Skill Best Practices
Write Skills that activate at the right time, spend context carefully, and produce repeatable results.
Write a trigger-focused description
The description is the catalog entry agents see before activation. It should describe the user intent, task shape, and boundaries, not just the topic.
# Weak description: Helps with PDFs. # Stronger description: Extract text from PDFs, split or merge pages, fill forms, and validate generated PDF outputs.
Spend context wisely
Once activated, every word in SKILL.md competes with the user's request, project context, tool results, and other active instructions. Put the high-value procedure in the main file and move rare details to references.
Activation rules, workflow, success criteria, defaults, and critical gotchas.
Long API references, exhaustive examples, style guides, schemas, and large templates.
Tell the agent when to read each reference, script, or asset.
Favor procedures over declarations
Agents follow operational steps better than broad principles. Replace "be careful with migrations" with a checklist the agent can execute.
List the files, fields, or signals the agent must examine before acting.
Describe the allowed actions and defaults for uncertain cases.
Specify tests, assertions, review checks, or output criteria.
Use templates for output
When the final answer has a required format, include a concrete template. Agents pattern-match against examples more reliably than prose descriptions.
## Final report Summary: - ... Findings: | Severity | Issue | Evidence | Recommendation | | --- | --- | --- | --- | Open questions: - ...
Bundle scripts carefully
Scripts are useful for deterministic work such as parsing files, extracting data, validating outputs, or generating reports. Design them for agents, not interactive humans.
--help so agents can inspect usage quickly.SKILL.md or compatibility metadata.