Agent Superpowers — Skills Package

Seven self-contained agent skills, compatible with the Agent Skills SKILL.md standard read by Claude Code, Codex, Cursor, Gemini CLI, Windsurf, OpenCode, and others.

Python 3.8+, standard library only. Zero pip dependencies.

The skills

Skill	What it does
ultra-instinct	Routes routine queries (status checks, lint, tests) straight to a mapped command, skipping the reasoning loop. Reasoning intent always vetoes a route match.
copy-ability	Parses an execution transcript into the commands that were run and emits a SKILL.md playbook.
invisible-attacks	Validates path mutations against an allowed root; queues, runs, and tracks speculative background tasks.
mastery	Serves per-stage (plan / build / audit / format) context budgets, tool allowlists, focus rules, and system prompts.
heat-mode	Classifies repeated tool failures and returns a concrete recovery plan.
cqc	Checks a command for perimeter escape before running it, then diffs the filesystem for side-effects.
animal-instinct	Reproducible mutation fuzzing of a target CLI, with the exact payload recorded for every finding.

Directory structure

Superpowers/
├── ultra-instinct/        # scripts/reflex_router.py
├── copy-ability/          # scripts/trajectory_cloner.py
├── invisible-attacks/     # scripts/sandbox_enforcer.py
├── mastery/               # scripts/context_adapter.py
├── heat-mode/             # scripts/anti_stall.py
├── cqc/                   # scripts/cqc_executor.py, scripts/boundary_validator.py
├── animal-instinct/       # scripts/mutation_fuzzer.py
├── tests/                 # stdlib regression suite
└── index_1.html           # landing page

Every skill folder has the same shape:

<skill>/
├── SKILL.md                     # frontmatter + agent-facing playbook
├── scripts/*.py                 # standalone argparse CLIs
└── references/<slug>_guide.md    # rationale, threat model, schemas
    references/<slug>_runbook.md  # condensed steps + failure modes

Install

Copy any skill folder into your agent's skill directory:

Agent	Path
Claude Code	.claude/skills/<skill> (project) or ~/.claude/skills/<skill> (global)
Antigravity	.agents/skills/<skill> at the workspace root
Cursor	.cursor/skills/<skill> (also reads .agents/ and .claude/)
Windsurf	.windsurf/skills/<skill> (reload the workspace after adding)
OpenCode	.opencode/skills/<skill> (also reads .claude/ and .agents/)
Others	.agents/skills/<skill> — the emerging cross-tool convention

Nothing to install or build. Each skill is portable on its own; copy one or all seven.

Tests

python3 -m unittest discover -s tests -v

71 tests, no runner to install. Each one pins a behaviour that was previously wrong in a way that produced a confident but incorrect result.

Scope and limits

These skills are honest about what they cannot do, and say so in their output rather than in a footnote:

• cqc and invisible-attacks are mistake detectors, not security sandboxes. Static analysis of a shell command is defeated by variable expansion, command substitution, eval, or any interpreter — those cases return UNDECIDABLE rather than "clean". Filesystem diffing only covers the roots it is told to watch, and every result names that scope. For an actual boundary use sandbox-exec (macOS), Landlock + seccomp or bubblewrap (Linux), or a container.
• mastery is advisory. It returns configuration for the calling agent to apply to itself; it cannot enforce a token budget or restrict a tool.
• animal-instinct is a blackbox fuzzer — no coverage feedback, no test-case minimization. For deeper work use Atheris or Hypothesis.
• copy-ability output is a draft. Generated commands come from a transcript and are unvalidated; review before running.

Runtime state

Skills that persist state write to $AGENT_SUPERPOWERS_STATE, else $XDG_STATE_HOME/agent-superpowers/, else ~/.local/state/agent-superpowers/ — never inside the skill folder, so a copied folder stays a clean read-only artifact.