Superpowers — agent workflow: plan → TDD → systematic debugging

When to use

Reach for Superpowers when you want an agentic methodology, not a one-off Cursor feature. It is a composable skill pack (plan → TDD → verify, plus systematic debugging) that teaches the agent how to work across a whole task.

Good fits:

• Greenfield features where jumping straight into files burns a session.
• Bug hunts where the agent keeps patching symptoms instead of isolating root cause.
• Multi-step work you want broken into a plan a junior engineer could follow — with real tests.

Contrast with Baonat’s other Cursor cards: cursor-create-rule and cursor-pr-review are single product features (/create-rule, /review). Superpowers is a workflow methodology pack you install once so planning, TDD, and debugging skills activate for the right phase.

Skip it for tiny one-file edits, or when you only need a durable project rule / first-pass PR review — use those narrower cards instead.

Access note

Host tools (Cursor, Codex, ChatGPT) may require a paid subscription to install and run plugins/skills. The GitHub artifact is MIT and free to clone from obra/superpowers. The upstream README also mentions Commercial Services (enterprise support / tooling). Baonat is recommending the free open-source methodology only — not Jesse’s commercial services, and not a free “whole stack” ChatGPT product. Running Superpowers inside a paid host is a host-plan choice; the skill pack itself is not a separately sold Baonat product.

Why

Bare “build X” prompts skip design, skip failing tests, and thrash when something breaks. Superpowers (by Jesse Vincent) packages an opinionated loop: tease out a readable design, get sign-off, write an implementation plan, then drive work with true red/green TDD and systematic debugging when reality disagrees.

Baonat marks this verified against the public repo and the Codex/ChatGPT plugin mirror. Link to upstream — do not paste wholesale SKILL.md bodies into chats; install the pack and let the skills load.

Install

1. Pick your harness and install Superpowers for that host (install separately if you use more than one).
2. Cursor: in Agent chat run /add-plugin superpowers, or search the plugin marketplace for “superpowers.”
3. Codex / ChatGPT plugins: install via the openai/plugins Superpowers entry (app Plugins UI or Codex CLI /plugins search).
4. Or clone the source of truth: github.com/obra/superpowers (MIT).
5. Before coding: ensure planning skills are available; for stubborn bugs, prefer skills/systematic-debugging over random edits.
6. Confirm host access — plugin/skill runtime may need your Cursor / Codex / ChatGPT plan.

Copy this

Kickoff — methodology before code:

Use Superpowers for this task. Do not write implementation code until I approve the plan.
1) Clarify the real goal and constraints
2) Produce a short design in digestible chunks for my review
3) After sign-off, write an implementation plan a junior engineer could follow
4) Implement with red/green TDD (failing test first)
5) If blocked, load systematic-debugging — no shotgun patches
Deliverables: approved plan → tests → code → verification notes

When a bug resists a quick fix:

Attach systematic-debugging. Reproduce first, isolate root cause, then one minimal fix + regression test.
Do not refactor unrelated code while diagnosing.

Concrete scenarios

1. Feature with unknowns. Agent brainstorms and plans in chunks; you approve; then TDD implementation instead of a 400-line speculative PR.
2. Flaky production bug. Systematic debugging forces reproduce → isolate → fix → verify, not “try another null check.”
3. Multi-agent / long session. Subagent-driven execution against a written plan so hours of autonomy stay on-rails.

Pitfalls / when NOT to use

• Confusing Superpowers with /create-rule or /review. Those are single Cursor features; this is a methodology pack.
• Implying it is a free ChatGPT product. Host runtime may be paid; GitHub is MIT; commercial services are optional and separate.
• Skipping human plan review. The value is the sign-off gate — do not rubber-stamp a vague plan.
• Tiny chores. Overkill for rename-a-symbol work.

FAQ

Q: Who wrote this?
A: Jesse Vincent (obra). Primary source: obra/superpowers.

Q: Is Baonat selling Superpowers or Jesse’s commercial services?
A: No. We recommend the free MIT methodology on GitHub. Commercial services mentioned upstream are optional and not what this card endorses.

Q: Cursor only?
A: Cursor is a primary fit for Baonat’s audience; the same pack also appears in the Codex/ChatGPT plugins ecosystem. Install per harness.

Q: Should I paste SKILL.md into chat?
A: No — install/link the pack and let skills load. Baonat links sources; it does not republish wholesale skill bodies.