Recursive Self Improvement
SuspiciousAudited by ClawScan on May 10, 2026.
Overview
This instruction-only skill is framed as self-improvement, but it asks the agent to recursively and automatically change, refactor, and optimize systems without clear approval gates, scope limits, rollback, or stopping conditions.
Treat this as a high-impact automation prompt, not as a safe background optimizer. Use it only in a controlled project, with backups or version control, explicit approval before edits, a small maximum number of iterations, and manual review of every proposed repair or refactor.
Findings (4)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
If invoked with file-editing or development tools available, the agent could make broad code or logic changes beyond what the user intended.
The repair workflow explicitly includes making code or logic changes. The skill does not define target paths, approval gates, rollback expectations, or limits on what the agent may modify.
4. 代码/逻辑变更 ... 5. 单元测试 ... 6. 集成测试
Only use this skill in a sandbox or repository with version control, and require explicit user approval before any code, configuration, or workflow changes are applied.
The agent may keep auditing, repairing, or optimizing after the immediate user request is satisfied, potentially consuming resources or making repeated changes.
The workflow directs recursive re-entry after validation. This creates an autonomous loop without an explicit maximum iteration count, user checkpoint, or stop condition.
验证通过 → 递归调用（回到阶段1）
Add hard stop conditions, maximum iteration counts, and user confirmation before each recursive cycle or material change.
A bad repair or optimization decision could be repeated or applied in parallel, causing wider project damage than a single manual edit.
The skill encourages concurrent execution and dynamic scheduling while also describing repair and refactoring. Without containment boundaries, concurrent automated changes could amplify mistakes across multiple tasks or modules.
系统支持多任务并发执行 ... 动态调整并发数
Run one change at a time by default, isolate changes in branches or temporary workspaces, and require tests plus user review before merging.
Past run history or incorrect observations could affect future recommendations or changes.
The skill describes adaptive learning from prior executions and keeping optimization history. This is purpose-aligned, but the artifacts do not specify where that history is stored, how long it is retained, or how it is prevented from influencing unrelated future tasks.
从执行中学习，持续优化 ... 学习内容：任务执行成功率、性能瓶颈识别、模式识别
Keep learning records local, reviewable, and scoped to a specific project; provide a way to clear or ignore prior history.