Vision

/vision

You are running the vision skill. Produce a VISION.md the author can approve: an acceptance policy for the project's future, grounded in what they actually build, and sharpened by hypotheticals they answer on a review board.

This is not a writing exercise. Follow this file top to bottom.

Provenance

Adapted from the upstream kunchenguid/vision Agent Skill (MIT). The pipeline, hard rules, output anatomy, and the review-board house style are upstream. Three things are localized to this repo and are the reason this is a first-class skill rather than a vendor-manifest.json entry — the vendor fetcher overwrites files in place and would clobber all three:

1. Evidence ladder (Step 3) — mines this repo's decision artifacts (OpenSpec specs and archived proposals, docs/decisions/ capability timelines, merge logs) ahead of PR titles, rather than treating merged-PR history as the primary source.
2. Review-loop transport (Step 6) — verdicts return through AskUserQuestion, this repo's established human-gate tool, instead of npx -y lavish-axi. No external service and no npm dependency sit between the author and their own vision.
3. Repo conventions — house frontmatter keys, the <skill-base-dir> path rule, the worktree mutation guard, and the tail-block contract enforced by skills/tests/vision/.

When syncing upstream improvements, re-adapt these three; do not overwrite them.

Hard rules

1. Evidence over vibes. Every principle in the draft must be traceable to concrete evidence: a named spec requirement, archived proposal, ADR entry, PR, commit, file, or the author's recorded answers. Generic engineering virtues ("we value quality") are banned unless the history demonstrates them specifically.
2. Check for an existing VISION.md first. If one exists on the default branch, switch to delta mode: treat it as the approved baseline, propose line-level candidate changes from evidence newer than it, and never write a competing document.
3. The author owns the vision. You draft, stress-test, and fold in their verdicts; you never approve, never soften a hypothetical to please, and never fold in a principle they did not state or demonstrate.
4. A vision is an acceptance policy. Write testable accept/resist criteria in declarative present tense, with explicit non-goals, so a future reader — human or agent — can apply them to a concrete change.
5. No softball hypotheticals. Each one must sit on a genuine fault line where yes and no are both defensible, with both sides steelmanned. If you can predict the author's answer, replace the hypothetical.
6. The review board is built from the shipped template, and verdicts come back through the host. The board is the shipped template with only its slots filled, never restyled or restructured; verdicts return via AskUserQuestion, never by you reading an answer into the record yourself. Mechanics in Step 6.
7. Iterate in batches, trace every edit. Each author verdict maps to a named edit in a changelog; the author must be able to see exactly how their answer changed the text.
8. Formatting of the VISION.md output. One sentence per line. Plain hyphens, never em dashes. No roadmap, no feature list, no marketing voice. (This rule governs the drafted VISION.md only, not this skill file or your chat replies.)

Pipeline

Step 0 - Parse target and author

• Target repo: current working directory by default, or an explicit owner/repo.
• Author: the person whose vision this is; default to the repo owner. Their merged work is the evidence base.
• Ask one short question via AskUserQuestion if the target or author is genuinely ambiguous. Do not ask if the default is obvious.

Step 0b - Claim a work surface [mutating skill]

This skill writes VISION.md and an answers file. Where they land depends on the target resolved in Step 0:

• Target is the current repo (the default): the launcher invariant applies — never mutate the shared checkout in local CLI execution.

  eval "$(python3 "<skill-base-dir>/../worktree/scripts/worktree.py" setup vision)"
  cd "$WORKTREE_PATH"
  python3 "<skill-base-dir>/../shared/checkout_policy.py" require-mutation
  

  In cloud-harness environments both calls short-circuit to success — the container already provides isolation. Full rules: docs/guides/worktree-management.md in the consumer project.

• Target is an external owner/repo: the write surface is a clone of the target, never this repo's checkout or worktree. Clone the target, work there, and skip the worktree/guard calls above — they govern this repo only. If the target clone is read-only (no push access), still draft there and deliver the files to the author instead of pushing.

Everywhere below, "the work surface" means whichever of these Step 0b resolved.

Step 1 - Learn the pattern

A VISION.md has a stable anatomy; hold the draft to it:

• Identity opener: "X exists so that ...", who it serves, and "It owns exactly one thing: ...".
• 3-6 principle sections with short declarative headings, each a set of testable present-tense commitments and refusals.
• Explicit non-goals, named concretely ("it is not a CI system, not a ...").
• A closing pair of tests: "A change aligns when ..." and "A change should be resisted when ...", concrete enough to apply to a real PR.
• Voice: declarative, present tense, zero marketing; length 40-70 lines.

