-teaming-your-own-work bounds a null by what the search could have seen.

Evolution — build and buy were not distinguished

Where it sits on the evolution axis (Wardley). Asks: is each component of this genesis (nobody has done it), custom (some have), product (you can buy it), or commodity (it is a utility)? Build what is genesis; buy or borrow what is commodity. Tell: something being hand-built that has a price list. Yields: the component that should be a dependency, and the one that should not. Owner: steelmanning-alternatives has buy-or-build as one of its generators.

Inversion — the negation was never tried

Owned in full by the against direction. The named forms, so they can be run by hand: via negativa (what would removing something fix?), do nothing (what actually happens?), the worst possible idea (state the worst answer, then invert each of its properties), relax the constraint (revalidating-decisions owns the case where the constraint is a recorded decision), and by hand, once (do it manually to learn whether it recurs).

---

Seats — the chairs seat <chair> can sit in

Each seat is one look, from one position, at what that person notices first. Pick the seat the request was not written from.

The main file's method was written for a request, and a request is one instance of the general case: anything that arrives carrying a frame it did not announce. A plan carries its author's judgments compiled out. A metric carries its instrument. A codebase carries "how it is done here". A project carries a charter nobody has compared to its behaviour lately. The run you are in carries the track and envelope you chose an hour ago, which have become premises.

The method is the same for every object — name the frame, restate without its nouns, look in the direction the tells select, name the observation that would tell two readings apart, route what you find. What changes per object is four things: where its frame comes from, the moment a look is cheap, the direction that pays first, and where a hit goes. That is this table, then the notes.

Notes on the objects that are not requests

A plan. Executing a plan exactly as written is correct discipline and also means adopting every judgment its author made without seeing one of them. The plan's frame is those judgments. The cheap moment is before task one, and the tell inside execution is a task that cannot be implemented as written — that is the plan's frame failing, not the implementer's, and working-a-plan-task-by-task routes it back to the plan for that reason. Sweep the plan, never the task.

A decision. A recorded decision has a principle and a premise, and only the premise decays. The sweep's question is ahead run backward: what has changed since this was written that its author could not have known? The finding is a premise that expired; revalidating-decisions owns what happens next, and the sweep must not overturn the decision itself — that is that skill's job, done with the record open.

A measurement. A number arrives with the authority of an instrument and the frame of whoever pointed it. Three questions, from the measurement lenses: what does this reading say about the instrument, what would it read with the change absent, and what does it stop measuring once it is a target. The finding is a tier mark — verified, inferred, assumed — on the number before it is acted on. diagnosing-before-fixing owns the case where the instrument never ran.

A codebase. "How it is done here" is the strongest frame in the list because it is usually right, which is what makes the cases where it is wrong expensive. The tell is not the pattern itself but its fourth special case. A hit is an observation about the pattern, routed to the record or to an ADR — never a refactor started from inside a sweep. steelmanning-alternatives owns what to do once the pattern is admitted to be a choice.

A project. The one object whose frame nobody wrote: it is the sum of what everyone assumes the project is for. The method for it is orbit.md, and its bound is stricter than any other object's, because the space around a whole project is unbounded and every pass can be made to produce something.

A conversation. Your previous turn is the fourth frame source in the main file, and a conversation is that source compounding. The tell is a conviction that has grown across several turns without new evidence entering — an approach adopted three responses ago that is now a premise. The sweep is one sentence: say what you are now assuming that you were not assuming at the start, and let the person correct it.

The run itself. A track and an envelope were chosen at the start of the run and have since become the water it swims in. orchestrating-work-end-to-end re-reads them at phase boundaries; this is the sweep that asks whether the process is the frame — a Diagnose run that should have been a Build, an express amplitude that has quietly become a campaign, a fix loop whose third iteration is the tell. A hit routes to the run record as an envelope change, which is a decision, and never to a silent downgrade.

A tool or instrument. Everything observed through a tool arrives shaped by what that tool can see. A search that returns nothing is a fact about the search. A truncated result is a fact about the pipe. The sweep's question is what this instrument structurally cannot show, and the hit is a tier mark or a second instrument — calibrating-confidence owns reading a hit list as the input to a check rather than as the check.

