Skip to main content
Superpowers imposes a complete development workflow on your coding agent. Skills are not suggestions — they are mandatory checkpoints the agent MUST pass through. Every step activates automatically at the right moment.
The agent checks for relevant skills before any task. If there is even a 1% chance a skill applies, the agent invokes it. This is not optional behavior.

The seven-step loop

1

Brainstorming

Before writing a single line of code, the agent explores the project context, asks clarifying questions one at a time, proposes 2–3 approaches with trade-offs, and presents a design in sections for your approval. A design document is saved to docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md and committed. Implementation is blocked until you approve the spec.Learn more about brainstorming →
2

Creating a worktree

After design approval, the agent creates an isolated git worktree on a new branch using the using-git-worktrees skill. This prevents any implementation from touching main or master. The worktree is verified to have a clean, passing test baseline before work begins.
3

Writing a plan

With an approved spec and a clean worktree, the agent writes a detailed implementation plan: exact file paths, complete code, TDD steps, verification commands, and commit instructions for every task. Tasks are 2–5 minutes each. No placeholders, no “TBDs”. The plan is saved to docs/superpowers/plans/YYYY-MM-DD-<feature-name>.md.Learn more about writing plans →
4

Executing the plan

You choose how tasks are executed: subagent-driven (recommended) or inline. Subagent-driven dispatches a fresh subagent per task with isolated context and automatic two-stage review after each one. Inline execution runs tasks in sequence with human-in-the-loop checkpoints.Learn more about executing plans →
5

Test-driven development

During execution, the test-driven-development skill enforces strict RED-GREEN-REFACTOR: write the failing test, watch it fail, write minimal code to pass, watch it pass, commit. Code written before a failing test is deleted. This happens for every task in the plan.
6

Code review between tasks

After each task, the requesting-code-review skill runs a review against the plan and reports issues by severity. In subagent-driven mode, this is a two-stage review: spec compliance first (did the implementation match the plan?), then code quality (is it well-built?). Critical issues block the next task.
7

Finishing the branch

When all tasks are complete, the finishing-a-development-branch skill verifies the full test suite, runs a final code reviewer subagent over the entire implementation, and presents exactly four options: merge locally, push and create a PR, keep the branch as-is, or discard. The worktree is cleaned up automatically for merge and discard.Learn more about finishing a branch →

Skill priority hierarchy

When skills conflict with other instructions, this order applies:
PrioritySourceExample
1 (highest)User’s explicit instructionsCLAUDE.md, AGENTS.md, direct requests
2Superpowers skillsbrainstorming, writing-plans, TDD
3 (lowest)Default system promptPlatform defaults
If your CLAUDE.md says “skip TDD for this project” and the TDD skill says “always enforce RED-GREEN-REFACTOR,” your explicit instruction wins. You are always in control.

How skills trigger automatically

You do not invoke skills manually. The using-superpowers skill instructs your agent to check for relevant skills before responding to any message. The rule: if there is even a 1% chance a skill applies, the agent invokes it. Skills load their full instructions at invocation time — the agent does not rely on cached knowledge. This means skill updates take effect immediately without reinstalling Superpowers.
Skills are rigid (like TDD and debugging) or flexible (like design patterns). Rigid skills must be followed exactly. The skill itself tells you which kind it is.

Brainstorming

Refine ideas, explore approaches, produce a validated design doc before any code is written.

Writing plans

Break the approved spec into 2–5 minute tasks with complete code and TDD steps.

Executing plans

Dispatch subagents per task with two-stage review, or run inline with checkpoints.

Finishing a branch

Verify tests, run a final review, and choose to merge, PR, keep, or discard.

Build docs developers (and LLMs) love