self update skills

Agent skills rot as APIs change, new features launch, or workflows fail. Most agents either ignore these drifts or make silent, destructive edits that break your production pipelines. This skill provides a controlled, evidence-backed maintenance framework to keep your agent library current without risking regressions.

What it does

• Audit existing skills by checking for stale citations, dead links, and outdated metadata against primary sources.
• Identify update triggers based on official API changes, security advisories, or reproducible execution failures.
• Establish authoritative scope to ensure edits stay within writable boundaries and preserve user-authored rules.
• Propose gated changes through a structured approval process that includes source links, risk assessments, and rollback plans.
• Validate and compare candidate versions against baselines to verify correctness and token efficiency before final deployment.

How it works

1. Identify a trigger such as a manual maintenance request or a discovered API deprecation.
2. Research primary sources to verify claims and document the specific evidence justifying a change.
3. Present a numbered proposal detailing the exact files, behavioral differences, and regression risks for human approval.
4. Apply non-destructive updates using a narrow patch approach that synchronizes instructions, tests, and examples.
5. Run structural validation and compare against the previous version to ensure no safety or performance regressions occurred.

Frameworks & tools

This skill is framework-agnostic and works with any SKILL.md or YAML-based agent definition. It utilizes version control systems for rollbacks and relies on primary documentation sources for verification.

Why this beats prompting it yourself

Generic prompts often lead to "style churn" or silent breaking changes where the agent rewrites instructions without proof of improvement. This skill enforces a non-negotiable contract that requires evidence, baseline comparisons, and explicit human gates for every single write operation.

Use cases

• Modernize agent libraries when a platform like OpenAI or Anthropic releases new tool-calling schemas.
• Fix recurring failures by identifying and documenting missing edge cases discovered during production traces.
• Maintain security hygiene by updating skills in response to new library vulnerabilities or policy changes.
• Prevent instruction drift by consolidating duplicated guidance and pruning contradictory rules across a multi-agent system.

Known limitations

Does not support unattended or autonomous rewriting. Requires access to version control or a backup mechanism for rollbacks. Performance is dependent on the quality of provided primary sources.