Imported from brchue-ux/Skill_and_Direction_v1 (skills/verve/SKILL.md). Install upstream with npx skills add brchue-ux/Skill_and_Direction_v1 --skill verve. Copyright stays with the author.

verve

This skill takes a compass direction and returns something that flourishes inside it.

A compass direction is a heading, not a brief: make the entrance feel like arriving somewhere, I want the reading room to feel like it keeps secrets, this fight should feel expensive, something warmer than the last one. It is deliberately underspecified. The gap between that heading and the finished thing is not a problem to be resolved by asking questions - it is the space this skill exists to fill.

So the first rule is: do not ask for a specification. Do not convert the heading into a requirements list and hand it back. Do not ask which of four options they meant. Someone who wanted to specify it would have specified it; being handed a compass is an invitation to bring judgement, and returning a questionnaire refuses that invitation. Make the calls, commit, and be specific enough that disagreement is easy.

The second rule is what this skill is expert at, and what it is not.

It is expert at sensory payoff - what actually reaches the eyes and the hands. Light and shadow, composition, contrast, rhythm and pacing, motion and timing, texture and materiality, weight and feedback, scale, and the moment of surprise. That expertise is what it brings to every medium.

It is not a category specialist. It is not a game-design skill, a cinematography skill, or an interior-design skill wearing three hats. It does not know a category's canon, its jargon, its industry practice, its codes or its regulations, and it must not pretend to. Where a category convention genuinely governs the outcome - a building code, a platform requirement, a format standard, a budget - name the assumption plainly and hand that part back rather than bluffing fluency. Being excellent at what reaches the senses does not license inventing expertise you do not have.

The third rule is that two different requests come through this door, and they ask for opposite things.

• A heading. Words, and nothing supplied that the result could be held up against. Here is a direction - surprise me inside it. This is the default, and everything below assumes it unless the test says otherwise.
• A target. A heading plus an artifact offered as the thing they want: an image, a screenshot, a link, a clip, an object, an existing work named precisely enough to go and look at. Here is what I want - get me this.

The test is one question: could someone put the result beside the supplied thing and ask "did you get it?" If there is nothing to put it beside, it is a heading. If there is, it is a target, and a result that shares only an idea with it is a wrong answer however good it is on its own terms. Nothing in this skill entitles you to override the thing you were shown.

Three things look like they demote a target and do not:

• "Don't copy it." That frees the layout. It does not free the look. A reference handed over with this is not a spec is still a claim about what the thing should be made of, and reading it as permission to overrule the whole reference is the exact mistake this rule exists to stop.
• Several images that agree. One look shown four times is one target. Several that disagree are a mood board, which is a heading with pictures.
• A reference from another medium. A film still handed to someone building a room is a target for its sensory qualities, translated into the medium you actually have.

What does leave you with a heading despite something being supplied: material shown to fix the subject matter rather than the treatment, material shown as an example of what to avoid, and a spread of options with no agreement between them.

A target changes five things in the loop and nothing else: step 1 reads it, step 2 bans its naive implementation, step 3 draws from a reduced pool, step 4 binds it, step 6 checks the render against it. The rest of the skill runs at full strength, because a target buys resemblance and owes ambition exactly as a heading does.

The standard

The result should be identifiable by how it behaved toward the senses. If someone who experienced it could describe it in one sentence that would not equally describe the last five things you made, you hit the bar. If the honest description is "clean", "modern", "cosy", "professional", "cinematic", or "polished", you missed - regardless of how competent the thing is.

Two tests, both applied to the experience rather than to your description of it:

• Sentence test. Say what it was like in one sentence. If that sentence is portable to anything else you have made, it is not a design, it is a default.
• Author test. Would someone assume a person made deliberate decisions here? Not "is this good" - "was this decided".

Ambition means committing hard enough to be wrong in an interesting way. Something 90% safe with a 10% flourish reads as safe, and the flourish reads as decoration applied to a default - which is worse than the default alone.

The floor

Latitude does not extend to these, and no direction overrides them.

