implement-spec
/implement-spec
Implement a spec (specs/NNN-…) with the agentic pipeline - independent spec critique and impact map, tests written first and locked, implementation, deterministic checks, independent audit before push. Use when the user asks to implement a spec with /implement-spec, or to "implement spec NNN with the pipeline". Add --light for small specs (critique + audit only).
- Author
- FRogora98 ↗
- Repository
- spec-driven ↗
- Licence
- MIT
- Last repository change
- 03.11.2026
Assessment and security
- Files in this package
- 4
- Scripts
- 2
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 03.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
- 2 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: scripts/checks.py, scripts/locked_tests.py
What it does and when to use it
Implement a spec (specs/NNN-…) with the agentic pipeline - independent spec critique and impact map, tests written first and locked, implementation, deterministic checks, independent audit before push. Use when the user asks to implement a spec with /implement-spec, or to "implement spec NNN with the pipeline". Add --light for small specs (critique + audit only).
Triggers from the description
Original author description
Implement a spec (specs/NNN-…) with the agentic pipeline - independent spec critique and impact map, tests written first and locked, implementation, deterministic checks, independent audit before push. Use when the user asks to implement a spec with /implement-spec, or to "implement spec NNN with the pipeline". Add --light for small specs (critique + audit only).
How it works
- Step 0 - Project configuration
- Step 1 - Spec critique (subagent spec-critic)
- Step 2 - Tests first (subagent test-writer) (skipped with --light)
- Step 3 - Lock the tests (skipped with --light)
- Step 4 - Implement, then deterministic checks
- Step 5 - Independent audit (subagent auditor)
- Step 6 - Push and report
- Rules
SKILL.md preview
---
name: implement-spec
description: Implement a spec (specs/NNN-…) with the agentic pipeline - independent spec critique and impact map, tests written first and locked, implementation, deterministic checks, independent audit before push. Use when the user asks to implement a spec with /implement-spec, or to "implement spec NNN with the pipeline". Add --light for small specs (critique + audit only).
---
# /implement-spec <spec path> [--light]
Pipeline for implementing a spec without self-review. Who writes the tests and who audits is never the session that
implements. The rules that matter are enforced by a hook and a script, not by good will.
Arguments: the spec folder or `spec.md` (e.g. `specs/056-…`). `--light`: skip steps 2 and 3-lock (for specs touching
one or two files with no new behaviour); steps 1, 4 and 5 always run.
## Step 0 - Project configuration
Read `.claude/implement-spec.json` at the repository root. If it is missing, stop and ask the user for the values,
showing this template (do not guess them). `standard` is the path of the project's coding standard; empty means the
standard bundled with this skill, `CODING-STANDARD.md` next to this file (.NET backends, with rules for LLMs and
agents).
The scripts are in `scripts/` next to this file; below, `<skill dir>` is this skill's base directory.
```json
{
"specsDir": "specs",
"standard": "",
"testCommand": "dotnet test --nologo",
"evalCommand": "",
"evalTriggers": ["**/Prompts/**"],
"testGlobs": ["**/*Tests/**", "**/*.Tests/**"],
"docGlobs": ["docs/**", "README.md"],
"fallbackPatterns": ["\\?\\? *(\"|default|new |0\\b|false\\b|string\\.Empty)", "catch *\\{ *\\}", "catch *\\(Exception\\)[^{]*\\{ *\\}"],
"branchPrefix": "exp/"
}
```
Then: work on a branch `<branchPrefix><spec folder name>`, never on the default branch. Read the spec, the repo
`CLAUDE.md`/`AGENTS.md` and the standard.
## Step 1 - Spec critique (subagent `spec-critic`)
Launch the `spec-critic` agent with the spec path and the config. It writes `<spec folder>/impact.md`:
the impact map (every consumer of what changes, every failure path, docs and evals to touch) as a checklist,
plus open questions.
- If `impact.md` has **open questions**: stop, show them to the user, and wait. Answers go into the spec
(not only into the chat), then re-run step 1 on the changed parts.
- If a question is about money, identity or legal matters, it is always the user's.
## Step 2 - Tests first (subagent `test-writer`) *(skipped with --light)*
Launch the `test-writer` agent with the spec, `impact.md` and the config. It writes tests for every acceptance
criterion and every failure path in the impact map, in the style of the existing tests, only under `testGlobs`.
Red gate: run the new tests. Every new test must fail (or not compile because the code does not exist yet)
for the reason the test describes. A test that is already green tests nothing new: send it back to `test-writer`.
