| name | palette-init |
| description | Opt a project into palette (the product-intent outer loop). Scaffolds the _palette/ backlog through a light intake and asks about git handling. Use when the user wants to start managing a project or roadmap with palette, or accepts the offer to set palette up. Does not scope a phase — the palette loop does that afterward. |
palette-init — opt a project into palette
This skill turns palette on for a project: it scaffolds _palette/ and seeds the backlog through a light intake. It does not scope a phase and writes no source code. After init, the palette loop (defined in the always-loaded claude-agent-kit--palette.md rule) takes over automatically, because _palette/ now exists.
Prefix messages with [palette-init]:.
When this runs
The user invoked palette-init, or accepted the one-line offer to set palette up. If _palette/ already exists, do not re-scaffold — report that the project is already palette-active, point at _palette/backlog.rst, and stop.
Step 1 — Light intake
Keep it short: a few questions, not an interrogation. Ask only what changes what gets built.
- What are you building, in one or two sentences?
- Who is it for?
- What platform? (web / mobile / desktop / CLI / library)
Then, only if not already clear: what is the core pain, or why is this worth building? Push back on vague terms ("simple", "basic", "standard") — ask what they concretely mean.
Intake posture — keep it light, but hold these lines:
- Ask what the product is, not how it is built. Do not ask about or decide stack, storage, auth, or architecture. If the user volunteers a technical directive ("use Postgres", "must be offline-first"), transcribe it verbatim into the relevant backlog item's context — do not act on it or evaluate it here.
- Do not propose architecture, libraries, or a decomposition.
- Scope from the user's answers, not by reading the codebase to enumerate work.
Step 2 — Seed the backlog
Decompose what you learned into backlog items. Use the _palette/backlog.rst schema and the RST house style from the always-loaded palette rule (claude-agent-kit--palette.md) — do not restate the schema here, follow it. Each discussed feature, requirement, or constraint becomes an item with :Status: backlog. Capture genuine durable product principles (product feel, interaction expectations, simplicity / performance / accessibility posture) in the Core Product Principles section — 3–7 max, never invented to fill space.
Create the _palette/ directory and write _palette/backlog.rst. Do not create phase folders, stories, or any optional artifact — those come later in the loop.
Step 3 — git handling (ask)
_palette/ defaults to not committed (see the palette rule). Ask the user:
Add a _palette/.gitignore so palette stays out of git? It is your personal planning record by default. Say no if you want to commit it — e.g. to share across a team.
- Yes → write
_palette/.gitignore containing a single line: *
- No → write nothing; leave
_palette/ untracked for the developer to handle.
Do not commit anything yourself either way.
Step 4 — Hand off to the loop
Confirm what was created, then hand off. Do not scope a phase here — the loop (now active because _palette/ exists) owns slicing the first thin phase with the user, when they are ready.
[palette-init]: palette is set up.
- _palette/backlog.rst — <N> items seeded
- git: <.gitignore added | left untracked>
Next: when you're ready we'll slice the first thin phase from the backlog. Tell me what you want to tackle first, or keep adding to the backlog.
Forbidden
- Do not write or fix source code — palette-init only scaffolds
_palette/.
- Do not scope or draft a phase, create stories, or create any optional artifact (tech-spec / ux-flow / design-brief / project-rules) — those belong to the loop and the pull-only
palette-* skills.
- Do not ask about or decide tech / architecture; transcribe volunteered directives verbatim, do not act on them.
- Do not commit
_palette/ to git.