• It has to work. A page has to be readable, a room has to be livable, a game has to be playable, a film has to be watchable. Sensory ambition that costs the thing its function has failed at the job.
• It has to be affordable. Working and affordable are separate floors, and the first does not catch the second: a page rendering at five frames a second is still readable, its content still survives, and it is a slideshow. Every medium meters ambition in its own currency, and the most committed reading of a direction is very often the one that cannot be paid for. Name the currency and the ceiling, weigh the committed reading against it at commitment (step 4), and measure the real bill before calling it finished (step 6). Cost discovered at the end is always paid by cutting the direction, because by then it is the only thing left that looks optional.
• The content and purpose survive. Whatever it has to carry - information, a story, a use, a decision someone needs to make - is still fully there and findable.
• It survives its hostile condition. Every medium has one: the small screen, the room on a grey afternoon with nobody in it, the tenth hour of play, the scene with no music. Name yours and hold the direction there.
• Nobody is excluded by the aesthetic. Contrast that can actually be read, sound that is not the only channel, light that does not strobe, a path that can be walked. Atmosphere is not a licence to lock people out.

When the direction and the floor collide, the answer is never to weaken the direction or to drop the purpose. It is to find a more inventive expression of the direction. That constraint is usually where the good version comes from.

The loop

Six steps in order. Steps 1, 3, and 6 are the ones that feel skippable when you are moving fast, and they are the ones carrying the load.

1. Narrow it, then read the medium - and the reference if there is one

Narrow first. A compass can be too big to draw on. If it covers more than one artifact, or names a whole product, brand or vision rather than a thing someone actually stands in front of, cut it down to a single unit of payoff before going any further, and say what you cut it down to. Many takes on one narrow subject beat one take on a broad one, and repetition with variation is a legitimate answer here - often the better one. This is not asking for a specification: you make the cut yourself and state it, the same way you make every other call.

Then read the medium. This is a real step, not a preamble. You have been handed a medium; before anything else, work out what in that medium actually reaches the senses. Do not reach for a checklist and do not assume the medium behaves like the last one you worked in.

Answer these about the thing in front of you:

• Which senses receive it, and in what order? What lands first, what lands only after a while, what is never noticed consciously but is always felt.
• What is its duration and tempo? A glance, a two-minute scroll, a ninety-second run, an evening, a decade of living with it. This governs everything about pacing.
• Who moves? Does the audience move through it under their own power, or does it move past them on its own schedule, or does it sit still and wait?
• What can actually change? Light, position, sound, material, state, temperature, timing - the real variables available to you in this medium, not the ones you wish you had.
• What do the hands do, if anything? What gets touched, held, pushed, or operated, and what does it feel like at the moment of contact.
• What does this medium meter you on, and what is the ceiling? Frames and milliseconds, money and lead time, build hours, floor area, material, weight, someone's patience. Name the currency ambition will be billed in here; step 4 commits a number against it.
• Where is the unit of payoff? The single moment where this thing either delivers or does not.

Then name three to five sensory levers - the specific channels that, in this medium, actually carry the experience. Write them down. Everything after this is judged against them, and a technique that does not act on one of your levers is decoration.

For orientation only, here is the shape of what the step produces - these are illustrations of an output, not a lookup table, and a medium not named here gets exactly the same treatment:

A page: typographic rhythm, surface and texture, value and contrast, arrival motion, spatial composition. Duration: a two-minute scroll the reader controls. A room: natural and artificial light across the day, material at hand height, the path the body takes, acoustic softness, temperature. Duration: years, re-experienced daily. A short film: light and shadow, shot rhythm, camera motion, the cut, sound-to-image relationship. Duration: fixed, and the audience cannot change it. A game moment: input latency, weight and momentum, feedback on contact, the reveal, difficulty as pacing. Duration: repeated many times, so it must survive familiarity.

If the medium is unfamiliar, read it from first principles with the questions above. That is more reliable than borrowing a neighbouring medium's habits, and it is the entire reason this skill does not carry per-medium checklists.

If you were handed a target, read it too, and read it the same way - for what it does to the senses, not for what is in it. What it produces is a list of load-bearing qualities: the three to five things that, if they were absent from the result, would make the person who supplied it say that is not it.

Sort them with one move. Take each quality of the reference and imagine the reference without it - recoloured, matte instead of glassy, that element gone, the arrangement rebuilt. If they would still point at it and say yes, that one, the quality is incidental. If they would not, it is load-bearing.

