Teaching AI to code like a real engineer
agent37.club
"Build me a login page"
Immediately starts writing code. No questions. No plan. No tests. Hopes for the best.
Without structure, AI agents skip thinking and jump to typing. They forget edge cases, ignore tests, and drift off course on long tasks.
A training manual for AI coding agents. Instead of letting the AI wing it, Superpowers gives it a step-by-step playbook for every phase of building software.
AI agents are smart but undisciplined. They need guardrails and processes, not more intelligence. Superpowers provides that structure.
AI asks you questions to understand what you really want
Creates a clean workspace so nothing else gets broken
Breaks work into tiny 2-5 minute tasks
Writes tests first, then code to pass them
Another AI agent checks the work
Address any problems found in review
Merge the code once everything passes
Instruction cards for a board game
Each card explains the rules for one specific activity. The AI reads the card and follows the rules.
There are 14 skills covering brainstorming, planning, testing, debugging, code review, and more.
# What a skill file looks like name: test-driven-development description: Use when implementing any feature or bugfix # The Iron Law: # Write the test FIRST. # Watch it FAIL. # Write minimal code to PASS. # If you write code before a # test, DELETE IT and start over.
Write a test that checks what you want. Run it. It should fail (because you haven't built anything yet).
Write the minimum code to make that test pass. Nothing extra.
Clean up the code while keeping the test green. Make it nice.
If the AI writes code before writing a test, Superpowers tells it to delete everything and start over. Harsh, but it works.
Tests prove your code actually works. Without them, you're just hoping. AI agents are especially good at appearing confident while producing broken code.
A construction site
The foreman (main agent) reads the blueprints and assigns tasks. Workers (subagents) each build one piece. An inspector (reviewer) checks each piece before moving on.
Long conversations confuse AI. A fresh agent on each task means no "context pollution" from previous work.
Oct 2025: First published
Jan 2026: Marketplace launch
Mar 2026: 102,700 stars
+1,867 stars in a single day
Built by Jesse Vincent at Prime Radiant. The insight: AI needs process, not more intelligence.
Superpowers proves that the missing piece in AI coding isn't smarter models — it's better methodology.
agent37.club