atelier-kit

name	atelier-kit
description	Opt-in research → plan → implement workflow for non-trivial changes, with a verifiable plan contract. Use when the user asks to "use Atelier", invokes /atelier, or continues a `.atelier/work/<slug>.md` file. Routes through research, design, and planning phases recorded in one Markdown work file.

Atelier Kit

Skill-first workflow that teaches the agent to research, decide, plan, implement, and review a non-trivial change through one Markdown file per task — .atelier/work/<slug>.md. It is opt-in and never blocks: it teaches the agent to work better, it does not gate execution.

Activation

Treat Atelier as off unless the user asks for it ("use Atelier", /atelier, or continuing an existing .atelier/work/<slug>.md). When off, behave natively.

Modes

Pick a mode by the task's weight — more process only where it pays off:

• quick — small, local, low-risk. Understand → change → validate. One file, no contract.
• standard — multiple files or real logic. Research → plan (slices) → implement → review.
• deep — architecture, data, security, billing, migration, performance. Add design + risks.

Phases (progressive disclosure)

Each phase has detailed instructions in a bundled reference file. Read the file for the phase you are entering — do not load them all up front.

Phase	When	Reference
Research	standard/deep (quick usually skips)	references/researcher.md
Design	standard/deep, when there is a real choice	references/designer.md
Plan + Review	standard/deep	references/planner.md

Instructions

1. Confirm the user invoked Atelier; if not, behave natively and stop here.
2. Choose the mode (quick / standard / deep) from the task's weight.
3. Create the work file: atelier new "<title>" --mode <mode>, or by hand from templates/work.md.
4. Research — read references/researcher.md, then fill ## Objective, ## Questions (the four buckets — blocking user questions, repository research, external research, safe assumptions), and ## Research.
5. Design (standard/deep with a real choice) — read references/designer.md, then fill ## Decisions.
6. Plan — read references/planner.md, then write ## Plan slices (and ## Risks in deep mode).
7. Run atelier validate and fix every error before implementing.
8. Implement natively, logging progress under ## Implementation and checks under ## Validation.
9. Review — run atelier review, then follow the review steps in references/planner.md to finalize ## Review.
10. At each phase hand-off (after research, design, plan, and review), if command -v plannotator succeeds, run plannotator annotate .atelier/work/<slug>.md and fold its notes back in before advancing.
11. In quick mode, skip the slices and the gate: understand → change → validate directly.

Plannotator (optional)

When Plannotator is installed, use it as a review surface before each hand-off: if command -v plannotator succeeds, run plannotator annotate .atelier/work/<slug>.md and fold the notes back into the work file. There is no separate Atelier command for this — it is opt-in and only runs when Plannotator is present. The phase references repeat this so it survives progressive disclosure.

Forbidden

• Do not gate execution — Atelier teaches, it does not block.
• During the research, design, and planning phases, do not edit project code; implement only after the plan is validated.