Skill Dispatcher — Universal Skill Router

§0 Configuration (READ FIRST)

Before this skill works, you MUST fill the routing table with YOUR skills:

1. Scan \x3Cavailable_skills> in your system prompt
2. Run scripts/coverage_check.py \x3Cskill1> \x3Cskill2> ... — lists uncovered skills
3. Fill each uncovered skill into the routing table below at \x3C!-- TODO --> markers
4. Add 2-5 trigger keywords per skill
5. Set frequency (⭐/⭐⭐/⭐⭐⭐) based on how often the skill is used
6. Re-run coverage check — should show "All skills covered"

Trigger keyword rules:

• Use phrases the user would naturally say
• Avoid single common words (e.g. "write", "search")
• Test: "If the user said ONLY this phrase, would it match?"

⚡ Direct Hits (FIRST — 2 seconds)

If user message contains an exact keyword match, skip Q1/Q2 and route directly:

If NOT a direct hit → proceed to decision tree.

🔴 Quick Decision Tree

Q1: Your primary domain?
    ├─ \x3C!-- TODO: add your most common task categories -->
    ├─ category_A / trigger_A → skill_A
    ├─ category_B / trigger_B → skill_B
    ├─ category_C / trigger_C → skill_C (chain if needed)
    └─ ...

    ⚠️ CHAIN RULE: "→" = read 1st skill, execute, THEN read 2nd.
    ⚠️ NO chain for non-multi-step tasks.

Q2: Anything else?
    → Scan Full Routing Table below.
    → No match? Proceed normally.

Q3: Enhancer layer?  ← auto-apply, not user-requested
    After routing to primary skill/tool, check if output needs enhancement:
    ├─ 生成图/图表/UI/配色 → aesthetic, Color Palette Generator, data-viz-palette
    ├─ 生成大段文字/解释   → humanizer
    ├─ 上下文/文件膨胀      → token-optimizer
    ├─ 复杂多步任务         → agent-planner (先规划再执行)
    ├─ 被用户纠正/发现错误  → Self-Improving + Proactive Agent
    ├─ 设计skill/prompt    → context-engineering, prompt-engineer
    └─ 无增强需求           → skip

    ⚠️ 🔴 MANDATORY: You MUST open the matched enhancer's SKILL.md.
    ⚠️ At minimum: read the first section. Skipping = execution failure.

Q4: Background health?  ← session-state triggers (every ~10 exchanges)
    ├─ Context usage >50% or >10 exchanges → token-optimizer
    ├─ Recent tool failures/timeouts       → healthcheck
    ├─ Bulk file create/delete/reorg       → lobster-workspace-guardian
    └─ All clear                           → skip

    💡 Check with: session_status (📊 session_status tool)
    💡 Not every message — every ~10 exchanges or when things feel off.

🚨 Anti-Forget Gates (backup — when dispatcher NOT loaded)

These are your belt-and-suspenders. If the dispatcher's description didn't match and you're about to take one of these actions raw, this section catches it. If the dispatcher IS loaded, the decision tree already handled routing — skip this.

🔼 Enhancer Skills (Q3 layer — auto-triggered by output type)

These are NOT standalone skills. They overlay on top of a primary skill/tool. User doesn't need to ask for them — check automatically when output matches:

📋 Full Routing Table

\x3C!-- TODO: Category 1 — your primary domain -->

\x3C!-- TODO: Category 2 — memory/knowledge -->

\x3C!-- TODO: Category 3 — browser/desktop -->

\x3C!-- TODO: Category 4 — content/search -->

\x3C!-- TODO: Category 5 — dev/tools -->

\x3C!-- TODO: Category 6 — workflow -->

\x3C!-- TODO: Category 7 — design/creative -->

\x3C!-- TODO: Category 8 — system/meta -->

🔄 Fallback Protocol + Coverage Gap Detection

If NO routing table match but you SUSPECT a skill exists:

1. Scan the \x3Cavailable_skills> list in current context
2. Coverage check: Compare every skill in \x3Cavailable_skills> against names in the routing table + Direct Hits.
   • If a skill exists in \x3Cavailable_skills> but NOT in this dispatcher → 🚨 uncovered skill detected → trigger New Skill Onboarding
3. If still unsure → read references/skill-capabilities.md for full capability details
4. If the skill doesn't exist locally → search Clawhub or other registries
5. If still nothing → proceed with built-in tools

🆕 New Skill Onboarding Protocol

When a new skill is detected, follow the full 6-step protocol: → references/onboarding.md — Detection → Classify → Draft → 3-Round Review → Commit → Log

🔗 System Integration

• Your agent workspace files (SOUL.md, AGENTS.md, TOOLS.md) should reference this dispatcher
• Add a line to your core rules: "每次回复前先问哪个 skill 能处理这个任务？路由细节见 skill-dispatcher"
• The system prompt already instructs "scan available_skills, read matching SKILL.md". The dispatcher makes this scan systematic.

📊 Self-Correction Protocol + Cortex Feedback Loop

When you realize you forgot a skill

1. Check if this is a repeat — search memory for prior forgets of this skill
2. If first time → no log needed. Just fix the immediate task.
3. If 2nd+ time → write a correction entry with: date, skill, task, root cause
4. Always fix the immediate task — don't wait for logging.

When you re-read this dispatcher

• Search memory for recent forget patterns
• If any repeated forgets → upgrade their frequency or add to Direct Hits

Why 2-strike threshold?

Writing memory on every single forget creates noise. Only repeated patterns persist.

📖 References

• skill-capabilities.md — Full capability descriptions for reference (fill with YOUR skills)
• onboarding.md — 6-step protocol for integrating new skills

📋 Covered Skills Registry

\x3C!-- Keep this list in sync with routing table + Direct Hits. Used for coverage gap detection. --> \x3C!-- Last sync: [DATE] --> \x3C!-- Fill with your skill names: skill1 skill2 skill3 ... -->