一键导入
subagent-driven-development
Use when executing implementation plans with independent tasks in the current session
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Use when executing implementation plans with independent tasks in the current session
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when implementation is complete, all tests pass, and you need to decide how to integrate the work - guides completion of development work by presenting structured options for merge, PR, or cleanup
Use when you have a written implementation plan to execute in a separate session with review checkpoints
Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes
Use when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback
Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions
Use when you have a spec or requirements for a multi-step task, before touching code
| name | subagent-driven-development |
| description | Use when executing implementation plans with independent tasks in the current session |
Execute plan by dispatching a fresh implementer subagent per task, a task review (spec compliance + code quality) after each, then two whole-branch passes at the end: a correctness review, then a refactor review (design lens — dead code, cross-task duplication, messy hardcodes, smells).
Why subagents: You delegate tasks to specialized agents with isolated context. By precisely crafting their instructions and context, you ensure they stay focused and succeed at their task. They should never inherit your session's context or history — you construct exactly what they need. This also preserves your own context for coordination work.
Core principle: Fresh subagent per task + task review (spec + quality) + broad final review = high quality, fast iteration
You orchestrate; you commit. Implementer and fix subagents NEVER commit, push, or touch git state — they leave their work uncommitted in the working tree. Only you commit, and only after a task's review passes and you have confirmed its changes do not collide with other tasks. You hold the whole-run overview; a subagent committing mid-stream would break per-task review isolation and could entangle parallel tasks. See Committing.
Parallelisation. Tasks run as concurrent subagents in the one shared
feature worktree — no extra worktree per subagent. Which tasks batch together
is computed by scripts/plan-batches, not decided by you (see
Pre-Flight Plan Review). To dispatch a batch, use
superpowers-custom:dispatching-parallel-agents and launch every task in the
batch in the background (non-blocking) in a single message — a blocking
dispatch serializes them and defeats the point. When the batch returns, review
and commit each task per its own files (Committing).
Narration: between tool calls, narrate at most one short line — the ledger and the tool results carry the record.
Continuous execution: Do not pause to check in with your human partner between tasks. Execute all tasks from the plan without stopping. The only reasons to stop are: BLOCKED status you cannot resolve, ambiguity that genuinely prevents progress, or all tasks complete. "Should I continue?" prompts and progress summaries waste their time — they asked you to execute the plan, so execute it.
digraph when_to_use {
"Have implementation plan?" [shape=diamond];
"Tasks mostly independent?" [shape=diamond];
"Stay in this session?" [shape=diamond];
"subagent-driven-development" [shape=box];
"executing-plans" [shape=box];
"Manual execution or brainstorm first" [shape=box];
"Have implementation plan?" -> "Tasks mostly independent?" [label="yes"];
"Have implementation plan?" -> "Manual execution or brainstorm first" [label="no"];
"Tasks mostly independent?" -> "Stay in this session?" [label="yes"];
"Tasks mostly independent?" -> "Manual execution or brainstorm first" [label="no - tightly coupled"];
"Stay in this session?" -> "subagent-driven-development" [label="yes"];
"Stay in this session?" -> "executing-plans" [label="no - parallel session"];
}
vs. Executing Plans (parallel session):
digraph process {
rankdir=TB;
subgraph cluster_per_task {
label="Per Task";
"Dispatch implementer subagent (./implementer-prompt.md)" [shape=box];
"Implementer subagent asks questions?" [shape=diamond];
"Answer questions, provide context" [shape=box];
"Implementer subagent implements, tests, self-reviews (NO commit)" [shape=box];
"Snapshot tree, write diff file, dispatch task reviewer subagent (./task-reviewer-prompt.md)" [shape=box];
"Task reviewer reports spec ✅ and quality approved?" [shape=diamond];
