samsantosb/ship-it/create-product

Skill: create-product

Entry point of the pipeline. Anyone — founder, dev, designer, PM, whoever — brings raw material and leaves with a product document that /create-tasks can split into technical cards. The pipeline makes no assumption about who writes product scope or where: the document is produced by interview, not by role. Respond in the conversation language (docs/agents/language.md; unset → the user's language).

Load before acting: ${CLAUDE_PLUGIN_ROOT}/shared/workflow-context.md, ${CLAUDE_PLUGIN_ROOT}/shared/sparring.md, and ${CLAUDE_PLUGIN_ROOT}/shared/guardrails.md — this skill writes product docs and may publish tracker items; the loading contract applies.

Accepted input: free text in the argument, a path to a rough doc, "this conversation", or nothing at all (start from zero).

Process

1 — Gather context and resolve roles

Collect the raw material: the argument, the referenced file, or the current conversation. If a repo is connected, read related docs and code before asking anything. List what the material already answers — those questions are skipped.

Resolve the Tracker role now, per ${CLAUDE_PLUGIN_ROOT}/shared/workflow-context.md (batched question if unresolved) — config questions happen before the sparring session starts, never in the middle of it.

2 — Sparring session

Run ${CLAUDE_PLUGIN_ROOT}/shared/sparring.md over the idea. Minimum topic checklist: the problem, who it's for, measurable success criteria, scope in/out, constraints, and the riskiest assumption. Depth is proportional to ambiguity — a well-written draft might need 3 questions; a one-liner might need 20. Stop when the sparring completion criteria hold: you could sketch the solution and a QA could write a test plan without guessing.

3 — Synthesize the document

Fixed template. Section headers are fixed English tokens (stable for /create-tasks to parse); the content inside them is written in the user's language. Product-level only — technical detail (files, tests, stacks) belongs to /create-tasks:

# <Product / feature name>
 
## Problem            — what hurts today, for whom, evidence if any
## Who it's for       — the concrete user/persona
## Success criteria   — measurable; how we know it worked
## Scope
### In
### Out               — explicit non-goals
## Acceptance criteria — high-level, verifiable (Gherkin-ish, no implementation detail)
## Constraints & dependencies
## Open questions      — table: question · owner · status

• Owner in Open questions = the role or person who must decide this before implementation (e.g. "tech lead — stack choice", not "TBD").
• Never fabricate an answer the user didn't give — unresolved topics go to Open questions, visibly.

4 — Publish

• Write the document to docs/product/<slug>.md in the project directory (works with or without git; create the folders if missing). One product doc per feature/product — no PRODUCT.md at root; that name is only honored as a legacy input by /create-tasks.
• Tracker MCP / GitHub Issues mode: additionally offer to publish a product item linking the doc, so the scope is visible where the team tracks work.
• Never overwrite an existing product doc silently — show what would change and confirm.
• Fold everything the user must approve — overwrite (if any), tracker publish, and any pending config write-back — into one final confirmation, not a chain of prompts.

5 — Handoff

Return the doc path (and item URL if published) plus the next step: /create-tasks <path-or-url>.

Guardrails (this skill, on top of shared/guardrails.md)

• Product document ≠ technical card: no file:line, no describe/it, no stack decisions here.
• Sparring answers are the user's decisions — record them faithfully; disagree during the interview, not in the document.
• AI-disclosure on anything published to a tracker — rule owned by ${CLAUDE_PLUGIN_ROOT}/shared/guardrails.md.