Auto Skill Distiller

Skill Distiller

Turn successful workflows into reusable skills — automatically.

> Inspired by Hermes Agent's learning loop, but with quality gates to prevent skill bloat.

When to Distill

Not every task deserves a skill. Evaluate these three criteria:

All three must be YES to proceed:

1. Novel? — Did this task require a workflow you haven't done before? (If you already have a skill for this, update it instead of creating a new one) 2. Successful? — Did the task complete with verified results? (Failed tasks produce lessons, not skills — write to memory/lessons-learned.md instead) 3. Reusable? — Will this exact workflow likely be needed again? (One-off tasks don't need skills)

Quick scoring:

Novel + Successful + Reusable = CREATE SKILL
Novel + Successful + One-off  = WRITE TO MEMORY (lesson learned, not a skill)
Novel + Failed                = WRITE TO LESSONS-LEARNED
Not Novel                     = UPDATE EXISTING SKILL (or skip)

Distillation Process

Step 1: Extract the Workflow

Look back at what you just did and identify:

Step 2: Generalize

Transform the specific instance into a reusable pattern:

Bad (too specific):

1. Read ch10-multi-agent-comm-patterns.md
2. Convert markdown to docx using python-docx
3. Upload to feishu folder nodcnxdXVfsiCVDuiigFVpnCPoc

Good (generalized):

1. Read source markdown file(s)
2. Convert to docx using python-docx (see references/docx-patterns.md)
3. Upload to target feishu folder

Step 3: Write SKILL.md

Generate the skill following the standard format:

---
name: 
description: ""
When to Use
<1-2 sentences on the trigger pattern>
Workflow
Key Decisions
Gotchas
References

Size target: SKILL.md body should be under 200 lines. If longer, split into SKILL.md (workflow) + references/ (details).

Step 4: Quality Check

Before saving, verify:

Step 5: Save and Register

Save to ~/.openclaw/skills//SKILL.md.

If the skill has reference materials, save them to ~/.openclaw/skills//references/.

After saving, verify the skill loads:

ls ~/.openclaw/skills//SKILL.md

Automatic Distillation Mode

When integrated with trinity-harness's Layer 3 (Compound), distillation happens automatically:

1. Task completes → Layer 3 Compound phase triggers 2. Evaluate Novel + Successful + Reusable 3. If all YES → run distillation process 4. If NO → write lesson to memory instead 5. Announce to user: "Distilled skill: . Review with read ~/.openclaw/skills//SKILL.md"

Never auto-distill silently. Always announce what was created so the user can review, edit, or delete.

Skill Maintenance

Update vs. Create

Before creating a new skill, check if a related one exists:

ls ~/.openclaw/skills/ | grep -i 

If a similar skill exists, update it (add the new pattern as a variant) rather than creating a near-duplicate.

Pruning

Periodically (during Dream Task), review skills:

Anti-Patterns

| Don't | Why | Do Instead | |---|---|---| | Distill every task | Skill bloat, noise drowns signal | Apply the 3-question gate | | Include conversation history | Wastes tokens, not reusable | Extract only the workflow pattern | | Write vague gotchas | "Be careful" helps no one | Specific: "API X returns 429 after 3 concurrent requests" | | Hardcode paths/names | Not portable | Use placeholders | | Skip quality check | Garbage skills waste future context | Always verify before saving |

Integration with Memory System

Distillation complements, not replaces, the memory system:

| Output | Goes to | When | |---|---|---| | Reusable workflow | ~/.openclaw/skills//SKILL.md | Novel + Successful + Reusable | | Lesson learned | memory/lessons-learned.md | Successful but one-off, or failed | | Quick note | memory/YYYY-MM-DD.md | Routine observations | | Core insight | MEMORY.md | Fundamental principle change |

<1-2 sentences on the trigger pattern>