"Dispatch fix subagent for Critical/Important findings" [shape=box];
"Controller verifies no file collision, then commits the task" [shape=box];
"Mark task complete in todo list and progress ledger" [shape=box];
}
"Read plan, note context and global constraints, create todos" [shape=box];
"More tasks remain?" [shape=diamond];
"Dispatch final code reviewer subagent (../requesting-code-review/code-reviewer.md)" [shape=box];
"Dispatch refactor reviewer subagent (./refactor-reviewer-prompt.md)" [shape=box];
"Refactor-Critical findings?" [shape=diamond];
"Fix loop: dispatch fix subagent, re-review; commit fixes" [shape=box];
"Write Refactor-Advisory items to refactor-report.md for human triage" [shape=box];
"Use superpowers-custom:finishing-a-development-branch" [shape=box style=filled fillcolor=lightgreen];
"Read plan, note context and global constraints, create todos" -> "Dispatch implementer subagent (./implementer-prompt.md)";
"Dispatch implementer subagent (./implementer-prompt.md)" -> "Implementer subagent asks questions?";
"Implementer subagent asks questions?" -> "Answer questions, provide context" [label="yes"];
"Answer questions, provide context" -> "Dispatch implementer subagent (./implementer-prompt.md)";
"Implementer subagent asks questions?" -> "Implementer subagent implements, tests, self-reviews (NO commit)" [label="no"];
"Implementer subagent implements, tests, self-reviews (NO commit)" -> "Snapshot tree, write diff file, dispatch task reviewer subagent (./task-reviewer-prompt.md)";
"Snapshot tree, write diff file, dispatch task reviewer subagent (./task-reviewer-prompt.md)" -> "Task reviewer reports spec ✅ and quality approved?";
"Task reviewer reports spec ✅ and quality approved?" -> "Dispatch fix subagent for Critical/Important findings" [label="no"];
"Dispatch fix subagent for Critical/Important findings" -> "Snapshot tree, write diff file, dispatch task reviewer subagent (./task-reviewer-prompt.md)" [label="re-review"];
"Task reviewer reports spec ✅ and quality approved?" -> "Controller verifies no file collision, then commits the task" [label="yes"];
"Controller verifies no file collision, then commits the task" -> "Mark task complete in todo list and progress ledger";
"Mark task complete in todo list and progress ledger" -> "More tasks remain?";
"More tasks remain?" -> "Dispatch implementer subagent (./implementer-prompt.md)" [label="yes"];
"More tasks remain?" -> "Dispatch final code reviewer subagent (../requesting-code-review/code-reviewer.md)" [label="no"];
"Dispatch final code reviewer subagent (../requesting-code-review/code-reviewer.md)" -> "Dispatch refactor reviewer subagent (./refactor-reviewer-prompt.md)" [label="correctness clean"];
"Dispatch refactor reviewer subagent (./refactor-reviewer-prompt.md)" -> "Refactor-Critical findings?";
"Refactor-Critical findings?" -> "Fix loop: dispatch fix subagent, re-review; commit fixes" [label="yes"];
"Fix loop: dispatch fix subagent, re-review; commit fixes" -> "Refactor-Critical findings?" [label="re-review"];
"Refactor-Critical findings?" -> "Write Refactor-Advisory items to refactor-report.md for human triage" [label="none / all fixed"];
"Write Refactor-Advisory items to refactor-report.md for human triage" -> "Use superpowers-custom:finishing-a-development-branch";
}
Before dispatching Task 1, scan the plan once. This scan has two required outputs — a conflict list AND a batch plan. Do both before any dispatch.
1. Conflict scan. Look for:
Present everything you find to your human partner as one batched question — each finding beside the plan text that mandates it, asking which governs — before execution begins, not one interrupt per discovery mid-plan. If the scan is clean, proceed without comment. The review loop remains the net for conflicts that only emerge from implementation.
2. Batch plan (mandatory — the script computes it, not you). You do NOT
decide batching by judgment. You DECLARE each task's files and dependencies in
a file; scripts/plan-batches computes the execution order deterministically.
This exists because the one judgement controllers reliably get wrong is
calling two tasks dependent because they are thematically "related" ("this
file changes something related to that one, so let's wait") — a phantom edge
that serializes work that could run in parallel. The script removes that
discretion: it honours only edges that name a real produced/consumed symbol,
and it does the layering so you cannot fall back to "all sequential to be
safe."
