Skill Creation Guide
Help the user author effective agent skills — single SKILL.md files (in the
open Agent Skills format) that auto-trigger by context and provide specialized
guidance. The format is shared across hosts (Claude Code, Cursor, and others);
only the install location differs.
This skill covers three things: deciding whether a skill is worth creating, discovering candidates from the user's work, and generating a well-formed skill.
BE SKEPTICAL
Not every idea makes a good skill. Evaluate before creating:
Recurrence — Will this be used repeatedly?
- ✅ Weekly or more → create
- ⚠️ Monthly → reconsider, may not be worth it
- ❌ One-time → skip
Automation — Can it actually be automated?
- ✅ Clear, repeatable process → create
- ⚠️ Needs significant judgment each time → reconsider
- ❌ Too context-dependent, each case unique → skip
Value — Is the cumulative time saved significant?
- ✅ Meaningful savings per use → create
- ❌ Slower to invoke than doing manually → skip
Duplication — Does it already exist?
- ✅ Novel, unaddressed need → create
- ⚠️ Similar to an existing skill → enhance that instead
- ❌ Already covered → skip
Don't create skills for one-time events, tasks that differ fundamentally each time, sub-30-second manual tasks, or things the user will forget exist.
Workflow A — Discover opportunities
When the user asks what they should automate, analyze their work patterns first.
Requires the Glean MCP tools. If they aren't visible, ask the user to set up
Glean by prompting the harness (for example, Set up Glean for me).
- Gather context with Glean:
memory(roles, responsibilities, active projects),user_activity(past ~2 weeks), andsearchfor process docs ("runbook OR checklist OR process owner:me"). - Spawn the
work-pattern-analyzeragent to identify repeated queries, frequent contexts, workflow sequences, and manual processes — and to vet each candidate against the recurrence/value tests above. - Present vetted recommendations grouped by cumulative value, and show the rejected candidates with reasons. A few high-quality candidates beat many weak ones — "no automatable patterns" is a valid outcome.
See using-glean/reference/ for the param shape of memory, user_activity,
and search.
Workflow B — Generate a skill
When the user wants to create a specific skill:
- Clarify (only what isn't already clear): purpose, trigger conditions, tools it uses, and output format.
- Spawn the
skill-generatoragent with the concept, requirements, and available-tools context. It returns a completeSKILL.md— or declines, with reasons, if the idea fails the skepticism tests. - Offer to save it to the user's host skills directory, or display it for review. Confirm the path and the trigger phrases that will activate it.
SKILL.md structure
---
name: skill-name-in-kebab-case
description: When this skill triggers — specific phrases, contexts, and use cases. Fold trigger phrases into this field; keep it thorough but under ~200 words.
---
# Skill Title
Brief overview of what this skill does.
## When This Applies
- Condition / trigger phrases
## Main Content
[The workflow, guidance, or instructions — imperative form]
## Output Format (optional)
[Template for what the skill produces]
Best practices
- Specific triggers. "Use when reviewing pull requests, or when the user says 'review this PR' / 'check my code'" beats "use for code review."
- Progressive disclosure. Essential action first; detail and edge cases
below. Push deep, load-on-demand material into
reference/*.md. - Actionable, imperative content. "Search for X", not "Searches for X."
- Name the tools the skill uses (e.g. Glean
search,memory; orGrep,Read). - Stay host-agnostic. Don't hard-code one host's slash-command syntax or install paths; describe the capability and let each host resolve it.
Where to save
Skills live in the host's skills directory — e.g. ~/.claude/skills/<name>/
(personal) or .claude/skills/<name>/ (project) for Claude Code, the equivalent
location for other hosts, or skills/<name>/ when authoring for a plugin
library like this repo.
Related skills
work-pattern-analyzer/skill-generatoragents back the two workflows above.