Product Vision

What this skill does

product-vision produces product/vision.md — the root WHY of the product. It works in three parts: it elicits the vision from human product intent; it reconciles that vision against any capabilities the project has already grown; and once the vision is approved it resolves the bootstrap vision-links left on capability stubs. It is the top of the pipeline — every capability ultimately cites the vision this skill writes.

Where the state tree lives

Generic skill — hard-codes no path. Resolve the state-tree root from the project's CLAUDE.md (project-declared paths, DESIGN_PRODUCT_PIPELINE.md §2); default docs/, and say which you used. {state-root} is that root, so the vision lives at {state-root}/product/vision.md and capability stubs at {state-root}/capabilities/{slug}/README.md.

STOP — how this skill works

Before writing vision.md you MUST, in order:

1. ELICIT INTENT. The vision is authored from the human's product intent —

what the product is for, who it serves, what it solves, what it explicitly is NOT. If you do not have that intent, ASK for it. You may not infer the vision from the codebase or from existing capability stubs.

1. RECONCILE against existing capabilities (see "Reconciliation"). Read the

capability stubs; produce a reconciliation report of where the drafted vision and the built reality disagree. Findings are flagged for a human, never silently fixed.

1. WRITE the vision (see "What vision.md must contain").
2. SEEK HUMAN APPROVAL — the vision is the highest-stakes artifact in the

pipeline; it gets an explicit human approval step.

1. RESOLVE BOOTSTRAP TOKENS, gated (see "Resolving bootstrap vision-links").

On approval, before writing status: approved, append a ## Changelog entry for this revision: the new revision number, the date, and one or two sentences on what changed and why. Bumping revision without the matching changelog entry is a defect — they are one act.

Hard rules:

• You author the vision from human intent. You never derive it from what got

built. A vision traced from the codebase can only ratify the product; it must be able to judge it.

• Reconciliation surfaces drift. It never edits the vision to fit the

capabilities, nor the capabilities to fit the vision.

• You do not approve the vision yourself. Approval is an explicit human act.

Reconciliation

A project may already have provisional capability stubs and shipped changes before this skill ever runs. Read them — every capabilities/{slug}/README.md — and hold the drafted intent-based vision next to them. Ask the tension questions:

• Is there a capability the vision, as drafted, does NOT justify? A real

finding: either the vision is incomplete, or something got built that should not have been. Flag it; a human decides which.

• Does the vision call for something NO capability is growing toward yet? A

gap between intent and reality. Flag it.

• Does a capability's description CONTRADICT the vision's framing? Flag it.

The output is the vision (from intent) PLUS a reconciliation report listing these findings. You do not resolve them — widening the vision or marking a capability misaligned is a human's call. Reconciliation is how this skill turns "write a vision" into "discover whether the built product and the intended product have drifted apart." Reading the stubs informs the CHECK; it never authors the vision.

What vision.md must contain

Write to {state-root}/product/vision.md.

Frontmatter:

---
revision: 1
status: draft
approved_by:
approved_on:
---

vision.md is the root why — it has no why-link.

Body — every section required:

• Purpose — what the product is for, in one or two plain sentences.
• Who it serves — the audience(s).
• What it solves — the problems it exists to remove.
• What it is not — explicit non-goals; the product's boundary.
• Vision goals — the named goals capabilities will cite. Each goal is a

short, stable, citable statement. These are the citation targets for every capability why-link.

• Reconciliation report — the findings from "Reconciliation": capabilities

the vision does not account for, intent no capability serves, contradictions. Each flagged for a human. If the project had no capabilities, say so.

• Changelog — per DESIGN_PRODUCT_PIPELINE.md §2b (Changelog); first entry

revision 1 — <date> — initial vision.

The vision is written from intent. It contains no capability decompositions, no technical content.

Resolving bootstrap vision-links

change-scope creates capability stubs during bootstrap with the vision-link why-link: BOOTSTRAP — no product vision yet, because no vision existed yet. Once vision.md is approved, resolve those tokens — but GATED on reconciliation:

• For a capability the reconciled vision JUSTIFIES — replace its BOOTSTRAP

token with a real citation to the specific vision goal it serves (path + goal).

• For a capability FLAGGED in the reconciliation report as unaccounted-for —

do NOT resolve its token. Leave the BOOTSTRAP marker, or escalate, until the human resolves the tension.

Resolving a BOOTSTRAP token is an assertion that the capability genuinely serves the vision. Never resolve a token by inventing a vision goal to point at — that would re-hide the very drift reconciliation just surfaced.

A re-run that changes the vision bumps revision and appends a ## Changelog entry naming the change — the same atomic act required at approval (STOP gate).

When this skill applies

• Establishing a product's vision for the first time, or materially revising

it — this skill.

• A single feature idea — NOT this skill; that is change-scope (stage 3).
• Decomposing the product into capabilities — NOT this skill; that is

capability-map (stage 2), which runs after the vision exists.

• Discussion about product direction — not a trigger; answer normally.

Precedence

vision.md is the root authority of the product pipeline — capabilities, changes, specs, and issues all ultimately cite it. A project's decision records (docs/adr/ etc., where the project keeps them) record TECHNICAL decisions and sit below this skill's concern; product-vision does not read or write them. If product intent and a technical decision record appear to conflict, that is a reconciliation finding for a human — flag it, do not resolve it here.