A dependency edge exists between two tasks ONLY when both are true; the script enforces the second:
Procedure:
.superpowers/sdd/deps.txt — one stanza per task: its task: id, the
files: it writes, and a needs: line per real dependency naming the
consumed symbol (needs: task-1 consumes=parseToken). The format and rules
are documented at the top of scripts/plan-batches. Declare an edge ONLY
with a nameable symbol; when in doubt, leave it out and let the file-collision
split and the per-task review catch a genuine conflict.scripts/plan-batches (defaults to that path). It prints the batch plan,
or exits non-zero naming a specific problem: a needs: with no
consumes=, a vague consumes= ("related", "similar", "theme", …), a
dangling edge, a dependency cycle, or a task with no files. Each of those is
a plan gap — fix the declaration (or the plan), do not work around it by
serializing.If you cannot derive a task's file set or its real dependencies from the plan, that is a plan gap: resolve it before execution, not by falling back to sequential.
Worked example. A 4-task plan. Task 2 calls a function Task 1 writes; Task 4 uses a constant Task 1 defines; Task 3 shares no symbol with anyone but writes a file Task 1 also touches. You write:
task: task-1
files: src/auth.py
task: task-2
files: src/handler.py
needs: task-1 consumes=parseToken
task: task-3
files: src/auth.py
task: task-4
files: src/config_check.py
needs: task-1 consumes=AUTH_TIMEOUT
Run scripts/plan-batches → it prints:
batch 1 (single): task-1
batch 2 (single): task-3
batch 3 (parallel): task-2, task-4
file-collision splits (same layer, overlapping files):
task-3 and task-1 both write src/auth.py — split into separate batches
Read what the script did that you would have gotten wrong by hand: task-2 and
task-4 both depend on task-1 but NOT on each other — no shared symbol — so they
run concurrently in batch 3. A cautious controller would have serialized
all four. task-3 has no dependency at all, yet it can't join task-1's batch
because they write the same file, so it lands in its own batch (2). And no edge
says consumes=auth or consumes=related — every edge names the actual
symbol, or it is not an edge.
Default for step-based plans: Sonnet implementers. When the plan is a
well-specified, step-by-step plan, dispatch every implementer subagent on
Sonnet (claude-sonnet-4-6) — the steps carry the design, so implementation
is execution, not architecture. Do not escalate the implementer to a more
capable model just because a step looks involved.
Call the advisor only when the plan itself has an issue. The "advisor" is a one-shot dispatch to the most capable available model, used to resolve a defect in the plan — a step that is ambiguous, internally contradictory, under-specified, or that an implementer hit as BLOCKED for a design reason (not a context reason). Ask the advisor the specific plan question, fold its answer back into the plan/brief, then continue on Sonnet. Routine implementation never calls the advisor. The final whole-branch review is the other most-capable dispatch (see below).
Beyond that default, use the least powerful model that can handle each role. Pick by what the task actually demands:
Turn count beats token price. The cheapest models routinely take 2-3× the turns on multi-step work, costing more overall — hence the mid-tier floor for reviewers and prose-fed implementers.
Always specify the model explicitly when dispatching. An omitted model inherits your session's model — often the most capable and expensive — which silently defeats this section.
Implementer subagents report one of four statuses. Handle each appropriately:
DONE: The implementer left its work uncommitted in the working tree. Capture it and review it:
AFTER=$(scripts/snapshot) — tree SHA of the working tree now.BASE is the tree before this task: HEAD^{tree} for a sequential task, or the snapshot you recorded before dispatching a parallel batch.scripts/review-package "$BASE" "$AFTER" (add -- <this task's files> when the task is one of a parallel batch) — it prints the unique diff path.DONE_WITH_CONCERNS: The implementer completed the work but flagged doubts. Read the concerns before proceeding. If the concerns are about correctness or scope, address them before review. If they're observations (e.g., "this file is getting large"), note them and proceed to review.
