skill
claude-code-feature-factory
Claude Code plugin: seven specialist subagents (researcher, story-writer, spec-writer, backend/frontend builders, test-verifier, validator) chained through research, spec, build, test, and validation
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/msdtk621/claude-code-feature-factory itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026Feature Factory
A seven-agent software factory for Claude Code. Describe a feature in one sentence. Seven specialist subagents take it from research to a spec, an implementation, acceptance tests, and a final validation report — pausing at three checkpoints where your judgment actually matters.
/feature-factory add a "forgot password" flow with email reset links
codebase-researcher → story-writer → spec-writer → backend-builder
→ frontend-builder → test-verifier → implementation-validator
Install
/plugin marketplace add msdtk621/claude-code-feature-factory
/plugin install feature-factory@code-and-chord
Then run /feature-factory <what you want built> from the root of a git repository that has a CLAUDE.md.
Why seven agents instead of one prompt
A single agent that researches, decides, writes the code, and then grades its own homework has nobody checking it. It quietly cuts a corner at every step: the failure paths never get tested, the scope creeps, and "done" is the model's own optimistic guess.
Splitting the work gives you the checks a real team has:
| Agent | Job | Hard boundary |
|---|---|---|
| codebase-researcher | Maps the relevant files, existing patterns, and risks before a line is written | Read-only — cannot edit |
| story-writer | Turns the idea into acceptance criteria, edge cases, and explicit out-of-scope | Never invents business rules |
| spec-writer | Writes the technical brief, listing every file that will change | Cannot touch files |
| backend-builder | APIs, services, migrations, jobs, unit tests | Never touches UI code |
| frontend-builder | Components, pages, hooks, UI tests | Never invents endpoints; consumes the API as built |
| test-verifier | Acceptance tests written from the story, from the outside | May only edit test files — cannot patch the implementation |
| implementation-validator | Reports gaps, security issues, and scope creep by severity, with file and line numbers | Fixes nothing — only tells the truth |
The boundaries are the point. The agent that writes the code is not the agent that decides whether it works.
The three checkpoints
- CP1 — approve the story. Acceptance criteria and edge cases, before any design exists.
- CP2 — approve the brief. Every file that will change, listed up front, before anything is touched.
- CP3 — read the verdict. A validation report grouped by severity. You decide commit or PR.
Between checkpoints the chain runs itself. Artifacts land in factory-work/<feature-name>/ so you can read exactly what each agent produced.
Night mode
Approve the story and the brief in the evening, then run with --night. The builders, the test verifier, and the validator run unattended and stop at CP3 with everything written to files. You read 05_final_report.md over coffee.
In night mode the agents never stall waiting for input: they choose the safe option and record the open question in the final report. Out-of-scope changes are still forbidden.
Rework rules
Acceptance criteria that fail at the test stage, and Critical/Important findings from the validator, go back to the responsible builder — never to whoever happens to be in context. Maximum three rework round-trips; past that the chain stops and writes the remaining problems honestly into the final report rather than declaring success.
Requirements
- A git repository (so rework and rollbacks stay safe)
- A
CLAUDE.mdat the project root describing your stack, commands, and conventions. If you don't have one,templates/CLAUDE.md.templatein this repo is a starting point — including a "wrong assumptions the AI made" ledger that turns each mistake into a permanent guardrail.
Honest limitations
- These are prompts, not magic. On a codebase with no tests and no
CLAUDE.md, the first runs will mostly surface that fact. That is useful, but it is not free features. - Night mode is only as safe as the brief you approved. Approve carefully.
- Seven agents cost more tokens than one. The projected cost is visible any time via
claude plugin details feature-factory.
License
MIT — see LICENSE. Use it in your own work, commercial or not.
There is also a paid add-on pack with anti-hallucination hooks (one hook logs every URL the model actually fetched; a companion script flags anything in a draft that was never fetched) and a prompt-design guide: gumroad. The plugin above is complete and fully functional without it.
Not affiliated with Anthropic. Claude Code is a product of Anthropic.