Skip to content

Claude skill

If you use Claude Code, recurve install drops two global slash commands into ~/.claude/skills/ that wrap the whole workflow with a guided prompt — so you never have to touch the raw CLI unless you want to. They work in any repo, with no per-repo setup.

Under the hood they run exactly the verbs on the Provider-agnostic page — nothing here is a different engine, just an ergonomic front door.

Install

From a clone of the recurve repo:

pip install -e .        # or: pipx install .   (installs the `recurve` CLI)
recurve install         # links the binary AND installs the two skills

recurve install prints installed global skills into …: /recurve-plan, /recurve-work. So a fresh clone is one line:

git clone https://github.com/bordumb/recurve.git && cd recurve
pip install -e . && recurve install

(recurve install --no-skills links only the binary; --skills-dir DIR installs the skills somewhere other than ~/.claude/skills.)

/recurve-plan — get to a gate-ready plan

Run it in the target repo. It walks you through, stopping at the checkpoints:

  1. verifies the recurve CLI is on PATH (and tells you how to install it if not);
  2. initializes the repo (recurve init) if it isn't already;
  3. finds or interviews you into a docs/PRD.md — kept decomposition-oriented, because claims get carved from it;
  4. drives that PRD through the admission gate (recurve admit) — recurve's front-door check for "is this goal concrete enough to become falsifiable claims?" — working its interview worklist with you until it passes.

You end with an initialized repo and a PRD that is ready to build claims from.

/recurve-work — burn down under the gate

Run it after /recurve-plan. It:

  1. checks the baseline is clean (recurve validate / matrix --gate / lock status);
  2. if the ledger is empty, offers to author the first claims from the PRD (RED-first probes + traps, then baseline) — so it never runs on an empty backlog;
  3. asks how you want to run —
    • Endless — cycle after cycle until the gate is complete, goes red, or hits a dead-end;
    • Per-claim — stop after each claim closes so you review and approve the next;
  4. drives the loop with the hard rules baked in: the gate is the arbiter, never fake or weaken a probe, per-cycle commits.

What you don't give up

The skills are a convenience layer, not a lock-in. The same repo is fully drivable from the CLI by any agent ($AGENT_CMD) or a human following RUN.md. Claude Code just gets a nicer door — the evidence, the gate, and the ledger are identical either way.