subagent-it

Prompt file imported from sealmindset/make-it (.claude/commands/subagent-it.md). Copyright stays with the author.

Turn an approved implementation plan into reviewed, committed code — one task at a time, each implemented by a fresh subagent with an isolated, file-based context, gated by a per-task review (spec + quality + safeguards) and a final whole-branch review. You coordinate; subagents do the work, so your context stays clean.

Full doctrine: ~/.claude/make-it/references/subagent-driven-development.md.

<expert_persona>

When constructing each task's implementer brief (and its reviewer brief), assign the specialist persona matching that task's domain — do NOT dispatch generalists. Infer per task from the task text + app-context.json signals, and inject the persona directive (identity + mental model + output standard) into that subagent's context so it implements/reviews to the expert's standard. No user prompt. See the reference below.

@~/.claude/make-it/references/expert-personas.md

</expert_persona>

/subagent-it -- Subagent-Driven Development

Core principle: fresh subagent per task + task review (spec + quality + safeguards) + broad final review. Sequential only — never run implementers in parallel (they conflict). Independent parallel problems are /dispatch-it's job.

Needs an approved plan as input (it executes, it doesn't brainstorm). A plan marked status: draft isn't approved: get the user's go, then mark it status: approved (Plan format: ~/.claude/make-it/references/backlog-board.md). Work on a branch/worktree, never main/master. Don't pause to check in between tasks — run the whole plan; stop only for an unresolvable BLOCKED, genuine blocking ambiguity, or completion.

Tooling

• ~/.claude/make-it/sdd/scripts/task-brief PLAN_FILE N → the task's brief file
• ~/.claude/make-it/sdd/scripts/review-package BASE HEAD → one diff file for a reviewer
• ~/.claude/make-it/sdd/test-writer-prompt.md / implementer-prompt.md / task-reviewer-prompt.md / adversarial-reviewer-prompt.md → dispatch templates
• Final review: /code-review. Finish: ship-it / wrap-it.

Steps

1. Resume check — cat "$(git rev-parse --show-toplevel)/.make-it/sdd/progress.md". Tasks marked complete are DONE; resume at the first unmarked one. Arm the ledger: append RUN: plan=<path> tasks=<N> (or RUN: resume if already armed), then register it from the repo/worktree that holds it: python3 ~/.claude/make-it/sdd/scripts/sdd-gate register "$(git rev-parse --show-toplevel)/.make-it/sdd/progress.md". A Stop hook (sdd-gate) then refuses to let the run end until you append STATUS: DONE | BLOCKED <why> | AWAITING_HUMAN <what> — write AWAITING_HUMAN before asking the user anything, and ask it readable cold: its title, what waits and why, 2-4 named answers, a batch as one grouped line, a high-stakes item never batched (approval-block.md, A list of asks).
2. Read the plan once — note global constraints, create todos, ensure you're on a branch/worktree.
3. Pre-flight review (once) — scan for tasks that contradict each other/the constraints or that mandate something the review rubric treats as a defect; present all findings as ONE batched question (finding beside plan text, "which governs?"). Clean → proceed silently.
4. Per task (sequential): task-brief → record BASE → dispatch the test-writer (template, Opus or Fable) → red check: run its tests, they must fail for the missing behavior (NOT_TESTABLE: skip to the implementer with no Tests section) → dispatch implementer (template, cheapest fitting model, ALWAYS specify it; it may not change the tests; a test it says is wrong goes back to the test-writer) → answer its questions → on DONE, tests untouched and green: git diff --name-only <TESTS_SHA> -- <test files> prints nothing and you run them yourself, all passing, none skipped (again after every fix) → review-package → dispatch task reviewer (template, with every SAFEGUARDS line the task produced) → fix-subagent loop for Critical/Important until spec ✅, quality approved, and safeguards ✅ (an undisclosed weakening is undone; a disclosed one is undone too, unless that changes behavior the plan didn't decide -- then the human's call; log each as Task N: safeguards undone|accepted: <item>; three consecutive rounds on the same root cause: stop patching and write STATUS: AWAITING_HUMAN with a decision-ready design choice) → adversarial review (tasks with tests, at most 3 rounds per task; the repo must be untouched after each): confirmed counterexamples only (its SAFEGUARDS line is separate: an undisclosed one is undone, in any round), which you adjudicate (ungrounded or plan-conflicting → the human; unsure severity → Important); in rounds 1–2 each becomes one test-writer test on its exact input (red check), the implementer fixes it, tests untouched and green, re-review, next round; round 3's are not fixed -- open Critical/Important → STATUS: AWAITING_HUMAN, Minor only → accept and log; after the human answers, the count doesn't reset -- apply their call and carry on where you stopped → append Task N: complete (commits <base7>..<head7>, review clean) to the ledger. Handle DONE_WITH_CONCERNS / NEEDS_CONTEXT / BLOCKED per the reference.
5. Final review — /code-review on the most capable model with review-package MERGE_BASE HEAD (git merge-base main HEAD) and every safeguards accepted ledger line; fix findings with ONE fix subagent carrying the full list.
6. Finish — after STATUS: DONE, offer learnings (~/.claude/make-it/references/learnings.md): queued agent lessons (e.g. reviewer findings that recurred across tasks), plus corrections the user made at acceptance — one message, each numbered with its evidence and what it would change; the user answers per item. Then ship-it / wrap-it.

Hard rules (see reference for the full list)

• Never pre-judge findings or tell a reviewer what not to flag.
• Hand artifacts as FILES (brief / report / package) — never paste plan text or diffs into your context.
• Use the recorded BASE, never HEAD~1. One task per dispatch — no pasted session history.
• All three verdicts (spec, quality, safeguards) are required; never skip a re-review. An undisclosed weakened safeguard is undone, never kept (approval-block.md).
• Trust the ledger + git log over recollection after a compaction/resume.

See subagent-driven-development.md for model selection, status handling, reviewer ⚠️ items, reviewer-prompt construction, file handoffs, durable progress, and red flags.