Coding agents
rqtk is built to be driven by coding agents as well as people: nothing prompts, every command prints JSON with --json, and exit codes are fixed. Four agent skills bring it into an agent’s normal workflow.
Skills
rqtk skills install # or, in a new repository: rqtk init --agents
| Skill | Invoked by | What it does |
|---|---|---|
rqtk-requirements | agent or you | Reading and writing .rqtk/: find the governing requirement, read its briefing, author items that lint clean. |
rqtk-verification | agent or you | The proof loop: link a test, run it, rqtk verify; done when rqtk lint and rqtk coverage --strict exit 0. |
/to-requirements | you | Turn a spec or discussion into needs and requirements with verification activities; you approve the draft before anything is written. |
/requirements-review | you | Review a diff against the requirements it touches: facts from rqtk impact, then a verdict per requirement, and the rqtk review commands that would settle what is still Suspect. |
The skills follow the open Agent Skills format, so they work with any agent that supports it:
| Agents | Read skills from | What rqtk installs |
|---|---|---|
| Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Amp, Cline, Zed, Warp, … | .agents/skills/ | the skill files |
| Claude Code | .claude/skills/ | a link per skill into .agents/skills/ |
rqtk skills install --for claude or --for universal installs for one family only; --copy writes copies instead of links; --dir installs anywhere else. The skills are compiled into the binary, so they always describe the rqtk you have. Reinstall after upgrading: untouched skills are updated, and skills you edited are kept.
rqtk also adds a short block to AGENTS.md (and to CLAUDE.md, unless it already imports @AGENTS.md) telling agents where requirements live and when work is done. It never creates those files unless you name one with --instructions.
Where the skills fit
The skills are small and slot into an existing flow rather than imposing one. They were designed alongside mattpocock/skills:
grill → /to-spec → /to-requirements → /to-tickets → /implement → /code-review + /requirements-review
/to-requirementsturns the spec into requirements with activity IDs.- Tickets cite those IDs in their acceptance criteria.
- The implementer’s tests cite the activity IDs, and
rqtk verifyrecords them. /requirements-reviewchecks the result on the requirements axis, next to the usual code review.
Commands agents use most
rqtk context REQ-SYS-0001 --json # the briefing: statement, links, tests, status, findings
rqtk impact main --json # what a change touches, what to re-verify, and with which tests
rqtk lint --json # findings with code, severity and file:line
rqtk explain RQ010 # what a rule checks and how to fix it
rqtk schema requirement # the JSON Schema of a file kind
rqtk add … --parent … --criteria … --activity … # a complete requirement in one command
Exit codes and the JSON contract are described under JSON output and exit codes.
Machine-readable docs
These docs are also available in plain Markdown for agents: /llms.txt indexes every page, and /llms-full.txt contains them all in one file.