using-superpowers

If you were dispatched as a subagent to execute one narrow task, skip this skill.

Rule

If there is even a small chance a research skill applies, load it before responding or acting.

User instructions still win. Skills define how to work, not what the human wants.

Platform Adaptation

Skills use Claude Code tool names. On other platforms, read:

• references/codex-tools.md
• references/gemini-tools.md

Research Skill Order

Use process skills before doing work:

1. paper-to-implementation for a paper idea or reproduction request
2. experiment-design before proposing changes to model, loss, data, augmentation, training, or evaluation
3. experiment-planning once the hypothesis and design are approved
4. experiment-worktree to create an isolated workspace before making experiment changes
5. experiment-execution when carrying out the plan (calls experiment-worktree automatically)
6. training-debugging for NaNs, divergence, OOMs, metric mismatches, or other failures
7. result-analysis when runs finish and you need to compare evidence
8. experiment-closeout when a run has ended and you must decide whether to keep or revert the code changes (calls finishing-experiment-branch for git cleanup)
9. finishing-experiment-branch to merge, push, pause, or discard an experiment branch
10. reproducibility-check before claiming an improvement or handing results to others

Red Flags

Stop and load the relevant skill if you catch yourself thinking:

• "I can just change the model and try it."
• "The metrics look better enough."
• "I'll explain the result without checking seeds/configs."
• "This failure is probably just learning rate."
• "I already know how to reproduce a paper."

What Good Looks Like

• Hypothesis before code
• Baseline and success metric before conclusions
• Smallest falsifiable experiment first
• Explicit keep-or-revert decision after each experiment
• Evidence attached to every claim
• Negative results recorded, not hidden