• This file routes work. Its linked files own detail. Use current facts, sources, product files, and named owners.
• Minimize wall time only after every rule and proof stays fixed. Do not skip work, cut proof, or overlap required actions.
• assets/speed-policy.json owns safe parallel work, serial work, reuse, and timing. assets/speed-policy.schema.json fixes its shape. evals/speed-budgets.json owns numeric bounds.
• assets/workflow.json is the only step order. assets/execution-ownership.json is the only owner map. Do not skip, merge, replace, or reorder a required step.
• assets/context-routing.json binds each action to full context. assets/context-routing.schema.json fixes route shape.
• assets/context-bundle.schema.json fixes packet shape. assets/section-support.json owns support for each heading.
• assets/section-support.schema.json fixes that owner shape. Read references/context-routing.md and references/section-support.md before work.
• Run

  mise run context-routing

  after a route change.
• The model has wide room only inside its owned work. It may add, split, reverse, mix, or reject options. Keep source rights, fixed rules, and vetoes.
• The model performs every eye, brain, and touch review. The model checks every bad design, bad output, and bad practice invariant.
• Mise tasks check structure. They do not make design judgments.

• help

  : Explain inputs, outputs, routes, tasks, and stops. Match examples/help.md. Do not start work.

• run

  : Create, review, or revise. Reviews change no product file. Revisions change only approved files.
• Before model work: Read examples/run.md before the first run and after any task, path, packet, or output change. Read examples/context-packets.md for full, missing, and unfit context.
• When context is incomplete: Compare evals/files/valid-context-record.json with evals/files/missing-context-record.json. These fixtures fix context accounting. Use examples/failure-missing-proof.md when a fact or proof field is missing. Do not guess.

• 1. Read assets/workflow.json, assets/execution-ownership.json, and assets/context-routing.json. Keep their order, owners, routes, and route hash fixed.
• 2. Read references/intake.md. Write facts in assets/run-intake.schema.json. Validate the intake. Stop only affected work when a needed fact is absent.
• 3. Run

  mise run run-start --intake INPUT --run-dir RUN

  .
• 4. Keep

  RUN

  outside this skill folder. Do not hand-build its files or fields.
• 5. Read every path named by the next packet.
• 6. Run both review tasks when the packet calls for human review.
• 7. Keep one writer for each run file.

The task writes fixed run forms and a route hash. It does not pick states, rules, options, or results.

• Follow every action in assets/workflow.json. Load each named source before acting. Save the named proof.
• Each runtime task rejects an action before its exact turn.
• Inside one action, batch independent reads, renders, captures, and checks after their inputs are fixed.

• run-start

  completes actions 1 and 2. Its next action must be

  source_meaning

  .

1. [CTX-FREEZE-INTAKE]

   freeze_intake

   .

   run-start

   completes this action. Apply references/intake.md and evals/files/missing-proof.json. Never invent a fact.
2. [CTX-LIST-CAPABILITIES]

   list_capabilities

   .

   run-start

   completes this action. Use

   mise run run-scaffold

   to inspect its output. Apply references/failure.md. Never treat structure as sight or touch proof.
3. [CTX-SOURCE-MEANING]

   source_meaning

   . Use

   mise run run-packet

   . Apply references/research.md, evals/source-mapping.json, and evals/context-cases.json. The model records source meaning.
4. [CTX-STATE-JUDGMENT]

   state_judgment

   . Use

   mise run run-packet

   ,

   mise run review-checklist

   , and

   mise run human-sweep

   . Apply references/product-states.md. The model finds states.
5. [CTX-SELECT-RULES]

   select_rules

   . Run

   mise run run-select-rules --run-dir RUN

   . Apply assets/simplicity-contract.json and references/decisions.md. Keep the fixed veto order.
6. [CTX-ATOM-JUDGMENT]

   atom_judgment

   . Use

   mise run run-packet

   . Apply assets/lineage.schema.json and evals/files/valid-lineage.json. The model judges proved atoms.
7. [CTX-PART-DESIGN]