The library itself. A skill that fired and turned out to be silent about the case in front of you is a frame too — the library's. That observation does not route to any gate in this run; it routes to tending-tbaguette's queue, in two sentences, and the run continues.

orbit is the view of a whole project from outside it: one bounded pass that looks at the thing everyone is standing inside and says what it is, who it is for, what moves through it, what is changing under it, and what it has never had. It is the closest this skill comes to the "God's view" people ask for, so the first thing to say is what that phrase cannot mean. There is no seat that sees everything. The view from above is assembled from several partial seats, chosen on purpose, each of which sees one thing the others cannot — and a person reading an orbit should be able to tell which seat every sentence came from. An orbit that claims omniscience has stopped being sight and become a manifesto.

Bounds first, because this is where the skill is most dangerous

An orbit is the one pass in this skill with no natural stopping point at all: the space around a whole project is unbounded, and every pass through it can be made to produce something. So the bounds come before the method.

• About five hundred words, counted. Seven panes of one to three sentences each, plus the findings block. This bound used to say "one page", and two measured runs of it came back at three times that while honouring every numeric bound in this file exactly — a limit stated in a unit the writer never measures is a wish. Count the words before delivering. If a pane needs more, the finding is in the excess and it goes to the routing table, not into the pane.
• At most three findings, each with the observation that would tell whether it is real, each routed. Everything else is description, and description is allowed to be all an orbit produces — an orbit that finds nothing is a project that is what it says it is, and that is a result.
• At named moments only. Not a mode, not a background hum, not something run because the work inside the frame got boring.
• Never during Respond. An incident has a clock. responding-to-incidents owns the order, and a whole-project view while users are broken is the "let me understand this properly first" red flag at its largest.
• Never a redesign. Anything found above the level of a rule — a goal, a paradigm, "this project should be a different project" — routes to the record and the closing offer. managing-scope-drift still wins, and it wins hardest here.
• Presence governs. With someone present, an orbit's findings are three questions. With nobody present, they are three lines in the record and the run continues on the original frame, per bounding-autonomous-work.

The moments

Between moments, an orbit stays valid until one of its panes changes, and the tell that one has is usually a finding from an ordinary sweep that does not fit the last orbit's description of the project. Re-run the pane, not the orbit.

The seven panes

Each pane is one seat. Write which seat, then what it sees, in one to three sentences. A lens from lenses.md is named beside each pane because it is the instrument that pane is made with; use it or a better one, not all of them.

1 — Altitudes. The project in one sentence at each of five heights: the line (what a function here does), the module, the system, the product (what a person gets), the world (what it changes outside itself). The finding, when there is one, is the height at which the sentence could not be written — the project that has no product sentence, or no world sentence. Lens: altitude ladder.

2 — Purpose, as revealed by behaviour. What the project says it is for, in its own words, and what it actually does — to its users, its operators, its data, its dependents — with the charter ignored. Where the two diverge, say which one the code is defending. Lens: the purpose of a system is what it does.

3 — The chairs. Who touches this: who pays, who operates, who maintains, who integrates, who is affected without knowing it, who could veto it. Then the empty ones — the seat nobody was in when the last three decisions were made. Lens: the empty chairs; threat-modeling the moment an adversarial chair is occupied.

4 — Flows across the boundary. Draw the boundary — what counts as this project — then what crosses it in each direction: money, data, trust, attention, dependencies, obligations. The finding is a flow with no owner on this side. Lens: flows across the boundary; mapping-dependencies for the graphs.

5 — Time. What is changing under this project — a dependency, a platform, a team, a market — and at what pace. Which of its layers churn and which hold, and whether anything is being built at the pace of the wrong layer. What was true when its oldest decisions were made that is not true now. Lens: pace layers; revalidating-decisions for the premise that expired.

6 — Negative space. What this project has never had, what it has refused, what it lost and did not replace, and what nobody asks for because everyone knows the answer. The closed pull requests and the wontfix list are half of this pane; the other half is what the domain has that the tree does not. Lens: the seen and the unseen; the history in the tree.

7 — Forces. The incentives, deadlines, budgets, rules, and habits that will decide what happens to any proposal here regardless of its merits — and which of the constraints everyone designs around would lift if someone asked. Lens: show me the incentives; the constraint audit from the main file.