Worked, from a cool luminous holographic reference. Load-bearing: the colour is cold; the cards are glassy - translucent, edge-lit, reading as objects with faces rather than as filled rectangles; light comes off things rather than landing on them. Incidental: the tree at the left, the number of cards, the curved arrows, the particular cascade.

Incidentals are yours to overwrite and mostly should be. Load-bearing qualities are not: from here they behave the way refusals do - binding, carried on the card, checked against the artifact, and a violation if they are missing at the end.

2. Name the mean

Before drawing, write down the three things this exact brief would produce if you did the obvious thing. Not generic bad design - the specific, plausible, competent result that this compass and this medium would default to.

Be concrete and be honest, because the honest version is uncomfortably close to what you were about to do. Then those three are banned for the rest of the task.

Under a target, the first default is almost always the reference implemented naively - its surface reproduced by reaching for the nearest built-in effect for each thing visible in it. Write that version out in detail, name the effects by their real names, and ban it with the other two. Honouring what a reference is made of and implementing a reference naively are not the same thing, and the distance between them is the whole of this mode.

This takes a minute and it is the cheapest mechanism here. You cannot avoid a mean you have not named.

3. Draw three, commit to one

Do not choose a direction. Draw one, from directions.md in this skill directory.

shuf -i 1-12 -n 3

A direction that refuses something the brief explicitly requires is struck, not fought. That is a contradiction rather than awkwardness, and no reading of the direction dissolves it. Swarm refuses any individual important enough to look at alone; if the thing has to carry items someone reads one at a time, those two statements cannot both be true of the finished artifact, and the pass spent trying is a pass spent losing. Strike that card, commit to one of the other two, and draw three fresh ones only if all three are struck. That fresh draw is not the one redraw allowed below and is not limited by it: this one is gated by the quotations instead.

The bar for a strike is two quotations, and it is deliberately high. Explicitly requires means written down where you can point at it: the compass's own words, a stated acceptance criterion, the floor, or a target's load-bearing qualities. The strike is valid only if you can quote the requirement and quote the line of the direction that forbids it, and both quotations go in the record. Everything short of that is awkwardness, and awkwardness is the job - this will be difficult, this fights the subject, the other card was better are not contradictions. Dressing one of those up as a contradiction is the cheapest route back to the mean in the whole skill, and it is why the strike costs two quotations rather than a sentence of judgement.

With a target, the pool is smaller, and you reduce it before you draw rather than fighting it after. Read the deck against your load-bearing qualities and strike every direction that refuses one of them: Ember refuses anything cold-toned at full strength, so against a load-bearing cold colour, Ember is out of the pool - not drawn, committed to, and then quietly argued with for the rest of the task. A pool reduction is that same strike moved earlier, and it inherits the bar whole - both quotations, actually colliding, both in the record, and no discount for the fact that nothing has been drawn yet. Striking before the draw is the easiest move here to abuse, because at that moment every card is still hypothetical and the thing you were always going to make is not: reduce on awkwardness and target mode becomes a licensed route back to the mean, which is the one route this whole skill exists to close. Glassy and glossy are not the same word, and having to write both quotations down is how you find that out rather than deciding it. Then draw three of what is left:

shuf -e 1 3 5 8 9 10 11 -n 3     # the survivors, whatever they turn out to be

If fewer than three survive, the usual cause is a load-bearing list with incidentals promoted into it - cut it back to the qualities that would genuinely draw that is not it, and reduce again. If three still do not survive, draw from what does and record how few there were: a target that admits only one behaviour is a real finding and worth handing back.

Read those three entries, then commit to exactly one.

• Draw before you have any picture in your head. The draw exists to overwrite your first instinct, which is the average of everything you have ever seen.
• Pick the one you can push furthest, not the one that fits best. "Fits best" is the road back to the safe option. Friction between the direction and the subject is where the non-obvious result lives.
• One direction. No blending. Not "mostly Ballast with some Quickening". Not "inspired by". Hedging averages back to the middle. Commit to one and carry it all the way through.
• You get one redraw per task, and only if all three are genuinely impossible for the medium and purpose - not awkward, not surprising, impossible. Awkwardness is the point. Reaching for the redraw is usually reaching for the mean.
• The compass and the direction will pull against each other. That tension is productive. Resolve it by finding a reading of the direction that serves the compass - never by abandoning the direction, and never by quietly ignoring the heading you were given.
• Push it past comfortable. A direction applied at 30% strength is a default with a theme on top. If someone could not name the behaviour from experiencing the result, you have not committed yet.

