Self-Improving Agent Learning Log
Paste the following prompt into your AI chat to install this skill:
Follow https://skillhub.cn/install/skillhub.md to install @user_15292d5a/yjkj-self-improving-agent.
About this skill
What problem it solves
When coding agents repeat the same command and hit the same error, or users repeatedly correct a convention, that knowledge often stays in one session. This skill writes errors, corrections, knowledge gaps, and feature requests into Markdown logs so future agent runs can learn from past mistakes.
How it works
On first use, it initialises a .learnings/ directory and routes entries into ERRORS.md, LEARNINGS.md, and FEATURE_REQUESTS.md. ERRORS.md is for non-zero exit codes, exceptions, timeouts, and external tool failures; LEARNINGS.md is for user corrections, knowledge gaps, and best practices; FEATURE_REQUESTS.md captures missing capabilities requested by the user. Each entry can carry an ID, status, priority, area tag, context, and remediation notes. Similar entries are linked with See Also, and recurring patterns accumulate a Recurrence-Count. When a learning becomes broadly applicable, it is distilled into a concise rule and promoted to agent context files such as CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md.
Boundaries and cautions
It is useful for project workspaces that need persistent agent guidance, but it does not replace documentation review or automated tests. Logs should avoid secrets, tokens, private keys, and full configuration files unless the user explicitly asks. Hook reminders, error detection, and cross-session sharing are optional and should be enabled only in controlled environments.
Use Cases
- When an agent repeatedly hits install failures, capture the failing command, context, and fix in an error log for later review by other sessions.
- When a user corrects an agent's assumption, record the correction, evidence, priority, and category in a learning log and status.
- When a workspace convention causes repeated failures, distill the verbose note into a concise project rule and promote it to agent files.
- When tool gotchas recur, search existing entries, increment recurrence, and link related entries before escalating in the project learning log.
Best For
- Engineers maintaining Codex workspaces who want recurring agent mistakes distilled into durable agent context.
- Platform engineers configuring OpenClaw workspaces who need behavioral rules and tool gotchas saved to workspace memory.
- Repository owners managing Copilot instructions who want project conventions synced into GitHub Copilot context files.
- Agent developers sharing cross-session experience who need sanitized learning summaries sent to other sessions, not raw transcripts.
Related Skills
A guide for writing Seedance 2.0 video prompts, covering @ references, camera moves, time segments, audio direction, and common pitfalls.
Collects daily conversations, task logs, and errors, then runs a nine-module review to surface blind spots, generate actions, and update memory.
Local long-term memory for AI agents that appends key facts and daily logs, supports cross-session context and search recall, and auto-archives old diaries to gzip files.
A hybrid-evolution agent for OpenClaw developers that offers tiered-confirmation and rollback-safe capability optimization.