The output

An orbit record is a markdown block with the seven panes as headings, one to three sentences each, and a final section:

## Findings (at most three)
1. <observation> — would be confirmed or refuted by: <the check> — routed: gate | ledger | record | discard (<reason>)

Where it goes depends on the envelope. In a run with a record, it is a section of the record, per finishing-what-you-started's ledger file. In a project that keeps design notes, it goes where those live, dated, so the next orbit can be diffed against it — an orbit is most valuable the second time it is run, because the diff between two orbits is the project's actual direction, which no single one can show. Handed to a person, it is the page plus at most one question, through the harness's own question tool, per offering-the-next-move.

How it fails

This is a worked record rather than an illustration. The skill was drafted, then swept with its own method at campaign size — all seven directions, once — and the draft was changed by what came back. Every edit named below is in the shipped file. It is kept for three reasons: it is the only honest way to claim the method does anything, it demonstrates the routing table on findings that were not chosen to be convenient, and it shows what a real yield looks like so that a run producing seven findings knows to be suspicious of itself.

The frame the draft inherited: the request arrived as make a flagship skill about seeing out of the box. Three commitments came with it — that the deliverable is a document, that "flagship" means large, and that more sight is the goal. The first two survived the sweep. The third did not.

---

Behind — what is this request a solution to?

Found: every skill in the library points inward at executing the request correctly. Nothing in it can notice that the request was wrong. So the outcome behind make a skill about creativity is narrower and sharper than the words: give the library an organ for noticing.

That reading has a consequence the draft was getting wrong. If the goal is the library gains this capacity, then the scarce thing is not prose — it is reachability. A flagship skill nobody routes to is a well-written file, and the draft was spending its effort on the document while the thing that decides whether it ever runs is the description, the spine, and the inbound cross-references.

Routed: to the gate. Remaining effort was re-weighted toward the orchestrator seat in all seven tracks and the inbound edges from the six skills whose failure modes hand off here, rather than toward more sections.

Also found: the draft was over-defended. It stated the null rule three times and had more lines on restraint than on method — which is what a skill written by someone worried about being annoying looks like. One restatement cut.

---

Under — what does my own vocabulary assume?

Found, and this was the sweep's best result: the draft said the frame, singular, throughout — after listing four separate sources it came from. That is precisely the failure this direction names, committed in the text that names it. A request does not have one frame. It has several, stacked, from different places, and they can disagree with each other.

Which turned out to be worth more than the correction. A disagreement between two frames is the cheapest sight available anywhere, because the noticing has already happened without you: you did not step outside the frame, two frames stepped outside each other. The reflex on meeting one — pick the more authoritative source, move on — discards the most valuable signal in the run.

Routed: to the gate. New subsection, Two frames disagreeing is the cheapest sight there is. It is now the most useful paragraph in the file and nothing in the draft would have produced it.

---

Above — what class is this an instance of?

Found: the class is an agent inherits a frame and has no organ for noticing, and it is wider than the request. Frames also arrive from a subagent's returned conclusion, a tool's output, a search that came back empty, and a memory written by a session nobody remembers. Each of those arrives as a finding with its framing compiled out — the same shape as a request, and trusted more, because it looks like evidence.

Routed: to the gate. A fifth source added to Where the frame comes from.

Discarded, with the reason: the general case would be a skill called something like examining inherited frames, which is more accurate and worse. The name was chosen by the person asking, "clairvoyance" is what someone reaches for by hand at the moment they need it, and accuracy that costs invocability is a bad trade in a library whose measured problem is skills not firing.

---

Beside — what stands next to this once it exists?

Found: the discard pile and tending-tbaguette's candidate queue are the same object seen twice. An observation discarded here because it is not about this project is, by construction, a lesson that generalizes past this project.

Found: managing-scope-drift now has a sibling that could be misread as licensing exactly what it forbids. Resolving that tension in this file only is half a fix — the reader who arrives from the other direction never sees it.

Routed: both to the gate. A line in the routing table pointing discards at the candidate queue, and a reciprocal edge added to managing-scope-drift saying that noticing is not drifting and naming where noticing goes.

---

Ahead — what does this commit its readers to?

Found: the second-order cost of a skill about seeing more is that it becomes the most sophisticated available excuse for not finishing. I looked into it and the request was the wrong shape is indistinguishable from insight until someone asks where the deliverable is, and this skill would be the thing that licensed it.

