Agent-Skills.md

Agent Skills: plan-loop

Draft an implementation plan for an epic or feature, then spar with Codex to harden it before saving. Use when the user wants to plan a sizable change and have a second model stress-test the plan, or says things like "plan this feature", "draft a plan and spar with Codex", "write an implementation plan for", or invokes /ca:plan-loop. Produces a saved task-by-task plan ready for the ca implement loop.

UncategorizedID: ken2403/.claude-paralell-dev-plugin/plan-loop
3

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ken2403/claude-parallel-dev-plugin/tree/HEAD/ca/claude/skills/plan-loop

Skill Files

Browse the full folder contents for plan-loop.

Download Skill

Loading file tree…

ca/claude/skills/plan-loop/SKILL.md

Skill Metadata

Name
plan-loop
Description
Draft an implementation plan for an epic or feature, then spar with Codex to harden it before saving. Use when the user wants to plan a sizable change and have a second model stress-test the plan, or says things like "plan this feature", "draft a plan and spar with Codex", "write an implementation plan for", or invokes /ca:plan-loop. Produces a saved task-by-task plan ready for the ca implement loop.

plan-loop

Draft a strong implementation plan, sharpen it by sparring with Codex, then save it. This is the planning half of the ca (Cooperate Agents) loop; the saved plan is later built by $ca-implement-plan and reviewed by /ca:review-pr.

Step 1 — Draft

Ground the plan in the codebase (read relevant files, dispatch Explore agents for unfamiliar areas) and write a task-by-task implementation plan following the superpowers writing-plans conventions: a clear goal, architecture, exact file paths, bite-sized TDD steps with real code/commands, and a self-review. Resolve ambiguities with the human before designing.

Include a ## Milestones section that groups the tasks into 2–4 natural checkpoints (layer or dependency seams — e.g. "data model", "API", "UI"); the implement loop pushes and gets a Claude checkpoint review at the end of each one, so defects surface before more code is built on top of them. A small plan (roughly 4 tasks or fewer) is a single milestone.

Step 2 — Spar with Codex (1–2 rounds)

Get an independent, adversarial critique from Codex. Codex exec calls are stateless — each call forgets the last — so write the FULL current draft plus your specific questions into a prompt file each round, then:

bash "${CLAUDE_SKILL_DIR}/scripts/spar-codex.sh" /path/to/round-N-prompt.md

Ask Codex to attack the plan: missing tasks, wrong sequencing, risky assumptions, simpler approaches, failure modes, tasks whose tests are missing or would not prove the behavior, and whether the milestone grouping/ordering is right (each milestone should leave the tree green and reviewable on its own). Incorporate what holds up; record (don't silently drop) what you reject and why. Repeat once more if the first round surfaced substantial changes.

Step 3 — Finalize and save

Apply the sparring outcomes, re-run the writing-plans self-review (spec coverage, no placeholders, type/identifier consistency), and confirm every task carries a failing-test spec and its test command — a task with no testable behavior must say so explicitly, because $ca-implement-plan's test-first step runs off exactly what each task specifies. Then save to the ca-owned plan dir (never docs/superpowers/ — that is superpowers' scratch namespace):

docs/ca/plans/YYYY-MM-DD-<feature-name>.md

Tell the human the saved path and that the next step is /ca:implement <plan> to launch the Codex implement loop. Open a plan PR only if the human asks.

Notes

  • CODEX_BIN overrides the codex binary if it is not on PATH.
  • Sparring runs Codex read-only (no edits) at high reasoning; it is advice, not authority — you own the final plan.