达尔文.skill
v1.0.0Autonomous skill optimizer inspired by Karpathy's autoresearch. Evaluates SKILL.md files using an 8-dimension rubric (structure + effectiveness), runs hill-c...
⭐ 0· 266·1 current·1 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
Capability signals
These labels describe what authority the skill may exercise. They are separate from suspicious or malicious moderation verdicts.
OpenClaw
Benign
medium confidencePurpose & Capability
Name/description claim an autonomous optimizer for SKILL.md files; the instructions explicitly read SKILL.md blobs, create test-prompts.json, run evaluations, and perform git commits/reverts. Asking to read/write skill files and use git is coherent and expected for this purpose.
Instruction Scope
The runtime instructions instruct scanning all skills (e.g., .claude/skills/*), editing other skills' SKILL.md, creating test-prompts.json, writing results.tsv, and spawning child agents to execute test prompts. Those actions are needed for optimization but give the skill broad file-modification scope across the user's skills directory and cause SKILL.md contents and test prompts to be sent to child agents (and therefore to the LLM provider). There are human-confirmation checkpoints described, but the SKILL.md grants the agent authority to modify other skill assets which is a sensitive capability.
Install Mechanism
Instruction-only skill with no install spec and no downloaded code or third-party packages. Low install risk — nothing is written by an installer step beyond what the instructions later direct the agent to do.
Credentials
No environment variables or external credentials are requested (proportional). However, the skill will read other SKILL.md files and test prompts and intentionally transmit those to subagents/LLM runs for evaluation — this can leak any sensitive content included inside SKILL.md files (rare but possible). The SKILL.md does not include explicit safeguards for secrets in SKILL.md content.
Persistence & Privilege
always:false (good). The skill requests write access to the user's skills tree and performs git operations (branching, commits, revert, stash). That level of file-system write privilege is necessary for its function but is a material privilege: combined with autonomous invocation it could make many edits. The SKILL.md documents human-in-the-loop checkpoints which reduce risk, but real enforcement depends on the agent runtime honoring those pauses.
Assessment
This skill is coherent with its stated goal but exercises broad write/read access over your skills directory and will send SKILL.md contents to child agents/LLM runs. Before installing or running it: (1) back up your ~/.claude/skills (or wherever your skills live) or run in a clone/branch; (2) inspect SKILL.md files for any secrets or tokens (do not store credentials in SKILL.md); (3) run the optimizer on a small, non-critical subset first and review each git diff/commit before accepting; (4) prefer to keep autonomous invocation disabled or require manual confirmation in your agent runtime to ensure the human checkpoints in the SKILL.md are actually enforced; (5) confirm the LLM/data handling policy of your provider if you are concerned about sending skill contents to external services.Like a lobster shell, security has layers — review code before you run it.
autoresearchvk973smc223dna26rvr1ffvr4c184rqafdarwinvk973smc223dna26rvr1ffvr4c184rqaflatestvk973smc223dna26rvr1ffvr4c184rqafoptimizationvk973smc223dna26rvr1ffvr4c184rqafskillvk973smc223dna26rvr1ffvr4c184rqaf
License
MIT-0
Free to use, modify, and redistribute. No attribution required.