Routed: to the gate, as a rule that outranks the routing table itself — a sweep is never a reason something did not ship, and a reframe raised and declined leaves the original ledger fully owed.

Also found and discarded: naming seven directions commits the file to that taxonomy. Checked and it is not a one-way door — an eighth would append. No action.

---

Against — what if the answer is the opposite?

Found: almost nothing, and the direction is recorded anyway because that is the point of recording it. Two inversions were tried honestly. Cut it to one page — rejected: without the tells the directions are a creativity listicle, and the tells are most of the length. Do not build it — rejected on the neighbour check done at design time, which found that steelmanning-alternatives and red-teaming-your-own-work both operate on a candidate that already exists, and neither has a move for the request itself being the wrong question.

Routed: discarded, with reasons. The only yield was a trim.

This is what six of every seven directions look like, and a sweep whose every direction produces a finding has stopped measuring and started performing.

---

Outside — who else has solved this?

Found: intelligence analysis has the load-bearing import, and it is the discipline of asking not which hypothesis is best supported but what observation would come out differently depending on which is true — on the grounds that evidence consistent with every hypothesis is evidence for none.

The draft had no equivalent, and its absence was the difference between a technique and a sensibility. Without it, competing readings of a request get argued rather than tested, which is how a sweep turns into an opinion nobody can check.

Routed: to the gate. It became the fifth move, and it is what closes the loop from seeing to knowing.

Discarded: aviation's instrument-versus-visual framing was already carrying the absence section and a second borrowing would have been decoration. TRIZ's contradiction was already in the draft. Medicine's differential diagnosis is the absence enumeration under another name.

---

The tally

Seven directions, six with a finding, one with a trim. Five findings routed back to the gate and changed the shipped file; four discarded with reasons recorded above; one — the reachability re-weighting — changed how the rest of the work was spent rather than changing the document.

The two that mattered most, frames disagreeing and the discriminating observation, arrived through the two directions most likely to be skipped: the one that examines your own vocabulary, and the one that looks like decoration.

---

# The second sweep — the enlargement, swept with itself

Twelve days after the record above, the skill was asked to grow: apply it to many more cases, broaden the horizons of thinking, produce ideas that would not otherwise come, a "God's view" on the request and the project, adapted for all possible cases. This is the record of sweeping that request with the method it was about to enlarge — the eight directions, the new moves, and one orbit pane — before a line of it was written. Same three reasons as before, plus a fourth: an enlargement is the moment a skill about seeing more is most likely to reintroduce the one commitment its first sweep rejected.

The frame the request arrived in: four commitments. That "massively" means bigger; that the skill file is the lever; that more cases means more triggers; and — for the second time — that more sight is the goal. The first survived in part, the second in full, the third was rewritten, and the fourth was rejected again, which is the finding that shaped everything else.

---

Restated without its own nouns

Make the agent notice, on purpose and at chosen moments, things about any object it is handed — not only a request — that the object's own framing hides; and make that noticing produce candidates it can check, without producing work. In that sentence "God's view" became many partial seats chosen on purpose, "ideas" became observations about the possibility space, and "all possible cases" became a table of objects with a default that always works. Nothing in the original wording contained the word object, and it is the noun the whole enlargement turned on.

Behind — what is this request a solution to?

Found: the outcome is not a larger file. It is sessions that produce reframes and ideas the person finds non-obvious, and that notice project-level problems rather than only request-level ones — without scope inflation. The first record found that the scarce thing was reachability, and it still is: a faculty that fires at the wrong moment, or never, is prose. So the enlargement spent its first effort on the routing surface — a description rewritten inside the cap rather than appended to, a selector in the first eight kilobytes because one harness truncates there, and reciprocal edges from the skills whose own moments now hand off here.

Routed: to the gate. In thirty seconds exists because of this direction.

Under — what does the vocabulary assume?

Found: "God's view" admits two readings. One is a seat that sees everything, which does not exist and which a skill claiming it would turn into a manifesto generator. The other is a view assembled from several partial seats, each chosen because it sees one thing the others cannot — which is buildable, boundable, and honest about which seat every sentence came from. The discriminating observation was cheap: does any discipline that does this for a living — intelligence analysis, systems engineering, safety — claim the first reading? None does. All of them assemble.

