Distill Skill

Distill Skill

Mirrors the self-improving skill loop from Hermes Agent: after solving something non-trivial, decide whether to save the approach so future tasks skip the trial-and-error instead of re-deriving it from scratch.

When to propose a skill

Propose distillation when the just-finished task matches any of:

• Took several tool calls / multiple steps to get right (roughly 4+)
• Involved recovering from an error, a wrong turn, or a failed attempt before succeeding
• The user corrected your approach mid-task
• The user explicitly asked to save/remember/turn this into a skill

Skip silently (don't mention this check to the user) when:

• The task was a single simple edit, a question, or a read-only lookup
• The approach is a one-off unlikely to recur (e.g. a throwaway debug print)
• It would just duplicate an existing skill's content (see below)
• The procedure would require hardcoding secrets, credentials, or machine-specific paths

Workflow

1. Check for an existing skill. List .cursor/skills/*/SKILL.md and

~/.cursor/skills/*/SKILL.md. If one already covers this workflow, prefer patching it (add a pitfall, tighten a step, add/replace an example) over creating a duplicate.

1. Draft a short proposal first, not a file. Include:

• Proposed name (lowercase, hyphens, max 64 chars)
• One-sentence description (what it does + when to use it)
• 3-5 bullets covering what it would capture: the procedure, key

commands, pitfalls hit, and how success was verified

1. Ask the user to confirm with a compact question, e.g. "这个方法要不要

存成一个 skill，下次同类任务直接复用？" Do not write any file before the user agrees — a wrong or premature skill is worse than no skill.

1. On yes, create (or patch) .cursor/skills//SKILL.md following

the create-skill skill's conventions: YAML frontmatter with name/description, concise body, concrete steps/commands over prose, one level of progressive disclosure if the reference material is long.

1. On no or no response, drop it. Don't create the file and don't ask

again for this same task.

What to capture in the skill body

• The concrete procedure (steps, commands, code snippets) that actually

worked — not a narrative of the exploration that got there

• Pitfalls hit and how to avoid them next time (if a wrong approach was

tried first, note briefly why it failed)

• A verification step (command/test) that confirms the task succeeded
• Links to relevant files only when they are stable, checked-in paths — not

scratch or generated output files

Curating over time

When patching an existing skill instead of creating a new one, keep it tight: fold the new case in as an example or a bullet rather than appending a new section per occurrence. If a skill has drifted from how the codebase actually works, correct it in place rather than leaving stale instructions next to the fix.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JustinGastby
• Source: JustinGastby/Hermes-Mind-Skill
• License: MIT

Install and usage instructions live in the source repository linked above.