   part_design

   . Use

   mise run run-packet

   and

   mise run lineage-file MANIFEST

   . Apply evals/files/skipped-lineage.json. The model builds linked parts.
8. [CTX-SCREEN-DESIGN]

   screen_design

   . Use

   mise run run-packet

   . Apply references/build.md and examples/context-packets.md. The model builds from proved parts.
9. [CTX-MOTION-JUDGMENT]

   motion_judgment

   . Use

   mise run run-packet

   and

   mise run human-sweep

   . Apply references/review.md. The model judges live motion. Never infer motion from a still.
10. [CTX-VISUAL-REVIEW]

    visual_review

    . Use

    mise run run-packet

    ,

    mise run review-checklist

    , and

    mise run human-sweep

    . Apply assets/review-record.schema.json. The model reviews current pixels and use.
11. [CTX-PLAIN-READBACK]

    plain_readback

    . Use

    mise run run-packet

    . Apply evals/cases.json. The model reads all visible words in context. Never accept clear words in a false state.
12. [CTX-CHECK-LINEAGE]

    check_lineage

    . Use

    mise run lineage-file MANIFEST --run-dir RUN

    . Apply assets/lineage.schema.json. The task checks links, records proof, and advances no design claim.
13. [CTX-FINAL-CHECK]

    final_check

    . Run

    mise run run-check --run-dir RUN

    , then

    mise run complete

    . Apply evals/contract.md and evals/rubric.md. Never turn task proof into design proof.

• Run

  mise run run-packet --run-dir RUN --action ACTION

  . Load the packet before that action.
• Follow assets/context-bundle.schema.json. Keep the packet's route, hash, support, output, and barred swaps.
• Record used paths in

  context_acknowledgements

  . Record unavailable or unfit paths in

  missing_context

  . Use

  BLOCKED

  when missing context prevents the claim.
• Create four unlike starts: known, product-shaped, reversed, and strange. There is no option cap. These starts are prompts, not boxes.
• Add, split, mix, or reverse starts. Explore state, structure, order, words, action, form, motion, sound, touch, access, platform fit, and service effects.
• Give each option a scene, claim, fit, proof, costs, veto check, test, and novelty note. Keep a bold option until evidence or a veto rejects it.
• The model chooses and explains the direction. A Mise task may check fields. It must not rank, score, or choose.
• Run

  mise run run-record --run-dir RUN --result RESULT

  . Save that result before any next action.

• Apply vetoes in this order: safety, access, understanding, agency, task success, convention, expression, novelty, then build cost.
• Do not trade away safety, access, understanding, or agency. Do not hide cost, risk, work, or control.
• Read references/failure.md for missing facts, rights, sources, sight, atom proof, context, clashes, and change.
• Return

  BLOCKED

  for the affected claim. Name the failed gate, missing proof, linked parts, safe work, and smallest way to resume.

• Read references/product-states.md, use assets/state-record.schema.json, and compare examples/product-states.md.
• Use evals/section-support-cases.json to reject closed lists and task-owned choices. Run

  mise run review-checklist

  and

  mise run human-sweep

  .
• Find product states instead of picking them from a closed list. States may overlap, nest, mix, conflict, or change during use.
• Check the person, goal, past act, data, system, place, device, input, time, content, people, and risk.
• Treat common states as prompts only. Skip unfit prompts. Add any state shown by proof.
• Make unlike response options for each state. Record causes, links, change, proof, doubts, tradeoffs, and the model's choice.
• Keep unknown, mixed, stale, short-lived, and concurrent states open. Do not ask for private chain-of-thought.

• Run

  mise run review-checklist

  for fixed check IDs. Run

  mise run human-sweep

  for open eye, brain, and touch prompts.
• Read references/review.md for proof rules. Apply every relevant human factor to every check.
• Eye: Check first sight, order, groups, change, text, color, focus, range, motion, and content.
• Brain: Check truth, action, cause, memory, choice, language, feedback, risk, context, and unknowns.
• Touch: Check targets, gaps, reach, inputs, cues, response, gestures, repair, settings, and effort.
• Objective failures: Reject text or image overlap, clipped meaning, blocked controls, lost work, dead ends, false focus, and false cues.
• Bad work: Check bad design, output, and practice. Include hidden state, false control, old proof, weak options, surface copy, and early choice.
• Answer shape: Give the scene, observation, proof, choice, reason, options, costs, vetoes, and doubt. Mark estimates.
• Evidence state: Keep fact, sight, inference, and change apart.