Also found: "more directions" assumes more directions produce broader thinking. The first record's own text says why that is wrong — a direction without a tell is a creativity listicle, and eight of them run every time is furniture. So every addition here arrived with its tell or it did not arrive.

Routed: both to the gate. The overview says what the phrase cannot mean; orbit is built from seven seats, and its bounds come before its method.

Above — what class is this an instance of?

Found: a single-moment tool becoming a general faculty, which this library has done twice before — orchestrating-work-end-to-end as a spine, and formidable as a default verb, a command table, and depth in reference files selected by the situation. The enlargement copied that shape rather than inventing one: verbs, a selector, a catalogue, and a per-object table.

Routed: to the gate — the structure. Discarded, with the reason: a ninth direction, within (the parts of the request, the one that carries the value). It is under applied to a decomposition, and scoping-before-building already owns splitting a request that is several projects wearing one description.

Beside — what already does most of this?

Found: the neighbours hold instruments a broader clairvoyance would want — steelmanning-alternatives' generator table, threat-modeling's actor tiers, mapping-dependencies' blast radius, revalidating-decisions' premise decay, orienting-in-unfamiliar-code's decisions that are not in the code. Restating them here would have been the duplication the library spends a test suite preventing. So the lens catalogue indexes them: each lens names the skill that owns it, and the lens is the reason to open that skill, not a substitute.

Also found: the description is shared routing budget with every other trigger the skill has, and it had thirty-eight characters left. A trigger appended past the cap displaces an older one — so the description was rewritten as a whole, and the verbs that need no description budget at all carry the rest of the new reach.

Routed: to the gate, both.

Ahead — what does this commit its readers to?

Found: cost. The file grew by half, and every invocation loads all of it. The mitigation is structural rather than a promise: the essentials sit in the first eight kilobytes, and everything a small sweep needs is in In thirty seconds. Also found: orbit is the most dangerous thing in the skill, because a whole project's surroundings are unbounded and a pass rewarded for yield will find something every time. The bounds — a word count, three findings, named moments, never during an incident, never a redesign — were written before the panes, and they are the part of that file a reader most needs.

Routed: to the gate. Bounds first, in orbit.md.

Against — what if the answer is the opposite?

Found: two real inversions. Split it — put the project-level view in a sibling skill and leave this one at request level. It buys a second routing budget and costs one method stated in two files; the person asking chose against it, and the ruling is recorded here so the next session does not re-derive it. Do not enlarge it — the first record's null was mostly right: the seven directions already covered most requests. What they did not cover was the object (a plan, a number, a project) and the field (the people), and those two are the enlargement. Everything else here is depth for those two.

Routed: the split to the record as a ruling; the "do nothing" as the filter that cut the enlargement to objects and the field.

Outside — who else has solved this?

Found: the intelligence community's structured analytic techniques — a catalogue of mechanical methods, each with a stated use and a stated failure, run one at a time when its tell fires — is the shape the lens catalogue took. Process safety's deviation guidewords became the seventh move, because they generate cases and ideas without consulting anyone's imagination. Systems thinking supplied the altitude ladder, the purpose-is-what-it-does pane, and the leverage-point bound. Pace layers supplied the time pane.

Discarded: de Bono's six hats — the seats table does the same job with chairs that name real readers rather than colours; and design thinking's divergence-convergence diamond, which is the routing table drawn as a shape.

Around — who has a stake in this skill?

Found: the person asking, who wants ideas; every reader of the library, who pays the file's cost on every invocation; two harness budgets that decide whether any of it is ever read — a listing budget that surfaces a fraction of trigger descriptions, and a body cap that truncates at eight kilobytes; and another session writing to the same tree that afternoon, which is why the change was landed through a throwaway worktree. The incentive acting on the skill itself is the one the first record named: a pass rewarded for finding things finds things. That is why the null rule is restated for orbit and imagine rather than assumed to carry over.

Routed: the budgets to the gate (the selector's position, the description's rewrite); the shared tree to the run's envelope; the incentive to the null section.