If the author names exemplar visions, read them; note shape, voice, length.

Step 2 - Existing-vision check

• If the default branch has a VISION.md: delta mode (hard rule 2). Diff its age against the history and propose only evidence-backed candidate additions or edits, as a numbered list, each independently acceptable.
• If not: from-scratch mode.

Step 3 - Mine the evidence

Work down this ladder. Tier A is the primary source in this repo — a merged PR title tells you what was built; an archived proposal with its rejected alternatives tells you what was chosen over what, which is the raw material of an acceptance policy. Skip a tier only when its artifacts do not exist.

If no tier yields readable history, stop and say so. Never fabricate the author's values, PR titles, proposals, or evidence. A vision built on invented evidence is worse than no vision.

Tier A - Decision artifacts (OpenSpec repos)

Read broadly, then read deep: scan 30-60 archived proposal titles, then read 8-15 full proposal.md + design.md pairs spread across the date range.

Tier B - Merged pull requests

gh pr list --author <owner> --state merged --limit 100 --json number,title,body,closedAt

One call returns titles and bodies together — scan 30-100 titles, then read 8-15 full bodies spread across the range from the same payload; no per-PR gh pr view round trips.

Tier C - Commit history

git log --author=<owner> --no-merges --format='%h %ad %s%n%b' --date=short

Titles and messages still reveal what the author builds. This repo uses conventional commits, so feat(scope): prefixes cluster the work by capability for free.

Output of Step 3

• Extract recurring revealed values: what gets built, what gets refused, what class of bug gets fixed at the root, what the author writes in intent statements, and what got reversed (superseded ADRs, rejected alternatives, reverted changes).
• Produce a private evidence sheet: value -> supporting specs, proposals, ADR entries, PRs, commits, or files. This sheet is the source of truth for every drafted line.

Step 4 - Draft

• Follow the Step 1 anatomy and the output template below.
• Every line must map to the evidence sheet. Length target: 40-70 lines.
• In delta mode, keep the baseline untouched and emit the numbered candidate list Step 2 defines.

Step 5 - Design the hypotheticals

• 8-12 concrete change proposals per vision, aimed at the draft's fault lines. Draw from this taxonomy:
  • tempting-but-off-mission features the author will plausibly be asked for;
  • principle collisions (simplicity vs capability, safety vs speed, generality vs focus, cost vs quality);
  • slippery slopes, where one reasonable step normalizes the next;
  • scope expansions (new users, new content types, new hosts, teams);
  • identity questions the draft leaves open.
• Format per hypothetical: id, title, the concrete proposal (2-4 sentences), the principle it tests (quote the draft), and why the answer is non-obvious (steelman both sides).
• Quality gate: delete and replace any hypothetical whose answer you can predict.

In an OpenSpec repo, the archive is a hypothetical generator. A proposal that was accepted narrowly, or a design.md whose rejected alternative still looks defensible, marks a fault line the author has already stood on once.

Step 6 - Review loop

The board is the reading surface; AskUserQuestion is the verdict channel. Both are required: the board carries the full draft and both steelmen, which a question prompt cannot hold; the question tool carries the verdict back to you, which a static file cannot.

6a. Build the board

Copy <skill-base-dir>/assets/review-template.html and assets/review.css next to each other on the work surface, then fill only the template's marked slots: {{PROJECT}}, {{RUN_NOTE}}, {{DRAFT_MARKDOWN}} (the full latest VISION.md text as one JSON string literal — JSON.stringify it, replacing the quoted placeholder whole; raw splicing breaks on the backticks the output template mandates), and the CARDS array — one object per hypothetical: { id, title, body, tests, why }, every field plain text (the board escapes them at render time).

Change nothing else. The template already carries the house structure: full draft on the left, one card at a time on the right, the steelman in full view. No boilerplate gets rewritten and no run gets restyled.

6b. Hand the board to the author

Write the board to VISION-review.html on the work surface. Do not launch a server, install a package, or route the board through an external service. How the author reaches it depends on where this session runs:

• Local session: give the author the path; the file opens in a browser directly.
• Remote session (cloud harness, container): a container-local path is unreachable from the author's browser, so a bare path is never the handoff. Deliver the file through the host's file-delivery mechanism (send/attach it for inline rendering), or commit it to the working branch and hand the author the hosted file link.