  NOT_APPLICABLE

  needs proof. Missing needed proof means

  BLOCKED

  for that claim.
• Invalidation: A source or product change invalidates affected captures and checks.

• Read references/build.md before atom or part work.
• Use

  token proof -> tokens -> atoms -> molecules -> organisms -> templates -> screens -> flows

  .
• Run the current

  dtcg-tokens

  skill for all source tokens and atoms. Do not restate its method.
• Stop at the atom gate when that skill or proof is absent. A screen may not invent a low rule.
• Mark linked high parts stale after a low change. Rebuild and review every affected part.
• Run

  mise run lineage-file MANIFEST

  . It checks links, order, loops, stale parts, and unused parts. It does not judge quality.

• Use assets/model-record.schema.json for normal work.
• Use assets/state-record.schema.json for states.
• Use assets/review-record.schema.json for visual review.
• Run

  mise run run-record --run-dir RUN --result RESULT

  .
• The task checks fields, context use, options, and owner lines. A filled record is not a good claim by itself.
• The model must support each result with current proof. Record visible reasons, options, costs, vetoes, and doubt.

• assets/reading-contract.json owns public grade and sentence checks. assets/reading-exceptions.json owns the sole exact-copy exception.

• SKILL.md

  uses standard YAML frontmatter and portable CommonMark only.
• Read references/generation-contract.md only when this public skill changes. Its unchanged copy is the sole full-file reading exception.
• assets/file-manifest.json owns public file roles. evals/source-lineage.json owns public hashes and source links.
• Run shape comes from assets/workflow.json and assets/execution-ownership.json. Run

  mise run run-scaffold

  to view it.
• mise.toml owns all executable tasks and links. .github/workflows/ci.yml runs the same check graph.
• Keep UTF-8, LF newlines, ATX headings, blank block spacing, one-line blocks, and relative local links. Every block stays clean, readable, and agent-parsable.
• Run

  mise run skill-info

  for package facts. Run

  mise run generate

  after public sources change.
• Run

  mise run complete

  for the full fixed gate. Its acyclic links force every needed check to run.
• Mise runs independent read-only checks in parallel and never caches or skips a required task.

• Load eval files only when an eval runs or changes.
• Read evals/manifest.json, evals/contract.md, evals/rubric.md, and assets/eval-case-template.json first.
• Read references/proof-ladder.md, assets/proof-ladder.json, examples/proof-ladder.md, and evals/pilot-cases.json. Run

  mise run proof-ladder

  .
• Run each pilot twice from clean context. Compare action order, context use, proof classes, vetoes, and status. Any clash is

  STALE

  . Wording and safe creative directions may differ.
• evals/cases.json and evals/trigger-cases.json own act and trigger cases.
• evals/evals.json, evals/trigger-queries.json, and evals/section-support-cases.json own local checks.
• evals/context-cases.json owns context-loss cases. evals/speed-budgets.json owns timing.
• evals/source-lineage.json and evals/source-mapping.json own source proof.
• Use evals/files/valid-intake.json and evals/files/missing-proof.json for run gates.
• Use evals/files/valid-lineage.json and evals/files/skipped-lineage.json for part gates.
• Fixture passes prove the task only. They do not prove sight, use, access, taste, product value, or a good choice.

• Return product files and run proof. Follow examples/run.md and the stop shape in examples/failure-missing-proof.md.
• Name sources, gaps, states, options, choice, reviews, atom proof, links, stale parts, context gaps, and vetoes.
• Use only

  PASS

  ,

  STALE

  , or

  BLOCKED

  . Never turn task passes into claims about pixels, use, access, or value.
• Run

  mise run run-check --run-dir RUN

  . Then run

  mise run complete

  .
• Return

  PASS

  only after

  mise run complete

  exits with code

  0

  .
• Confirm that every record and context path is present. See each needed view. Leave no veto open.