One orbit pane, as a check on the method

Purpose, as revealed by behaviour. The library says it is a set of project-agnostic judgment skills. What it does, measured on one harness that week, is surface a fraction of its own trigger descriptions and invoke a skill in under half of the sessions it is loaded into. The gap between charter and behaviour is reachability, which is the finding the first record reached by argument and this pane reaches by measurement. The check that would refute it is the same harness's listing after the descriptions are shortened, and it is recorded as a reopen condition rather than as work.

The tally

Eight directions and one pane; seven with a finding, one — above — with a structure and a discard. Six findings routed to the gate and changed the shipped files; one ruling (the split) routed to the record; three discards with reasons. The move that mattered most was the restatement: object was not in the request and is now the axis the enlargement runs on. And the commitment rejected twice — that more sight is the goal — is the reason the largest file in this change is a catalogue of tells rather than a catalogue of questions.

Measured, the same day

The enlargement was then tested blind against the version it replaced: four scenarios, two arms, two replicates each, prompts identical apart from which guidance was read, a rubric written before any result, and one blind judge per scenario. The enlarged version scored 104 to 89 and won six of eight pairs with one tie — and the whole margin was discipline: findings marked as inference, each carrying the observation that would test it, bounded, ranked by the cost of finding out, delivered as one decision. On the two criteria that measure insight — a reframe specific to the request, and a point a senior engineer would not have raised — every one of the sixteen answers scored full marks, in both arms. So the honest claim for this skill is not that it produces thoughts the model would not otherwise have had. It is that it makes those thoughts checkable and bounded, and hands them over as a choice. The enlargement also cost about a quarter more tokens per run, all of it reading.

The test found one defect in the enlargement itself: orbit's bound said "one page", and both measured runs came back at roughly three while honouring every numeric bound in the same file exactly. The bound is now a word count. A limit stated in a unit the writer never measures was, it turns out, a wish.

orchestrating-work-end-to-end routes every request to one of seven tracks. Clairvoyance has a seat in all seven, and the seat is genuinely different in each — a sweep run at the wrong moment in a track is not merely lower-value, it is one of the two ways this skill does damage. The other is skipping the routing table.

The shared rule: a sweep goes at the last moment before the frame becomes expensive to change, and never at a moment when something more urgent is running. Everything below is that rule applied — first for the sweep, then for the two larger passes, orbit and imagine, which have fewer seats and stricter ones.

---

Build

Seat: between phase 1 (frame) and phase 2 (design). After the ledger exists — you need the request's own words written down before you can ask whether they are the right words — and before the first approach is proposed.

Why there: it is the only moment in the whole track when a reframe is free. At phase 2 it costs a paragraph. At phase 4 it costs the plan. At phase 6 it costs the implementation, which is why a brilliant reframe at review is usually declined and should be.

Directions that pay: behind first — the ledger you just wrote is a list of mechanisms, and it is worth one question whether they are the outcomes. Then whichever tells fired.

A hit routes to: the gate, row one. The design has not been written yet, so reopening costs nothing but the paragraph you were about to write anyway. Say the reframe out loud, name what it displaces, and let the phase-2 approval decide it — a reframe adopted without the gate is the same defect as a design adopted without the gate.

Second seat: phase 8, restricted to ahead and beside, feeding offering-the-next-move. Run it before the run record is torn down: the record's rulings and surrendered lines are the sweep's best input, and once they are deleted the closing offer gets reconstructed from memory, which reliably yields the obvious next step and nothing that was actually learned.

---

Diagnose

Seat: when the hypothesis list empties, or when the third fix does not hold. Not at the start — a bug with an untried obvious cause does not need a sweep, it needs the reproduction diagnosing-before-fixing asks for.

Why there: three failed fixes is not three bad patches. It is one wrong model, applied three times, correctly. The patches were fine. The thing they were patching was chosen by a frame nobody examined.

Direction that pays: under, near-exclusively. The assumption is in the reproduction, not in the fix — most often that the failure is where it surfaced, that it is deterministic, that it is one bug rather than two, or that the component named in the report is the component at fault.

A hit routes to: back to reproduce, with the assumption named. A sweep that produces a new patch has done the wrong thing; it produces a new hypothesis, which then earns a reproduction like any other.

---

Respond

Seat: at stabilize, and again in the postmortem. Nowhere earlier.

