agentsclimarketplace

Write plan

Skill taipt1504/claudehut/skills/write-plan

Use in the Plan phase after the spec is approved - dispatches the planner agent to draft the executable plan from the standard template (decision summary, T-xxx task breakdown with test-first + verify per task), gets the user's approval, records the plan (opening the write gate), and mirrors the breakdown into Claude Code's native task list. Runs inline on the main thread (it owns the approval gate, the state write, and the task mirror).From its SKILL.md

Install
npx -y skills add taipt1504/claudehut --skill write-plan

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.8 KB, as published. Nobody here has run it

Write Plan (Plan phase)

Convert the approved spec into an executable, test-first plan. Runs inline on the main thread — the planner drafts in isolation; this skill owns the user gate, the state write, and — where task tools exist — the task mirror (a subagent cannot).

Flow

flowchart TB
    s(["spec approved (set-spec recorded)"]) --> draft["dispatch claudehut-planner → draft plan.md<br/>(T-rows + §3 flow + per-task Sketch)"]
    draft --> check{"§1 restates decision, §3 traces end-to-end,<br/>every behavior task RED-first + verbatim verify?"}
    check -- "no" --> draft
    check -- "yes" --> smart{"full tier AND (≥5 tasks OR sensitive keyword)?<br/>the SAME predicate set-plan gates on"}
    smart -- "no" --> ask
    smart -- "yes" --> rev["dispatch claudehut-plan-reviewer → REFUTE vs spec<br/>writes plan-review.md (SubagentStop blocks empty return)"]
    rev --> verdict{"Verdict == APPROVE?"}
    verdict -- "REVISE (route items back)" --> draft
    verdict -- "APPROVE" --> record["set-plan-review APPROVE --evidence plan-review.md<br/>(byte-identical plan; edit forces re-review)"]
    record --> ask{"interactive? user Approves?"}
    ask -- "Request changes" --> draft
    ask -. "headless -p" .-> bypass(["record approval: non-interactive in header"])
    ask -- "Approve" --> setplan["set-plan plan.md (opens PreToolUse write gate)"]
    bypass --> setplan
    setplan --> mirror["IF task tools available: TaskCreate per T-row + TaskUpdate addBlockedBy<br/>(absent → skip; plan.md is the source of truth)"]
    mirror --> phase["set-phase implement → REQUIRED NEXT claudehut:implement"]

Process

  1. Dispatch claudehut:claudehut-planner. Task dir is derived from the recorded spec path (dirname of the set-spec path) — never recompute NNNN. Inputs: spec, reuse-scan + brainstorm (same dir), template references/plan-template.md. It writes …/tasks/NNNN-<slug>/plan.md; it does NOT write state. (set-plan REJECTS a plan with no | T-xxx rows.) Summer KB (when the project has .claude/summer-kb/): every T-xxx row whose files touch Summer wiring MUST carry the KB citation from the spec (.claude/summer-kb/<module>.md §<section>) in its verify column — the implementer verifies against it, and the plan-reviewer REVISEs a plan whose Summer tasks cite nothing.
  2. Dispatch claudehut:claudehut-plan-reviewer — on the SAME predicate set-plan gates on, not on every plan. It fires when the plan is substantial (≥5 | T- rows) OR sensitive — security, /auth, migration, flyway, liquibase, password, secret, owasp, permitAll, deserial anywhere in plan.md (mirror it exactly — dispatch on less and set-plan refuses a plan you chose not to review). Below the predicate the §1/§3 check above IS the doc gate: no dispatch, no verdict. When it fires it gates the doc BEFORE the user sees the plan and writes its coverage table + Verdict: APPROVE|REVISE to tasks/NNNN-<slug>/plan-review.md. Loop on REVISE — cap 2 rounds: each round re-drafts and re-reviews the whole plan, so a third REVISE means the spec is the problem, not the plan. Take it to the user with AskUserQuestion (the surviving gaps + the two attempts) instead of looping. Then record the verdict (only the main thread writes state):
    claudehut-state --session ${CLAUDE_SESSION_ID} set-plan-review APPROVE --evidence .claude/claudehut/tasks/NNNN-<slug>/plan-review.md
    
    The wire that makes the reviewer fire: set-plan REFUSES a full-tier plan meeting that same predicate unless plan_review==APPROVE is recorded for the byte-identical plan (smart-gate; editing forces re-review via content-hash). Headless -p, no Agent budget: A human may unblock both set-plan and the write gate with claudehut-state set-bypass true --reason '<why>' — ask, do not run it yourself; the reason is persisted in state.
  3. Get approval (this opens the write gate — not before). Interactive: AskUserQuestion with the decision + T-xxx list, Approve / Request changes. Non-interactive (-p): record approval: non-interactive run — proceeded with draft in the header. Only after approval (unlocks the PreToolUse gate):
    claudehut-state --session ${CLAUDE_SESSION_ID} set-plan .claude/claudehut/tasks/NNNN-<slug>/plan.md
    
  4. Mirror into the native task list — ONLY if you have task tools. Availability is per session and not guaranteed; TaskCreate/TaskUpdate are absent in many. No task tools → skip this step entirely, the same way claudehut:implement skips its orchestration section when there is no Agent tool. plan.md is the source of truth; the mirror is a convenience view, never a gate. With task tools: one TaskCreate per T-xxx row (subject: "T-001: <goal>", description: files + test-first + verify + req-ref, activeForm: present-continuous), then TaskUpdate addBlockedBy per Depends-on. plan.md stays the durable source of truth. Then enter Implement: claudehut-state --session ${CLAUDE_SESSION_ID} set-phase implement.

REQUIRED NEXT: claudehut:implement (test-first; the enforcement-set rules auto-load by path).

Keep looking

Skills are one crate of 609,689. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.