Frontier Mind
The discipline that separates a frontier model from an eager intern — as installable skills. From Jesse Vincent's 257k-star Superpowers project: brainstorm before building, write plans another engineer could run, execute in checkpointed batches, test-drive every change, debug to root cause instead of guessing, refuse to say 'done' without proof, request and absorb code review, parallelize across subagents, and even write new skills of its own. Works on any Claude model — point Opus, Sonnet, or Haiku at these and watch it slow down, think, and verify like a bigger model. Every skill credited to its original author.
Install the whole bundle
10 skills into ~/.claude/skills/. Free.
Paste into your terminal. Restart Claude Code afterwards — new skills are picked up on startup.
This copies each skill's full folder (SKILL.md plus any supporting scripts, reference docs, or asset templates) so every skill works out of the box. Safe default.
Faster alternative (instruction-only skills)
Skips the git clone and downloads only each SKILL.md. Avoid if any skill in the bundle ships Python scripts, reference markdowns, or asset templates — they'll be missing and the skill will fail when it tries to load them.
What's inside
Brainstorming
obra
Force Claude to ask questions and pressure-test the idea before any code gets written.
Writing Plans
obra
Turn any task into a step-by-step implementation plan another engineer — or another model — could execute.
Executing Plans
obra
Execute a written plan in controlled batches with review checkpoints, instead of one giant leap.
Test-Driven Development
obra
Enforce strict RED → GREEN → REFACTOR: the failing test comes first, every time.
Systematic Debugging
obra
Replace guess-and-check debugging with a four-phase root-cause investigation.
Verification Before Completion
obra
Never let the model claim 'done', 'fixed', or 'passing' without running the proof first.
Requesting Code Review
obra
Have the model request a real review of its own work before it merges anything.
Receiving Code Review
obra
Take review feedback rigorously — verify each claim, push back with evidence, fix properly.
Dispatching Parallel Agents
obra
Split independent work across parallel subagents and merge the results, like a frontier-model orchestrator.
Writing Skills
obra
The meta-skill: teach the model to write, test, and improve new skills of its own.