| name | agent-skill-best-practices |
| description | Research-backed quality framework for creating and upgrading SKILL.md-based agent capabilities. Use when designing new skills, auditing existing skills, reducing context bloat, improving trigger precision, adding safe tool boundaries, or migrating skills across Codex, Claude Code, and Copilot-style environments. |
Agent Skill Best Practices
Overview
Use this skill to engineer better skills with measurable quality gates: precise activation, efficient context usage, executable workflows, and safe operational boundaries.
This skill is intended for skill authors and reviewers.
Research-Backed Standards
Apply these baseline standards before any custom style decisions:
- Keep
SKILL.md lean and move deep detail to references/.
- Prefer progressive disclosure (metadata -> body -> resources on demand).
- Keep reference depth shallow (link directly from
SKILL.md, avoid deep chains).
- Use imperative instructions with explicit completion conditions.
- Treat scripts and command guidance as production code with risk controls.
Load citations in references/evidence-2026.md when needed.
Workflow
- Determine mode.
create: build a new skill.
upgrade: improve an existing skill.
audit: score and report findings.
migrate: adapt to another agent platform.
- Define target contract.
- Required behaviors.
- Trigger phrases/contexts.
- Supported platforms.
- Risk tolerance and execution boundaries.
- Run structured audit.
- Score with
references/audit-rubric.md.
- Record severity-ranked findings using
references/review-template.md.
- Implement fixes in strict order.
- Frontmatter precision.
- Body structure and progressive disclosure.
- Deterministic workflow and output contract.
- Security/governance boundaries.
- Cross-platform portability.
- Validate and finalize.
- Run
quick_validate.py where available.
- Return change summary, remaining risks, and next iteration items.
Non-Negotiable Rules
- No TODO placeholders in production skill files.
- No broad/vague description that can trigger everywhere.
- No monolithic SKILL body when detail can live in references.
- No unbounded command execution guidance for risky operations.
- No hidden platform assumptions (paths, tool availability, shell behavior).
Required Outputs Per Engagement
- Updated skill files.
- Findings list ordered by severity.
- Validation result.
- Residual risk statement.
References
Load only relevant references for the current job.
- Evidence and source links:
references/evidence-2026.md
- Scoring framework:
references/audit-rubric.md
- Frontmatter design:
references/frontmatter-guidelines.md
- Body design and progressive disclosure:
references/body-design-patterns.md
- Security and governance:
references/security-governance.md
- Portability/migration:
references/interoperability-migration.md
- Anti-patterns:
references/anti-patterns.md
- Review output format:
references/review-template.md