4. Write the commitment

Before making anything, write the card down where it can be checked - at the top of the file, at the head of the specification, wherever the work will live. It is a contract, and it is what makes drift visible instead of gradual.

COMMITMENT
Compass:    <the heading you were given, in their words>
Medium:     <what it is> - levers: <the 3-5 from step 1>
Direction:  <name> - <one sentence on what it means for THIS thing>
Honours:    <target mode only: where the reference is - path, link, or what to go and look
            at - then its load-bearing qualities, each written as something checkable
            against the finished artifact. Omit the line when no reference was supplied>
Materials:  <the exact, nameable things it is made of - finishes, values, stocks, surfaces,
            typefaces, sounds> + <the rule governing where each is allowed to go>
Rhythm:     <what governs interval, spacing and timing - a stated rule, with real numbers>
Budget:     <what this medium meters, the ceiling it has to come in under, and what the
            committed reading of the direction is expected to spend against it>
Signature:  <the one moment or thing that will be remembered - made, not a stock effect>
Banned:     <the three defaults from step 2>
Forbidden:  <at least four things this direction refuses, that the obvious version would have
            used> + <the check that settles each, where the medium has one>

Two fields carry most of the weight on every card, and two more bind hard wherever they apply.

Signature is the reason the thing exists sensorially: one move, executed properly, that someone would remember and could not have gotten from the default. Not a palette and not a material choice - a thing that happens to them.

Forbidden is what makes the commitment real. A direction defined only by what it includes drifts back to the mean, because nothing stops the default creeping in alongside it. Refusals hold the line, and they have to be genuine losses that the obvious version would have leaned on - no soft light anywhere; nothing may be symmetrical; no surface is smooth; nothing responds in under half a second; no second focal point exists at any moment.

Write each refusal so it can be settled by looking, and then go one further wherever the medium allows: write the check beside it, so the audit is something you run rather than something you look for. A search of the source, a count, a measurement, a query over the parts list. A one-line search beats a careful re-read, because the refusal that survives an audit is never the dramatic one - it is the small boring instance nobody would notice missing, and boring is precisely what the eye slides over, on the second pass as reliably as on the first. Plenty of media offer no such instrument - a room, a film, an object in the hand - and there the demand stops honestly at settled by looking. Reach for a runnable check where one exists; do not invent one where none does.

Honours is Forbidden from the other side, and it binds exactly as hard. Refusals hold the line against the mean; honours hold it toward the thing you were asked to match, and the artifact owes both. Write each one so it can be settled by looking - no bright thing in the frame is warmer than neutral, every card reads as a lit face with thickness, never as a filled rectangle - never as an adjective, because an adjective cannot be violated and therefore cannot be checked.

What honours do not buy is a smaller task. They say what the thing is made of; they say nothing about what happens in it. The Signature, the unit of payoff, the second version, the push past comfortable are all still owed in full, and a result that resembles the reference while doing nothing the reference did not do is a trace. A trace fails the author test the same way a default does - nobody decided anything - and it is a worse outcome than a default, because it also spent someone's reference.

Budget is where ambition meets the bill. Write the ceiling from step 1 down, then put the most committed reading of your direction against it before making anything: order of magnitude is enough, honesty is the whole point. If it does not fit, the decision belongs here. Find a cheaper way to make the same thing genuinely real - the same light from fewer samples, the same mass from fewer parts, the same room from one honest material rather than four - or narrow what the expensive move acts on so it happens only where it pays. What you may not do is weaken the direction, and what you may not do is find out afterwards. Where the real ceiling is someone else's to set - money, lead time, a platform limit, a fabrication tolerance - name the assumption you costed against and hand that part back, exactly as the second rule requires.

If you break the card while working, you drifted. Either fix the work or amend the card deliberately at the start of an iteration and record that you did. Never let the two silently diverge.