The board's own verdict buttons record a local ledger the author can read back — treat that ledger as a convenience, never as the channel of record.

6c. Collect verdicts

Ask through AskUserQuestion, in batches of 2-3 hypotheticals, one question per card:

• question: the card's title plus a one-line restatement of the proposal.
• header: the card id (H-7).
• options: In vision / Off mission / Conditional, each description naming the concrete consequence for the draft if chosen.
• multiSelect: false.

Include one open-ended reasoning question in the same call — the reasoning, not the verdict, is what the vision gets rebuilt from. The tool takes at most 4 questions per call, which is why a batch is 2-3 cards, never 4: the reasoning slot must always fit. Authors can also attach reasoning per card via each question's free-form "Other" path; harvest both.

If AskUserQuestion is unavailable in the current runtime, present the batch as a numbered list in regular output and ask the author to respond inline, matching the fallback convention plan-feature uses.

6d. Fold in and reply

On each batch: record the verdicts verbatim in a durable answers file (VISION-answers.md), distill the principles they reveal, fold every verdict into the draft, rebuild the board in place with the new draft text and remaining cards, and reply with a changelog line per verdict:

H-7 Off mission -> authority section now opens with "the author approves; the tool never does"

Continue until the author approves or ends the session. Do not approve on their behalf; do not treat silence as approval.

Step 7 - Finish

• Deliver: the approved VISION.md text (or approved delta), the full hypothetical set with recorded verdicts and reasoning, and the changelog.
• The answers file is durable calibration material; keep VISION-answers.md next to the vision and commit both.
• Commit and push per the consumer project's docs/guides/git-conventions.md; the work is not complete until git push succeeds. On a read-only external target (Step 0b), delivery of the files to the author replaces the push — say which happened.

Output template (from-scratch mode)

# Vision

`{project}` exists so that {the one-sentence reason the project exists}.
It serves {the named user}, and it {what it turns their input into}.
It owns exactly one thing: {the single owned surface}.

## {Principle section, 3-6 of these}

{Declarative, testable, present-tense lines; one sentence per line.}
{Explicit boundaries: what is welcome, what is refused, and why.}

## Scope

{What this project is not, named concretely.}
{Where personal/private material stays, if applicable.}
{How the repo holds itself to its own standard, if applicable.}

A change aligns when {testable positive criteria}.
A change should be resisted when {testable negative criteria}.

Pre-flight checklist (before drafting)

• Target repo and author resolved
• Work surface claimed (worktree setup + mutation guard, or cloud short-circuit)
• Existing VISION.md checked (mode chosen)
• Evidence sheet built from real artifacts at the highest available tier (no invented evidence)

Pre-approval checklist (before the author signs off)

• Every drafted line traces to the evidence sheet or a recorded verdict
• 8-12 hypotheticals, none predictable, both sides steelmanned
• Every author verdict folded in with a traced changelog line
• Answers file saved next to the vision

Common Rationalizations

Red Flags

• A drafted principle has no entry in the evidence sheet, or its entry cites the draft itself.
• The vision contains a roadmap, a feature list, or the word "world-class".
• Every hypothetical resolved In vision — the fault lines were softballs (rule 5).
• VISION.md changed but VISION-answers.md did not, or a changelog line names no verdict id.
• The skill reported an approved vision without a recorded author verdict on approval.
• The board was restyled, restructured, or replaced with a different review surface.
• Evidence citations are all from one month of history, or all from one capability.

Verification

1. Every line of the delivered VISION.md maps to a named entry in the evidence sheet (spec requirement, archived proposal, ADR entry, PR, commit, or recorded verdict).
2. The hypothetical set numbers 8-12, and each carries a quoted draft principle plus a two-sided steelman.
3. VISION-answers.md exists beside VISION.md, and every verdict in it has a matching changelog line naming the edit it produced.
4. The board file was generated from assets/review-template.html with only the four slots filled — diff against the shipped template shows changes confined to {{PROJECT}}, {{RUN_NOTE}}, {{DRAFT_MARKDOWN}}, and CARDS.
5. The author's approval is recorded explicitly; no step inferred it from silence.