# Skill: Skill Incubator

**Agent:** Devin (Delivery Ops)
**Trigger:** After completing any client deliverable, or during weekly skill-incubation-review cron
**Output:** Updated `knowledge/areas/skill-tracker.md` + new skill files when graduated

---

## Purpose

Evaluate completed client deliverables for reusable skill extraction potential. When a deliverable passes the reusability threshold, extract it into a deployable skill file that can be offered to future clients — enabling faster delivery and higher margins.

## Process

### Step 1: Evaluate Deliverable

For each completed deliverable, answer these questions:

1. **Parameterizable?** Can this automation work for different clients by changing inputs (API keys, ICP filters, CRM credentials)?
2. **Agent-executable?** Can an AI agent run this end-to-end without human intervention?
3. **Clear I/O?** Are the inputs and outputs well-defined (e.g., "input: ICP filters → output: enriched leads in CRM")?
4. **Repeatable demand?** Have we delivered this type of automation before, or is there clear market demand?

**Scoring:**
- 4/4 = Strong candidate → proceed to extraction
- 3/4 = Promising → log as candidate, revisit after next similar delivery
- 2/4 or below = Not reusable → log and move on

### Step 2: Log in Skill Tracker

Update `knowledge/areas/skill-tracker.md`:

- Add entry to **Completed Deliverables** log with: client (anonymized), automation type, date, tools used
- Increment the **Category Tally** for the automation type
- If 3/4+ score: add to **Skill Candidates** section with status "Candidate"

### Step 3: Check Graduation Threshold

A skill candidate **graduates** when:
- Same automation type has been delivered **3+ times**
- The core logic is consistent across deliveries (not heavily customized each time)
- A documented SOP or pattern exists

### Step 4: Extract Skill

When graduating a candidate:

1. Create new skill directory: `~/.openclaw/skills/{skill-name}/`
2. Create `skill.md` with:
   - **Name and description**
   - **Required inputs** (API keys, configuration parameters)
   - **Steps** (numbered, agent-executable instructions)
   - **Expected outputs**
   - **Error handling** (common failures and recovery)
   - **Testing checklist**
3. If the skill requires tool configurations, add `tool-config.json`
4. Update skill-tracker.md: move from "Candidate" to "Incubating"

### Step 5: Test and Validate

1. Run the skill end-to-end in a sandbox/test environment
2. Verify outputs match expected results
3. Test with different input parameters (at least 2 variations)
4. Fix any issues found
5. Update skill-tracker.md: move from "Incubating" to "Graduated"

### Step 6: Deploy

1. If the skill should run on a schedule, propose a cron job to Alfred
2. Update the relevant agent's SERVICES.md with the new skill
3. Update skill-tracker.md: move from "Graduated" to "Deployed"
4. Report deployment to Alfred via Telegram

## Automation Categories

Track these categories in the skill tracker:
- Speed-to-Lead
- Review Automation
- Appointment Scheduling
- Lead Qualification
- CRM Automation
- Reporting Dashboard
- Content Repurposing
- OpenClaw Setup
- Invoice/Billing
- Client Onboarding
- Testimonial Collection
- Competitor Scanning
- Data Enrichment

## Anti-Patterns

- Do NOT extract a skill after only 1 delivery — wait for the pattern to emerge
- Do NOT create skills that require heavy per-client customization — those are services, not skills
- Do NOT skip testing — an untested skill deployed via cron can cause client-facing issues
- Do NOT forget to update skill-tracker.md — the tracker IS the system of record