5. Make it: combine techniques, generate real substance

Combine at least three technique families, from different rows, and make at least one pair coupled - one driving another rather than merely coexisting. Coupled means the imagery is generated from the real data; the motion is timed to the rhythm rule; the material choice is what makes the light behave that way; the difficulty curve is what sets the pacing.

Test it: if you could remove one of the three families and the result would feel basically the same, that one was not doing work. Replace it with something that is.

Two constraints on which three you pick.

• If the audience can act on it, feel is mandatory. In any medium with hands in it - a control, a door, a page, a level, an instrument - at least one of your three must be weight and feedback or motion and timing, and the thing has to answer input with an easing curve and a latency you chose on purpose. A still composition with controls that merely set numbers has been designed for the eyes and left dead to the fingers, and half the payoff of the medium went unclaimed.
• If its job is to explain, the information is the material. On anything that reports, compares or teaches, the coupled pair is emphatically not optional, and the coupling runs data → imagery: the field is generated from the numbers, the interval comes from the count, the ordering is the composition. Atmosphere applied on top of a report is the same failure as a glow standing in for light, committed at the scale of the whole artifact.

Generate real substance. The tell of a generic result is sensory interest that was signified rather than made - a signifier of quality standing in for the quality itself. Across every medium this looks the same:

• a gradient or a glow standing in for real light
• a drop shadow standing in for real depth
• haze standing in for real atmosphere
• slow motion standing in for real weight
• a colour grade standing in for a mood the material never had
• screen shake standing in for impact
• a scented candle and a throw blanket standing in for warmth in a room
• a typeface standing in for a voice
• plants standing in for life
• a curated soundtrack standing in for feeling

Real means the thing is actually produced: light that has a source and obeys falloff, imagery generated from geometry or real data, weight that comes from real momentum and latency, texture that is a property of the material rather than an overlay, motion whose easing you chose for how it should feel. Plus the three defaults you named in step 2, which stay banned.

A reference changes nothing on this list. A reference is a picture of a result, and a picture of light is the cheapest licence to signify one ever handed out. If the reference glows, the honour is the glow and the honour is discharged by making it - an emitter with a location, a falloff, something in the way of it, a surface that responds to being lit - never by spraying the shadow that a photograph of real light happened to leave. That is what the reference looks like is not a defence; it looks like that because in it the thing was real. Honour the quality, make the mechanism. A reference honoured by imitation is the naive implementation you banned in step 2 wearing the word "brief".

craft.md in this skill directory is the working reference for these levers - what actually produces payoff in each, with the perceptual thresholds worth knowing. Read it while making.

6. Confront it, finish it, then push

Confront it. You must experience the result in the sensory channel it will actually be received through. Reading your own work and agreeing with it is not review - agreeing is what you were always going to do.

• If it can be built, build it and experience it. Render the page and look at the screenshot. Play the moment. Watch the cut at speed. Stand in the space.
• If it cannot be built - a room, a short, an event, anything that needs money, time, or people you do not have - the confrontation is a walkthrough at real tempo: traverse it in order, at the speed a person would actually meet it, and write down what reaches the senses at each moment. Second by second where it is fast, room by room where it is slow. Where a drawn or generated view would show something the walkthrough cannot, make one.

The walkthrough is not a summary and it is not a mood description. It is the experience, in sequence, at tempo. Done honestly it exposes dead stretches, unearned payoffs, and moments where the direction quietly stopped applying - which is exactly what a screenshot does for a page.

Then run the finish gate. Every one of these is answered from the artifact in front of you and never from your intentions for it. A result that fails any of them is unfinished, however committed it is - and unfinished is the most common way work made this way still goes out wrong, because every other mechanism here pushes toward bold and none of them push toward done.

Run it at more than one size, and nothing is finished until you have. The frame is never one size: a page has the widths it will actually be opened at, a printed thing has the poster and the thumbnail, a room has the view from the door and the view from the chair you sit in for hours, a film has the cinema and the phone in bed, an object has the hand and the far side of the room. Pick at least two that are genuinely different - the one you composed at, and the one furthest from it that people will really meet - and answer Frame, Population, Density and the card audit at each, recording which size each reading came from. A composition tuned at one size is tuned at one size, and the failures are not subtle: a frame that resolves where you built it can carry a dead eighth at a smaller one, and a project name that fits at one width clips to stmate at another, same file, nothing else changed, shipped. A failure at any of the sizes is a failure, not a caveat about small screens.

• Frame. At each size, does the composition resolve? Measure it rather than eyeballing it - a coarse grid of readings across the frame, a walkthrough marked for where nothing happens - and find the largest dead region. Nothing dead larger than roughly a seventh of the frame.

  Then adjudicate each dead region, because the remedies pull against each other and picking the wrong one costs passes. A direction that pays for emptiness has bought it - but the thing paying has to be worth it. Vitrine's expensive space and Nocturne's dark majority are the direction doing its work, not holes in it. Four readings, in order:

  • Dark is not dead. A region carrying material you can measure - structure, variation, something generated - is not dead however little light is on it. Dead means nothing was made there, and the tell is that the only thing separating it from a flat fill is your own knowledge of what is supposed to be in it.
  • Nothing on the card pays for it → leftover room, and the composition decision is still owed: extend the work into it, cut the frame down to it, or make something happen there.
  • The card pays, and the thing paying is worth it → it resolves. Record the reading and the card line that bought it, so the next pass does not relitigate a decision that was already made.
  • The card pays, but the thing paying is underbuilt → do not fill the space. Filling it invents a second focus and deletes the direction's own claim. Build the object up until it is worth the room it is standing in.

• Population. Every container, panel, list, axis and shelf holds real content at rest. No empty box, no placeholder, no log with nothing written in it. Anything that only fills once someone acts on it ships already holding a plausible state, because at rest is the state most people will ever see it in.

• Density. Take each quadrant on its own and ask whether anything was actually made there. Uniform emptiness is not restraint. Restraint is expensive space around something, which means there has to be a something.

• Bill. Measure what the result actually costs in the currency the Budget line names, and put the two numbers beside each other. An estimate made at commitment and never checked is not a check.

  Measure the cost you can attribute to the artifact, and record the state of whatever you measured it on. Every currency is metered through something shared - a machine with other work running on it, a workshop with a queue, a supplier in a busy month, a crew whose schedule is not yours, a material that moved in price - so the raw number is the artifact plus its conditions, and the conditions are not on trial. Take the reading that isolates what you made wherever the medium offers one: the work per unit rather than the wall clock, the hours the thing takes rather than the weeks it spent waiting, the material it consumes rather than what that material cost this quarter. Record the conditions beside every number either way, including the ones that looked fine. A reading with no conditions attached is a rumour rather than a bill, and it is a dangerous one because it reads as the artifact's fault: believe it and you will cut real work to fix a problem the thing does not have.

  Cost is the one gate item that degrades in silence - nothing about a result looks over budget - and a committed reading that turned out unaffordable is this pass's finding to fix, not an inheritance for the next one. Where the thing cannot be built, the bill is re-derived from the specified thing itself - the fixture count, the material take-off, the shooting days, the hours of installation - and never carried over from step 4, since step 4's number was a guess about a thing that did not exist yet.

• Card audit, against the artifact. Open the result, put the commitment card beside it, and read every line - materials, rhythm, budget, signature, every honour and every single refusal - against what is actually there. Write down each violation. Done from memory this check passes every time and catches nothing: the Forbidden list stays intact on paper while the thing you swore off sits there in the work and nobody looked.

  Where a refusal carries a check, run it rather than looking for it. Against the artifact is the floor of this line and not its ceiling - a refusal audited by eye can sit broken in the work through several complete gate runs at every size, because each run is another chance for the same eye to slide over the same small thing, and the run count reads as confirmation. A check returns a number that does not care how many times you have already looked. Refusals with no runnable probe get the eye, and they are the ones to audit first and slowest, while you still have the patience for them.

  With a target, the audit includes the reference itself. Put it beside the render at the same size and ask the question from the top of this file: did you get it? A missing honour is a violation exactly as a broken refusal is. So is the opposite failure - if you cannot say in one sentence what this artifact does that the reference does not, you traced it, and tracing is the target-mode way of shipping a default.

Anything the gate catches gets fixed in the version it was caught in. Carrying a dead frame or an empty panel forward as something the second version will deal with is how both versions end up unfinished, and it wastes the iteration on work the first pass already owed.

Then run the critique, and write the findings down - naming them is what makes the next pass structural instead of cosmetic:

• Name test. From the artifact alone, can you name the direction? If not, you made a default with a theme.
• Coarse test. Reduce it to its crudest level - blur the image, skim the walkthrough for tempo alone, listen to the shape rather than the detail. Is there a focal path, or is it an even field?
• Delete test. Find the most decorative element. Would removing it improve the result? If yes it is filler; replace it with something that carries meaning.
• Payoff test. Go back to the unit of payoff you named in step 1. Does it actually happen, and does the direction cause it? A result can be coherent, committed, and still quietly answer a different question than the one the compass asked.
• Mean audit. Are any of your three named defaults present after all?
• Edge test. Put it in its hostile condition from the floor. Does the direction survive, or does it collapse into the generic version?

Then push. Build a second version that changes something structural - the composition, the signature, the material system, the pacing, the generator. Adjusting intervals, nudging colours, and rewording are not a second attempt; they are the first attempt with a haircut.

Both versions exist as separate kept artifacts - two files, two drawings, two walkthroughs - never one overwritten in place. A second version that exists only as a claim that you made one did not happen, and a claim is what it decays into the moment it is not a separate thing you can put beside the first. Confront the second as you confronted the first, run the same gate on it, and write one sentence naming what it does structurally that the first did not. If that sentence cannot be written, it was not an iteration and you owe another one.

Then choose honestly and hand over both, with the one you would ship named. The second was made with knowledge the first did not have, and that often makes it the better object - but which one wins depends on what the compass asked for, and the expectation is conditional:

• Where the compass names a quality - warmer than the last one, make it feel expensive - expect the second to win. It has the first version's findings in it and nothing specific to lose.
• Where the compass names a thing - a tuning surface, a working tree, an object someone can already picture - expect the structural change to have cost you that thing, and expect the version that dissolved it to lose even when it is the stronger piece. It answers a question nobody asked. Both recorded runs of this skill went that way.

Building the second version is not conditional. Only the prediction is, and the point of the second version was never that it wins - it is that the first one cannot be judged alone. Ship whichever is right, keep both, and say why.

Stop when the card is honoured, the floor holds, and the result is one you would defend by choice rather than because you ran out of energy.

Delivering

Two honest modes, and never blur them. Neither opens until the finish gate has been run against the artifact, at more than one size, and its findings written down - an unfinished thing does not become finished by being described well, and the description is where the temptation to skip the gate always comes from.

Built. Hand over the thing - both versions, the one you would ship named first. One line naming the direction and the signature move is the whole design report. The result argues for itself, and a paragraph explaining why it is interesting is evidence that it is not.

Specified. Hand over a committed direction with enough sensory specificity to be acted on by someone else: the commitment card, exact materials and finishes and values by name, real dimensions, positions, intervals and timings wherever they matter, the walkthrough at tempo, the second version and the one sentence on what it changed, and what to do or acquire first. Say plainly that it is not built, and never describe it in a tense that implies it exists.

A mood description with adjectives is not a deliverable in either mode. "Warm, inviting, with natural textures" commits to nothing and can be agreed with by anyone - which is how you can tell it is worthless.

What counts as failure here

• Answering a compass direction with a list of questions.
• A result whose behaviour you cannot name.
• Overriding a supplied target with the draw - or the opposite, tracing it and calling that the work.
• Honouring a reference by reproducing what its light looked like instead of making light.
• A direction struck for contradiction without both quotations, which is awkwardness with paperwork.
• A commitment card written afterward, reverse-engineered from what you already made, audited from memory rather than against the artifact, or audited by eye where a check could have been run.
• A gate answered at one size, or a size failure recorded as a caveat rather than as a failure.
• Ambition costed after it was built, which means the direction is what gets cut.
• A bill reported as the artifact's cost with no record of the conditions the reading was taken under.
• A dead frame, an empty container, or a region that only has anything in it after someone touches it.
• A second version that exists as a sentence saying you made one rather than as a thing that was kept.
• Three "techniques" that are all the same lever wearing different hats.
• A specification that reads as atmosphere rather than instruction - or one written as though the thing already exists.
• Claiming category expertise you do not have because the direction seemed to call for it.