This one is a prohibition, and it is the most important line in this file. An incident has a clock and an audience, responding-to-incidents inverts the usual order deliberately — mitigate before you diagnose — and a frame question asked while users are broken is precisely the "let me understand this properly first" red flag that skill names. Nothing a sweep can produce is worth a minute of ongoing harm. The mitigation may well be crude, wrong-headed, and the thing a sweep would have improved. Ship it anyway.

Directions that pay, once the harm has stopped: under on the cause, ahead on what the mitigation itself has now committed you to — a crude mitigation is a decision, and it is usually still in production a quarter later. And in the postmortem, above: an incident of this shape has often happened before under a different name, and that is the finding worth more than the timeline.

A hit routes to: the postmortem's action items, not into the live stabilization.

---

Investigate

Seat: after gathering, before the conclusion is written.

Why there: an investigation's characteristic failure is answering the question asked with great rigour when the question that mattered was one sideways from it. The evidence is all in hand at this point, which is exactly what makes the reframe checkable rather than speculative.

Directions that pay: behind — why is this being asked, and what decision hangs on the answer? An investigation commissioned to support a decision should be answering the decision's question. And outside, which is where an unfamiliar-feeling question turns out to have a standard name.

A hit routes to: the report, as a named second question with its own confidence marking, per calibrating-confidence. Both questions get answered: the one asked, fully, and the one you believe matters, marked as your inference. Substituting the second for the first is not a reframe, it is not answering.

---

Review

Seat: during coverage, before judgment. orchestrating-work-end-to-end already makes coverage of the diff this track's gate; this is what makes coverage mean more than "every changed file was opened".

Directions that pay: the absence pass from the main file, aimed at the diff. What is not in this change — the caller not updated, the state not handled, the test for the case the author did not think of, the migration for the data already in production. reviewing-code-deeply owns weighing absence alongside presence; this is the method for finding it.

A hit routes to: a finding, phrased as a question about a case rather than as an accusation. And the sweep applies to the review itself: a review that generated forty comments and no observation about the change's shape has inspected without seeing.

---

Author

Seat: twice. At frame the reader, and again before verify.

Why twice: the first is about who this is for, which determines everything downstream and is nearly free to change at that moment. The second catches the document that is now internally consistent and answers a question nobody has.

Directions that pay: beside — who else reads this, and what do they need that the primary reader does not? The on-call engineer, the person evaluating this in a year, the newcomer with none of the context. And ahead — what does this document make impossible to say later? A document that overstates a certainty commits its author's successors to defending it.

A hit routes to: the draft, directly, since the artifact is the deliverable and the frame is the work here rather than a precondition to it.

---

Change in place

Seat: before deciding-reversibility, not after.

Why there: reversibility is a question about a chosen operation. Asked first, it silently accepts that the operation is the right one. The sweep goes one step earlier, at the operation itself.

Direction that pays: against, first and hardest. The cheapest migration is the one that does not happen because the thing being migrated can be deleted; the cheapest upgrade is the one avoided by removing the dependency. Then behind on the thing itself, which is Chesterton's fence in this library's vocabulary: code-archaeology recovers why it exists, and a thing whose reason has expired is a deletion rather than a migration.

A hit routes to: a different operation entirely, which then re-enters this track at reversibility. That is a track staying put, not a downgrade.

---

Where orbit and imagine sit

The sweep has a seat in every track. The two larger passes do not — an orbit is a whole-project view and an imagine pass is deliberate divergence, and both are wrong at most moments of most tracks. Their seats:

Both route exactly as a sweep does. An orbit's findings and an imagine pass's candidates are observations; the moment one becomes work, managing-scope-drift has the table and it wins.

---

The envelope, dial by dial

The track says where the seat is. The envelope says what the sweep may do from it.

The autonomous stop, stated once more

An unattended run may sweep, and may not act on what it finds by changing what is being built. Building something other than what was asked, with nobody present to agree, is not a bounded risk that confidence can retire — it is the definition of the failure bounding-autonomous-work exists to prevent, arrived at through a door marked insight.

The substitution is the standard one: the gate keeps its job and changes its mechanism. Write the observation into the record, name the reading that lost, add a stop condition that fires if the original frame turns out wrong in a way the run can detect, and continue on the original frame. The reframe is then waiting for whoever reads the report — which is the run finishing correctly, not the run being blocked.