## Step 3 - Lock the tests *(skipped with --light)*
Commit the tests alone (`test: <spec> acceptance tests (red)`), then lock them:
```
python "<skill dir>/scripts/checks.py" lock <test file> [<test file> …]
```
This writes `.locked-tests` at the repo root with the lock commit. From now on the `locked_tests` hook rejects edits
to those files (the plugin registers it). If a locked test is wrong, do not work around it: stop and ask the user. Only the user unlocks
(`python "<skill dir>/scripts/checks.py" unlock`).
## Step 4 - Implement, then deterministic checks
Implement until all tests are green. Follow the standard and the repo rules; docs in the same change.
Tick every item of `impact.md` with the file that covers it (`- [x] … → path:line`), or explain why it does not apply.
Then run:
```
python "<skill dir>/scripts/checks.py" run <spec folder>
```
It writes `<spec folder>/checks.md` and fails if: a locked test changed since the lock commit, the test command
fails, a prompt changed and the eval command did not pass, `impact.md` has unticked items, or code changed with no
doc change. Silent-fallback patterns in the added lines are reported as warnings for the audit. Fix until it passes.
## Step 5 - Independent audit (subagent `auditor`)
Commit the implementation on the branch (not pushed). Launch the `auditor` agent with: the spec, `impact.md`,
`checks.md`, the standard and the commit range. It is read-only and returns findings ordered by severity with
file:line.
- Blocker/high findings: fix them (tests first), re-run step 4, then re-run the auditor **on the new commits only**.
- At most 2 audit rounds. If a blocker/high is still open after round 2, stop and escalate to the user with the list.
- Medium/low: fix what is cheap, list the rest in the report.
## Step 6 - Push and report
Push the branch (never the default branch, never force). Final report to the user, short:
spec, branch and commit, impact map (n items), tests (n new, red → green), checks result, audit rounds and remaining
findings, decisions taken, open points. Never merge: merging is the user's decision.
## Rules
- Never edit a locked test, `.locked-tests` or `checks.md` by hand.
- Never skip a step silently: if a step does not apply, say why in the report.
- The spec is the contract: a behaviour not in the spec is a question for the user, not a free choice.
Files in this package (4)
- CODING-STANDARD.md 23 KB
- SKILL.md 5 KB
- scripts/checks.py 7 KB
- scripts/locked_tests.py 3 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · FRogora98/spec-driven | 100 | MIT |
| cravingmaker/skills | 100 | MIT |
| OutlineDriven/outline-driven-development | 100 | Apache-2.0 |
| gosuda/dalcom-skills | 95 | Apache-2.0 |
| dzhng/skills | 90 | MIT |
| dzhng/jevgrep | 90 | MIT |
| dzhng/skills | 90 | MIT |
| aomni-com/open-duet | 90 | MIT |
| FradSer/dotclaude | 86 | MIT |
Database snapshot: 13.11.2026 · package version c89587ee9beb (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100spec-write
Start spec-driven work on Device by writing specs/NNN-slug/spec.md (problem, behaviour, requirements, acceptance checks, out of scope, open questions). Use when someone asks for a new feature, a behaviour change, cross-layer or performance work, or a format or C ABI change,… - 90implement-spec-with-codex
Run [implement-spec](../implement-spec/SKILL.md) with Codex doing the implementation passes while you orchestrate, integrate, and review. Use when the user asks to implement a spec with Codex, or to run implement-spec with Codex as the implementer. - 100om-implement-spec
Implement a specification (or specific phases) using coordinated subagents with unit tests, integration tests, docs, and code-review compliance. Tracks progress by updating the spec. Triggers on "implement spec", "implement phases", "build from spec", "code the spec". - 98tdd-spec
Test-first workflow with locked, number-based spec tests for any repo with a .claude/tdd/ pack. Use when starting any new change, feature or bug fix in such a repo; when asked to write spec tests, lock tests, implement against locked tests, raise a dispute about a test, or… - 90om-auto-implement-spec
Implement an existing spec and ship a verified, reviewed, ready PR. Resolves the spec by path, name, issue, or spec-PR number (clean stop with candidates when not found). A spec PR stays design-only — implementation ships on its own PR referencing it. Delegates to… - 85spec-implement
按 docs/specs 下的一份 spec 逐条 AC 跑红绿循环,并把测试名、状态、失败证据回填进 spec。当 spec 已是 ready 或 MVP 状态、要落成代码时使用。