Vision

/vision

Create or update root VISION.md: the compact project north star that gives cold agents, maintainers, and future contributors the same answer to "what are we building, why, and what kind of excellence matters here?"

Conversational first. Artifact always.

Contract

• Ask the operator what the deal is unless the request already contains enough project intent. One good question at a time; use the interrogate-first lens (harnesses/shared/references/interrogate-first.md) when hidden decisions matter.
• Read the live repo before drafting: existing VISION.md, AGENTS.md, README*, positioning/product docs, roadmap/backlog, manifests, examples, demos, tests, screenshots, and repo-local skills or prompts that encode product behavior.
• Research just enough. Start local, then sibling/adjacent projects when they explain fit, then web or external exemplars only when category, audience, competition, or public-facing positioning is unclear.
• Clarify lifespan. Is this a spike, internal utility, consulting artifact, product substrate, long-lived product, or public standard? The answer changes tone, maintenance bar, backlog ambition, and non-goals.
• Write the root artifact at VISION.md. Do not tuck the canonical artifact in docs/ unless repo evidence proves root is wrong.
• Let the project determine the structure. No required headings, no house template. Keep it tight enough that agents will actually read it.
• Capture what is uniquely load-bearing: intent, philosophy, audience, category, fundamentals, standards, non-goals, strategic bets, and what excellent looks like over the horizons the project needs.
• Wire consumers, not copies. When VISION.md is created or materially changed, update relevant repo-local AGENTS.md, skills, prompts, or runbooks with pointer lines to VISION.md; never duplicate vision prose into them.

Quality Check

Load references/vision-quality.md when drafting, reviewing, or updating the artifact. Use it as a taste/checklist reference, not a template.

Output

Finish with:

1. What changed in VISION.md.
2. What questions were answered, deferred, or still need the operator.
3. Sources read: local files, sibling projects, external references, and what each changed.
4. Consumers wired: AGENTS.md, skills, prompts, runbooks, or explicit none.
5. Residual risk: stale assumptions, missing competitive context, or unclear lifespan.

Verification

For repo edits:

test -f VISION.md
rg -n "VISION\\.md" AGENTS.md .agents .codex .claude .pi .antigravitycli skills 2>/dev/null

Then run the repo's named gate. In Harness Kit itself, that is:

cargo run --locked -p harness-kit-checks -- check --repo .

Gotchas

• Generic mission statement. If it could describe three other repos, it failed.
• Over-prescribed structure. A beautiful template can erase the project's real shape.
• Backlog dump. Vision decides what belongs; it is not a sorted task list.
• Marketing voice by default. Use marketing language only when that is the repo's real surface.
• docs/vision.md drift. Root VISION.md is the default canonical path.
• Stale consumers. If local harness primitives keep making direction calls from old prose, the vision is decorative.
• Research theater. A giant competitive map is waste unless it changes a decision.