NEEDS_CONTEXT: The implementer needs information that wasn't provided. Provide the missing context and re-dispatch.
BLOCKED: The implementer cannot complete the task. Assess the blocker:
Never ignore an escalation or force the same model to retry without changes. If the implementer said it's stuck, something needs to change.
The task reviewer may report "⚠️ Cannot verify from diff" items — requirements that live in unchanged code or span tasks. These do not block the rest of the review, but you must resolve each one yourself before marking the task complete: you hold the plan and cross-task context the reviewer lacks. If you confirm an item is a real gap, treat it as a failed spec review — send it back to the implementer and re-review.
Per-task reviews are task-scoped gates. The broad review happens once, at the final whole-branch review. When you fill a reviewer template:
scripts/review-package path
from Handling Implementer Status, never
pasted into your own context. The package has no commit list (the work is
uncommitted) — that is expected. Without bash: git diff --stat BASE AFTER
and git diff -U10 BASE AFTER redirected to one uniquely named file.scripts/review-package MERGE_BASE HEAD (MERGE_BASE = the commit the
branch started from, e.g. git merge-base main HEAD) and include the
printed path in the final review dispatch, so the final reviewer reads
one file instead of re-deriving the branch diff with git commands.After the final correctness review comes back clean, dispatch one more whole-branch pass: the refactor reviewer (./refactor-reviewer-prompt.md). It is a different lens from every review before it. The per-task reviewer and the final correctness reviewer answer "does it work / match spec." This one answers "now that it works, is it well-shaped, or should something be refactored before merge" — design smells, cross-task duplication, messy hardcodes, dead code, missing or leaky abstractions.
Why it must be whole-branch and late. These are properties of the assembled branch, not of any one task's diff. A symbol added in Task 1 and consumed in Task 5 looks dead in Task 1's scoped review; duplication between two tasks is invisible until both land; a hardcode is only "messy" relative to a constant another task defined. Independent task subagents never saw each other's code, so they reinvent helpers and constants — only a pass over the whole branch catches it. Unlike the per-task reviewer, this reviewer MAY read the whole codebase (it must, to grep every new symbol for a real consumer).
Hybrid by severity — two buckets, two fates:
.superpowers/sdd/refactor-report.md and carry them into
superpowers-custom:finishing-a-development-branch, where the human triages
each: fix now, ticket, or accept. A roll-up nobody reads is a silent discard.A clean branch with zero findings is a valid, expected result — this reviewer
does not manufacture refactors, and YAGNI makes a speculative abstraction a
smell, not a fix. Dispatch it on the most capable available model (design
judgment, same tier as the final correctness review), and reuse the same
branch package the final review used (scripts/review-package MERGE_BASE HEAD)
rather than regenerating it.
Everything you paste into a dispatch prompt — and everything a subagent prints back — stays resident in your context for the rest of the session and is re-read on every later turn. Hand artifacts over as files:
scripts/task-brief PLAN_FILE N — it extracts the task's full text to a
uniquely named file and prints the path. Compose the dispatch so the
brief stays the single source of requirements. Your dispatch should
contain: (1) one line on where this task fits in the project; (2) the
brief path, introduced as "read this first — it is your requirements,
with the exact values to use verbatim"; (3) interfaces and decisions
from earlier tasks that the brief cannot know; (4) your resolution of
any ambiguity you noticed in the brief; (5) the report-file path and
report contract. Exact values (numbers, magic strings, signatures, test
cases) appear only in the brief.…/task-N-brief.md → report …/task-N-report.md) and put it in
the dispatch prompt. The implementer writes the full report there and
returns only status, files changed (it does not commit), a one-line test
summary, and concerns.Implementer and fix subagents never commit — you do, and only here. For each task, after its review verdict is clean:
git diff --name-only HEAD). For a parallel batch, confirm no two tasks
wrote the same file. If a collision exists, do not commit — reconcile it
(re-dispatch the loser sequentially, or merge by hand) first.git add -- <task files> && git commit -m "...") so history stays
one-commit-per-task and the ledger can name a single SHA.After a sequential task's commit the working tree is clean again, so the next
task's BASE is simply HEAD^{tree}. Never let a subagent commit on your
behalf; never push until the whole branch is done and you reach
superpowers-custom:finishing-a-development-branch.
