|
2 | 2 |
|
3 | 3 | Never make assumptions about what the user wants. Always ask clarifying questions before proceeding. |
4 | 4 |
|
| 5 | +# MANDATORY: Feature Planning & Implementation Workflow |
| 6 | + |
| 7 | +**This workflow is non-negotiable for any non-trivial feature or multi-step change.** Do not implement multi-task features directly in a single session. Do not write a monolithic plan and start coding from it. The workflow below exists because implementation plans that live in a single file are either too large for a context window or too shallow for independent execution — both lead to poor results. |
| 8 | + |
| 9 | +### The workflow has three stages. Execute them in order, every time. |
| 10 | + |
| 11 | +**Stage 1 — Plan.** Discuss the feature with the user. Ask clarifying questions. Understand requirements, technical constraints, and acceptance criteria. This happens naturally in planning mode. |
| 12 | + |
| 13 | +**Stage 2 — Create the spec.** Once planning is complete, invoke the `create-spec` skill. This converts the planning conversation into a structured `specs/{feature}/` folder with: |
| 14 | +- Self-contained task files (one per task, in a `tasks/` subfolder) |
| 15 | +- A README with a dependency graph and parallel execution waves |
| 16 | +- Requirements and manual action items |
| 17 | + |
| 18 | +Do not skip this step. Do not proceed to implementation without a spec folder. The spec is what enables parallel agent execution — without it, you're back to single-threaded implementation. |
| 19 | + |
| 20 | +**Stage 3 — Implement via orchestration.** Once the spec folder exists and the user approves it, invoke the `implement-feature` skill. This orchestrates parallel coder agents wave-by-wave: |
| 21 | +- Each task is dispatched to its own coder agent |
| 22 | +- Tasks within a wave run in parallel |
| 23 | +- A code review gate runs between waves |
| 24 | +- Issues from review are dispatched back to coder agents |
| 25 | +- The cycle repeats until everything passes |
| 26 | + |
| 27 | +The main agent (you) does NOT write code during Stage 3. You orchestrate. Coder subagents implement. The code-review agent verifies. You manage progress and commit completed waves. |
| 28 | + |
| 29 | +### When does this workflow apply? |
| 30 | + |
| 31 | +- Any feature with 2+ tasks or files to change |
| 32 | +- Any work that came out of a planning conversation |
| 33 | +- Any time the user says "implement", "build", "create this feature", or similar after discussing what to build |
| 34 | + |
| 35 | +### When does it NOT apply? |
| 36 | + |
| 37 | +- Single-file bug fixes or trivial changes |
| 38 | +- Quick config edits or one-line patches |
| 39 | +- Tasks the user explicitly asks to do inline ("just fix this here") |
| 40 | + |
| 41 | +If in doubt, ask the user: "This looks like it could benefit from a spec — should I create one, or would you prefer I implement it directly?" |
| 42 | + |
| 43 | +--- |
| 44 | + |
5 | 45 | # Project Agent Instructions |
6 | 46 |
|
7 | 47 | ## Skills |
|
0 commit comments