• Prefer merged-PR history via a GitHub-class CLI (gh, gh-axi).
• If PRs are not accessible, fall back to git commit history on the default branch (git log): titles and messages still reveal what the author builds.
• Only if no real history is readable at all, stop and say so. Never fabricate the author's values, PR titles, or evidence. A vision built on invented evidence is worse than no vision.

npx -y lavish-axi

1. Evidence over vibes. Every principle in the draft must be traceable to concrete evidence: named PRs or commits, files, docs, 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 loop runs on lavish-axi, from the shipped template. Draft and hypotheticals are presented as one board built from

   assets/review-template.html

   +

   assets/review.css

   , used as-is: black ink on white paper set like literature, full draft always fully visible, one hypothetical at a time in a card stack. Fill the template's slots; never restyle or restructure it, and never substitute another review surface.
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. One sentence per line. Plain hyphens, never em dashes. No roadmap, no feature list, no marketing voice.

• 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 if the target or author is genuinely ambiguous.

• 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 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, each independently acceptable.
• If not: from-scratch mode.

• Repo analysis: README identity claims, architecture, stated non-goals, refusal paths, test discipline.
• History mining: list the author's merged PRs, aim for 30-100 titles, and read 8-15 full bodies spread across the range (for example

  gh pr list --author <owner> --state merged --limit 100

  , or the gh-axi equivalent). If PRs are inaccessible, walk default-branch commit history instead (

  git log --author=<owner>

  ), reading messages for the same signal.
• 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.
• Produce a private evidence sheet: value -> supporting PRs, commits, or files. This sheet is the source of truth for every drafted line.

• Follow the step 1 anatomy and the output template below.
• Every line must map to the evidence sheet. Length target: 40-70 lines.
• Delta mode instead yields: baseline unchanged + a numbered list of candidate line additions/edits, each with its evidence.

• 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.

• Copy

  assets/review-template.html

  and

  assets/review.css

  next to each other, then fill only the template's marked slots: project name, run note, the full DRAFT markdown, and the CARDS array (id, title, proposal, tested principle, both-sides steelman per card).
• 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, one queued verdict per card - so no boilerplate is rewritten and no run is restyled.
• Launch with

  npx -y lavish-axi <board.html>

  , report the URL, then wait on

  npx -y lavish-axi poll <board.html>

  ; answers arrive as queued verdicts.
• On each batch: record the verdicts verbatim in a durable answers file, distill the principles they reveal, fold every verdict into the draft, update the board in place (new draft text, remaining cards), and reply through

  poll --agent-reply

  with a changelog line per verdict ("H-7 no -> authority section now opens with ...").
• Continue until the author approves or ends the session. Do not approve on their behalf; do not treat silence as approval.

• 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; tell the author to keep it next to the vision.

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}.

• Target repo and author resolved
• Existing VISION.md checked (mode chosen)
• Evidence sheet built from real PRs or commits (no invented evidence)

• 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