Conversation memory does not survive compaction. In real sessions, controllers that lost their place have re-dispatched entire completed task sequences — the single most expensive failure observed. Track progress in a ledger file, not only in todos.
cat "$(git rev-parse --show-toplevel)/.superpowers/sdd/progress.md". Tasks listed there
as complete are DONE — do not re-dispatch them; resume at the first task
not marked complete.Task N: complete (commit <sha7>, review clean).git log over your own recollection.git clean -fdx will destroy the ledger (it's git-ignored scratch); if
that happens, recover from git log.You: I'm using Subagent-Driven Development to execute this plan.
[Read plan file once: docs/superpowers/plans/feature-plan.md]
[Create todos for all tasks]
Task 1: Hook installation script
[Run task-brief for Task 1; dispatch implementer with brief + report paths + context]
Implementer: "Before I begin - should the hook be installed at user or system level?"
You: "User level (~/.config/superpowers/hooks/)"
Implementer: "Got it. Implementing now..."
[Later] Implementer:
- Implemented install-hook command
- Added tests, 5/5 passing
- Self-review: Found I missed --force flag, added it
- Files changed: install-hook.sh, install-hook.test.sh (left uncommitted)
[Snapshot tree, run review-package BASE AFTER, dispatch task reviewer with the printed path]
Task reviewer: Spec ✅ - all requirements met, nothing extra.
Strengths: Good test coverage, clean. Issues: None. Task quality: Approved.
[Verify only the task's files changed, then YOU commit the task]
[Mark Task 1 complete with the commit SHA]
Task 2: Recovery modes
[Run task-brief for Task 2; dispatch implementer with brief + report paths + context]
Implementer: [No questions, proceeds]
Implementer:
- Added verify/repair modes
- 8/8 tests passing
- Self-review: All good
- Files changed: recovery.sh (left uncommitted)
[Run review-package, dispatch task reviewer with the printed path]
Task reviewer: Spec ❌:
- Missing: Progress reporting (spec says "report every 100 items")
- Extra: Added --json flag (not requested)
Issues (Important): Magic number (100)
[Dispatch fix subagent with all findings]
Fixer: Removed --json flag, added progress reporting, extracted PROGRESS_INTERVAL constant
[Task reviewer reviews again]
Task reviewer: Spec ✅. Task quality: Approved.
[Mark Task 2 complete]
...
[After all tasks]
[Dispatch final code-reviewer]
Final reviewer: All requirements met, ready to merge
[Correctness clean — dispatch refactor reviewer with the same branch package]
Refactor reviewer:
Well-shaped: config loader, error types
Refactor-Critical: parseTimeout() in retry.ts:40 duplicates parseDuration()
in util.ts:12 (both ranges cited); MAX_RETRIES literal 5 in retry.ts:8
duplicates existing RETRY_LIMIT in config.ts:20
Refactor-Advisory: Task 3's Handler is growing into a God-object — consider
splitting transport from routing
Refactor before merge? Critical fixes only
[Dispatch one fix subagent with both Critical findings; fixer re-runs covering
tests; re-review Critical bucket → clean; commit fixes]
[Write the advisory God-object note to .superpowers/sdd/refactor-report.md]
Done! (carry refactor-report.md into finishing-a-development-branch)
Never (each expanded in the section named):
needs: edge on a theme instead of a nameable produced/consumed
symbol (Pre-Flight)scripts/plan-batches — dispatch the
batches it computes, don't decide parallelism reactivelyscripts/review-package BASE AFTER)scripts/task-brief;
skip the scene-setting context that says where the task fitsgit log after any compaction (Durable Progress).superpowers/sdd/refactor-report.mdRequired workflow skills:
Subagents should use:
Alternative workflow: