A free, open-source pack of 14 skills that teaches your AI a disciplined way to build software: design first, plan, build, debug properly, verify before claiming done. One-click install in the Codex app. Use it on your own projects after Demo Day.
Not an app. Not a plugin that adds features. A set of written instructions that change how your AI works.
Each “skill” is a markdown document that tells the AI how to behave for one kind of task — how to brainstorm, how to plan, how to debug. When you invoke one, the AI stops improvising and follows the method. The whole pack encodes one opinion:
Slow down at the start. Design before code, plan before build, find the root cause before the fix — and never say “done” without proof.
That is the opposite of how most people use AI (type a wish, accept whatever comes back). It is also why it works.
⌁ github.com/obra/superpowersThe source — every skill is a readable markdown file. MIT license.Wait until after Demo Day to install this. During the sprint we keep gstack as the only skill pack — two packs at once doubles your $ list and the names overlap. Superpowers is for the projects you build after this course.
Unlike gstack, Superpowers is in the Codex app's public plugin directory. No commands at all.
Install it once, from Public. If you ever see every skill listed twice (“Brainstorming” and “Superpowers: Brainstorming”), you installed it from two sources — go back to Plugins and remove one.
Same habit as gstack: type $, then pick from the list.
$ in the chat box → the skill list appears. Type a few letters ($brain, $debug, $test) to filter, then pick one. You never memorise names.Six stages. Real work loops back — that is normal, not failure.
Fourteen skills exist. Start with these six — each card: when, what, and a line to try.
When: before you build anything new — a feature, an app, a change of behaviour.
What it does: refuses to write code. It interviews you one question at a time, then produces a design document you approve before any building starts. This is its hard gate — and it is the whole point.
When: you have an approved design (usually straight after Brainstorming) and the work has multiple steps.
What it does: turns the design into a numbered implementation plan — small steps, each checkable — so a long build cannot silently drift.
When: the plan exists and it is time to build.
What it does: works through the plan step by step with review checkpoints — it stops for your approval instead of sprinting to the end and presenting a mystery.
When: anything breaks — an error, a failing test, weird behaviour.
What it does: bans guess-and-patch. It reproduces the bug, isolates it, and names the root cause before proposing any fix. Slower per bug, far faster per evening.
When: building any feature you would be embarrassed to break next week.
What it does: writes the test first, watches it fail, then writes the code that makes it pass. The test becomes proof — not decoration.
When: the AI (or you) is about to say “done”, “fixed”, or “passing”.
What it does: demands evidence — run the commands, show the output — before any claim of success. The cheapest honesty tool in the pack.
Skip these until the six above feel natural. One line each.
| Skill | One line |
|---|---|
requesting-code-review | Get a focused review of your work before merging it |
receiving-code-review | Evaluate review feedback before blindly applying it |
finishing-a-development-branch | Wrap up: tests pass → choose how to integrate the branch |
using-git-worktrees | Isolated copies of your repo for parallel work |
dispatching-parallel-agents | Split independent tasks across several AI agents at once |
subagent-driven-development | Run a plan with staged sub-agent implementers and reviewers |
using-superpowers | Teaches the AI when to reach for which skill |
writing-skills | Write and test your own skills — yes, you can make your own |
One small, real thing you actually want. Not a to-do app.
$brainstorming + one sentence about your idea. Answer its questions honestly — short answers are fine. Stop when it presents a design and asks for approval. Read the design before saying yes.$writing-plans. Read the steps. Cut any step you do not understand — smaller is better.$executing-plans. Approve or redirect at each checkpoint. If something breaks, $systematic-debugging — do not just retype the wish.$verification-before-completion. If it cannot show proof, it is not done.The AI asks you more questions than usual, and writes code later than usual. That delay is the method. The first ten minutes of questions save the two hours of rebuilding you would have spent otherwise.
You now own two packs. They answer different questions.
| gstack (used in class) | Superpowers (this guide) | |
|---|---|---|
| Core question | “Should this exist at all?” | “How do we build it well?” |
| Voice | A skeptical investor | A disciplined senior engineer |
| Shines when | You have an idea and no evidence | You have a validated idea and real code to write |
| Star skills | office-hours, plan-ceo-review | brainstorming, systematic-debugging, TDD |
| Install | Paste a paragraph (script install) | One click in Plugins |
Natural order for a real project: gstack first to decide whether the idea deserves your weekend — then Superpowers to build it without making a mess. If the combined $ list feels crowded, toggle one pack off in Plugins while you use the other.
| Symptom | Likely cause | Fix |
|---|---|---|
Typing $brain shows nothing | App not restarted after install | Quit the app fully and reopen |
| Every skill appears twice | Installed from two sources | Plugins → remove one copy (keep Public) |
$ list is overwhelming | gstack + Superpowers both active | Plugins → toggle one pack off for now |
| It keeps asking questions instead of coding | That is Brainstorming working as designed | Answer them — or skip the skill for tiny throwaway tasks |
| Want it gone | — | Plugins → Superpowers → uninstall. Nothing else to clean up |
github.com/obra/superpowers → skills/RELEASE-NOTES.md)