Menu
Full development cycle — plan a feature, implement it, and open a PR
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Commit remaining changes, push the branch, open a pull request, and run Codex review
Process unhandled PR review comments on the current branch's open PR — answer questions, modify code, or push back
Like /develop, but skips test-writer invocation and post-impl verification (for repos without a test suite — see ADR-014 §12)
Run deep-learning experiments as disciplined hypothesis tests — frame the question, set up a self-contained per-experiment folder, launch training/eval runs (confirming before heavy GPU jobs), track metrics, analyze results against a baseline, and write a human-readable report with tables and plots. Use this whenever the user is doing experimental ML work: launching or preparing a training/finetuning/sampling/eval run, an ablation, or a hyperparameter sweep; saying things like "let's try X and see if it helps," "does this change improve FID/accuracy/loss," "compare these two runs/checkpoints," "track this experiment," "analyze the results," or "write up what we found." Trigger even when the user doesn't say the word "experiment" but is clearly testing whether a change moves a metric, or wants results organized, compared, or reported reproducibly.
Post-merge cleanup — update main, prune the remote, and print one paste-ready block of follow-up commands
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
| name | develop |
| description | Full development cycle — plan a feature, implement it, and open a PR |
| allowed-tools | Agent, Bash, Read, Edit, Write, Glob, Grep |
| argument-hint | [--no-test] <feature description> |
You are running the full development cycle: planning, implementation, and PR creation.
The feature to develop: $ARGUMENTS
Operating principle: Only ask the user when a step explicitly requires it, or when it is strongly necessary (e.g., blocking errors, ambiguous requirements). Do not stop just to report progress — keep moving through the steps autonomously.
Run /start-dev $ARGUMENTS to kick off planning.
Wait for /start-dev to complete fully (plan approved, issues created, branch ready, plan committed). Do not proceed to Phase 2 until /start-dev has finished all its steps.
Note the planning doc path — you will need it in Phase 2.
Read the planning doc from docs/plans/ (created in Phase 1).
Multi-issue scope rule: If /start-dev created multiple sub-issues, only implement tasks belonging to the current sub-issue (the one linked to this branch). Do not work on tasks for other sub-issues.
Flag check. --no-test is recognized only when it is the first whitespace-delimited token in $ARGUMENTS (e.g., /develop --no-test add foo handler). A literal --no-test later in the feature description (e.g., /develop add --no-test option to bar) is part of the description, not a flag. When the leading-token flag is present, set the no-test mode; skip test-writer invocation and post-impl verification per ADR-014 §12. Spec authoring is NOT skipped. See Phase 2.NT below.
Read the spec-bearing record. Look for the header line matching ^\*\*Spec-bearing:\*\*\s+(yes|no)\b near the top of the planning doc. Cross-check spec file presence: if the header says yes, the spec file referenced in the header MUST exist at docs/specs/NNN-<slug>.md; if the header says no, no spec file should exist for this feature. If the header is missing/malformed, OR header and spec-file presence disagree, halt with plan/spec out of sync — header says <X> but spec file <exists|missing> and ask the user.
The test-writer reads only the spec and the AST manifest. Any planner notes or reviewer comments forwarded into its prompt MUST be paraphrased first:
bot/, .claude/hooks/, .githooks/, scripts/; line numbers; identifier names that do not appear in the spec.bot/cogs/foo.py:42" does not.docs: spec edit, then re-run sanitization.No regex-based sanitizer is mandated — paraphrasing is done by the orchestrator inline. Verbatim reviewer text is never forwarded.
Resolve target paths. Read the planning doc's task list to determine the impl module path(s). A feature may touch multiple impl files — collect all. Pick test paths by mirroring each under tests/:
tools/foo.py → tests/tools/test_foo.pyscripts/foo.py → tests/scripts/test_foo.pyFor multi-file features, a single spec covers the whole unit and a single test-writer invocation generates tests across all target test paths.
Extract the API manifest. Run:
uv run python scripts/extract_api.py <target-paths>
Capture stdout.
Paraphrase planner claims into spec vocabulary. Apply the Sanitization rules above before forwarding any planner/reviewer text to the test-writer.
Invoke the test-writer subagent. Use the Agent tool with subagent_type: test-writer. Build the prompt from this template:
## Spec
{full contents of docs/specs/NNN-<slug>.md, inlined verbatim}
## Target language
Python (pytest)
## API manifest
{stdout of `scripts/extract_api.py <target-paths>`, inlined verbatim}
## Target test paths
{test paths chosen in step 1}
## Constraints
- Do NOT Read implementation files under `bot/`, `.claude/hooks/`, `.githooks/`, or `scripts/`.
- Do NOT Read existing tests for the same target unless explicitly named below as spec-derived.
- Every new test MUST fail on first run. If any new test passes, halt and report — do not commit.
- Map tests to spec labels (`test_b1_…`, `test_e1_…`, `test_er1_…`, `test_i1_…`).
## Spec-derived existing tests for the same target
{paths of existing tests explicitly named as spec-derived; empty list if none}
## Paraphrased planner notes
{paraphrased claims from Phase 1 / planning doc — already stripped of impl-leak references}
Verify failing-test deliverable. Read the agent's reply.
Commit failing tests: git commit -m "test: add failing tests from spec NNN-<slug>". Stage only the new test files (not the whole tests/ tree).
Implement. Now write the implementation against the failing tests. Standard TDD: red → green incrementally. You may Read the test files.
Post-impl verification.
uv run pytest <test-path> --tb=short — all-green required.Commit impl: git commit -m "<feat|fix>: <description>". Use feat: for new behavior, fix: for bug fixes.
Commit ordering invariant. The branch's commit log for this feature MUST show: docs: spec → test: failing tests → <feat|fix>: impl. Three separate commits. If tempted to squash locally, don't — the order is auditable evidence the spec-driven flow was followed.
Skip the test-writer flow entirely. Implement the prose change. No tests required per ADR-014 §6 / §8. Proceed straight to Phase 3.
--no-test mode (spec-bearing feature without new-test authoring)When $ARGUMENTS starts with --no-test:
--no-test means don't author new tests, not don't run tests. If the repo has no test suite at all (e.g., downstream DL experiment repo), the test command finds nothing and exits clean — that's fine.Commit order collapses to two: docs: spec → <feat|fix>: impl.
See ADR-014 §12 for the portability rationale.
Once all in-scope tasks are checked off, proceed to Phase 3.
Run /create-pr to commit any remaining changes, run code review, push, and open the PR.
Wait for /create-pr to complete fully.
If /start-dev created multiple sub-issues, after the PR is created, print:
/finish-task, then run /develop again targeting the next sub-issue