obra/superpowers · skills/subagent-driven-development/SKILL.md
subagent-driven-development
“A strict subagent foreman for running coding plans: smart rules, few check-ins, half its manual elsewhere”
What it does
Runs a written implementation plan by sending each task to a fresh subagent, then a reviewer for spec and quality. It loops fixes, logs every step and decision to a ledger file that survives context loss, and ends with a whole-branch review.
- When the agent uses it
- When the agent is executing an implementation plan whose tasks are mostly independent, in the current session, with a subagent tool available.
- Who it is for
- For superpowers users with written plans and a subagent tool. Skip it without the rest of the repo, or if you want the agent to ask before it decides plan conflicts.
- What it costs
- 278 lines, about 5.3k tokens once the agent loads it.
Use when executing implementation plans with independent tasks in the current session
What it tells the agent to do
- Work in a worktree, run scripts/sdd-workspace, then create or resume the plan's progress.md ledger
- Read the plan and any spec once, make a todo per task, and run the preflight conflict scan
- Per task: record BASE, extract a brief with scripts/task-brief, dispatch one implementer with an explicit model
- On DONE, build a review package from BASE..HEAD and dispatch a task reviewer for spec and quality
- On findings, follow references/fix-loop.md, re-reading it before every round
- Log the task complete in the ledger and move on without pausing to check in
- After the last task, follow final-review.md, delete the plan workspace, and hand off to finishing-a-development-branch
Lines worth reading
Straight from the file, with why each one matters.
line 3Use when executing implementation plans with independent tasks in the current session
The whole trigger. It says when to use the skill but never mentions subagents, reviews or the ledger, which are what it actually does.
line 23Ruling: <what you decided> — <why> — <what it costs if wrong>
A neat format for autonomous decisions. Each judgment call is written down with its cost, so you can audit it later.
line 222a real session's dispatch hit 42k chars of which 99%
A war story with numbers: 99% of that prompt was pasted history. It is why each dispatch must carry only its own task.
line 160Always specify the model explicitly when dispatching a subagent.
A sharp cost trap. An omitted model inherits the session's expensive one, which silently defeats the model-tiering advice.
line 290pre-judging, usually to spare yourself a review loop.
The skill names the agent's laziness to its face and lists the exact phrases that give it away.
Watch out for
- Medium
The agent decides plan conflicts and defects itself, even raising its own caps, and does not stop to ask unless an action is destructive, security-sensitive or shared.
line 20a cap you would have asked to exceed — decide
- Medium
It runs bundled bash scripts that were not available to review, so what they actually execute is unverified.
line 114bash scripts/sdd-workspace PLAN_FILE
- Low
It deletes the plan's scratch workspace at the end. That folder is git-ignored and holds the ledger, briefs and reports.
line 80Final review clean: delete this plan's workspace
The scorecard
- TriggerSays when to use it5/10
Says when, not what. No mention of subagents, reviews or the ledger.
- FocusOne clear job8/10
One job, plan execution through subagents, clearly bounded against executing-plans.
- LeanLight to load6/10
About 340 lines and ~5k tokens. Detail is offloaded, but anecdotes and the model section bloat it.
- ConcreteSteps you can follow7/10
Real scripts, ledger formats and statuses. The fix loop and final review live in unseen files.
- SafeSafe to run7/10
Clear stop list for destructive, security and shared actions, but autonomy is the default.
What it does well
- The ledger plus git log survives context compaction, which fixes a real and costly failure: redoing finished tasks.
- Dispatches and diffs are handed over as files, keeping the controller's context small. A thoughtful token budget.
- The 'Common Rationalizations' table pre-empts the shortcuts an agent will reach for, each with a firm reply.
How it could be better
Say what it does in the description
The agent only sees the description, and this one never mentions subagents, per-task review or the ledger.
Adddescription: Executes a written implementation plan task by task: dispatches a fresh subagent per task, reviews each for spec and quality, logs progress to a ledger, and ends with a whole-branch review. Use when asked to execute, run or implement a plan with mostly independent tasks in this session.
Announce autonomy once before Task 1
The skill never checks in and decides conflicts itself. One upfront notice lets the user opt out cheaply.
AddBefore dispatching Task 1, post one line: "Running all <N> tasks without pausing; rulings go to <workspace>/progress.md. Say stop to halt."
Move Model Selection to a reference file
Thirty-plus lines of tiering advice load on every use; a short rule plus a link would keep SKILL.md lean.
Add## Model Selection Always set the model explicitly. Mid-tier is the floor for reviewers; use the cheapest tier only when the brief contains the full code. Details: [references/model-selection.md](references/model-selection.md)
Inline the fix-loop cap
The loop's limits live only in an unseen file, so a reader of SKILL.md can't tell how long a stuck task may churn.
AddFix loop: at most <N> rounds; rounds 1-3 resume the implementer, rounds 4-<N> escalate a tier; then park findings with a Ruling.
Kin reviews against Anthropic’s Extend Claude with skills and agentskills.io’s the Agent Skills standard. It reads SKILL.md only, so a script beside it can still do things the file does not say.