ChaseSunstrom/directions

---

name: directions description: Use whenever a new UI needs a look: six style directions, the pick saved as DESIGN.md.

Sampo: style directions

One sentence about a product in; six distinct style directions out, each a card that fits on one screen; one is picked and saved as DESIGN.md with an audacity level. Start it with /sampo:directions <sentence> (in Claude Code or a Smith (sampo smith) session), or start it yourself when the user asks for looks or styles to choose from before a UI exists. It is for web and app screens; a terminal UI or a Claude Code mod takes the playbooks' terminal-ui.md instead.

It widens the UI playbook ../playbooks/references/ui/design-direction.md from one direction to six to choose from: that file holds the taste words, the direction bank, the audacity dial, "Defaults to avoid" and the DESIGN.md template. Read it, then references/cards.md (the card format, the rules for the set, the generic defaults, the preview page recipe and a worked example) before drafting.

Directions come from three shelves: the playbook's twelve-row bank; the twenty looks (sampo lib list --area looks, then sampo lib show look-NAME: measured palettes, type, layout, motion and a contrast test); and the shader looks (sampo lib show shader-NAME) as an optional texture on one surface, never as the direction itself.

The user's taste orders the looks: sampo looks ranks the twenty by what they kept, reverted, called "not like that" or picked in a calibration (/sampo:looks calibrate). Draw the looks you offer from its top, never one below 0 unless the user names it or none at 0 or above fits, else the least negative, and say which card the ranking favoured; with no signal yet every look is 0. Nothing detects a kept or reverted look on its own: when the user keeps or reverts one, run sampo looks keep or sampo looks revert yourself (step 5 says when).

Procedure

1. A brief, and an existing direction. Saving DESIGN.md is an edit, so work inside a focused task; with none, sampo task new "Style direction for <product>" --type FEATURE --tier S --ac "DESIGN.md names a direction, a runner-up and an audacity :: grep -qE 'audacity (safe|bold|wild)' DESIGN.md" --focus. If the repo already has a direction (a DESIGN.md with a ## Direction section, a brand guide or a shipped design system), it wins (the playbook's step 1): name it and stop, unless the user asked for a new look. Then draft against it (its direction is not offered again) and replace the file; git keeps the old one, and the new Direction section says which direction it replaced. A token file, or a DESIGN.md holding only the Tokens section sampo design extract writes, is a palette, not a direction: draft as usual, each card saying whether it keeps that palette.
2. Read the sentence. Take the product, the audience and any taste words (the playbook's table; its empty words set no knob). Answer interview questions 1 (who, and what they were doing a minute before) and 3 (what it must never be mistaken for) from the sentence, else from the repo (README, existing copy), marked assumed. Don't ask question 2: the six cards are its answers, one physical thing each. If the sentence names no product at all ("make it pretty"), ask one question, what it is and for whom; under full autonomy read that from the README.
3. Draft six cards in the format of references/cards.md, then hold the whole set to its rules:
   • at least four from the bank or the looks, at most two own;
   • spread: a dark ground and a mid or tinted one, three type classes, a sparse and a dense card, a still and a lively one, and each audacity setting (safe, bold, wild) at least once;
   • any two differ on at least three of the five knobs and never share a physical thing or a signature move;
   • each card's body text pair measured at 4.5:1 or more (a look's palette is measured already; otherwise use the script in sampo lib show look-lido --section "Test first"), its accent 30° or more from the status hues;
   • none is a generic default from cards.md, none breaks "Defaults to avoid", and no look whose Not for: names this kind of product. Redraft a card that fails, at most twice per slot, then fill the slot with another bank row or look; keep one line on each card you dropped and why.
4. Show them. All six cards in one message, then the comparison table (one row each: ground, type, density, motion, radius, audacity) so the spread shows, the dropped ones, and your recommendation and runner-up, each with its reason. Recommend at the playbook's default dial for the surface (safe for settings, admin and forms; bold for a product's home; wild for games, portfolios and events) unless the user's words move it. Preview pages only when asked ("show me", "mock them up") or when the user is torn between two: cards.md's recipe, outside the repo, then sampo design lint over them. Diverge before you converge when the user is torn or every card feels alike: a three-variant round (sampo design diverge 3 <the screen> --out DIR: briefs in looks far apart, each built in its own lane, the winner kept with sampo design diverge pick LOOK); a taste duel or four-look tournament the user answers (/sampo:looks tournament); opposite day (sampo looks opposite LOOK, the looks farthest from the current one); a lookbook page (sampo looks book --out FILE); or five historical idioms of one layout (references/idioms.md).
5. The pick.
   • Standard autonomy: one AskUserQuestion with two questions. "Which direction?": four options, your recommendation first, then the next three; its free-text answer takes any number from 1 to 6, or a blend of two (one owns type, the other colour; never three). "How bold?": the picked card's audacity first, then the other two settings. A picked look goes in the ledger: sampo looks keep look-NAME. A shown look the user turns down ("not like that"): sampo looks not look-NAME --why '<their words>' (words with no quote character; else leave out --why), then redraft that slot. A saved look the user later replaces or reverts: sampo looks revert look-NAME. An AskUserQuestion that is denied records nothing.
   • Full autonomy (sampo autonomy says full): no question. Take your recommendation at its audacity and record it: sampo decide "Style direction: <name>, audacity <setting>" --task <ID> --why "<why it fits the sentence>" --rejected "<the other five, a reason each>". Name the pick and the runner-up in the final report, so the user can switch with one word. Record nothing in sampo looks: it holds the user's choices only.
6. Save DESIGN.md at the repo root (or beside an existing design-system package), filling the playbook's DESIGN.md template with every heading in its order: Point of view, Interview, Direction, Knobs, Signature moves, Banned, Allowed exceptions. An existing Tokens section stays as it is (rerun sampo design extract <ui root> once the token file follows the pick); with none, Tokens waits for the token file and sampo design extract writes it.
   • Point of view and Interview from step 2; "Feels like" is the picked card's physical thing, marked user when the user picked it and assumed under full autonomy.
   • Direction: <bank row, look-NAME or "own: NAME">, audacity <safe | bold | wild>., then the runner-up and why it lost, then the taste-word split. A blend names both and what each owns: Gallery (type) + Negative file (colour), audacity bold.
   • Knobs: the card widened to the template's detail; every colour with its role, and the measured contrast of each text and background pair.
   • Signature moves: as many as the audacity row allows (safe 1, bold 2, wild 3), the card's first. A look that calls for sound takes it from references/sonic.md (off by default, never the only signal).
   • Banned: the anti-reference, the generic default the set came closest to, and each "Defaults to avoid" line that applies. Check it, and record the check as the step's evidence (sampo task evidence <ID> --step N --run "<cmd>"): grep -cE '^## (Point of view|Interview|Direction|Knobs|Signature moves|Banned|Allowed exceptions)$' DESIGN.md prints 7, and grep -E 'audacity (safe|bold|wild)' DESIGN.md finds the Direction line.
7. Hand on. Building follows the playbook's step 7: build, then its Checks (the cover test included). Preview pages stay out of the repo. Close the task through the Smith loop, or carry on when the look is one step of a larger task.

A second run finds the DESIGN.md the first one saved and stops at step 1, unless the user asks for a new look.