Skill Routing
You do not need explicit /command invocation. Auto-select based on task signals:
| Signal |
Skill |
Rationale |
| "fix", "broken", "debug", error context |
/repair |
Debugging router (auto-detects web vs mobile) |
| "build", "implement", complex task |
/melt |
Autonomous execution with verification |
| "clean up", "tech debt", "slop" |
/burndown |
Debt elimination |
| "improve design/UX/perf" |
/improve |
Recursive improvement loop |
| "analyze", "think deeply", "evaluate" |
/heavy |
Multi-perspective analysis |
| "create video", "promo video", "hero video", "AI cinema" |
/video |
AI-cinematic video generation |
| "good morning", "briefing", "catch me up", "what did I miss" |
/briefing |
Morning digest across all platforms |
| "check discord", "discord messages", "discord DMs" |
/discord-digest |
Discord activity summary |
| "check X/twitter", "timeline", "bookmarks", "what's on X" |
/x-digest |
X timeline + Grok analysis |
| "reply on discord", "respond in channel", "send discord" |
/discord-reply |
Context-aware Discord reply |
| "tweet", "post on X", "draft thread" |
/x-post |
Voice-consistent tweet drafting |
| "engage on X", "find conversations", "grow audience" |
/x-engage |
Strategic X engagement |
| No clear task / research only |
No skill |
Just answer directly |
Parallelization Strategy
| Condition |
Strategy |
| Single focused task |
Single-agent execution |
| 2 independent work items |
Parallel Task() calls in a single message |
| 3+ independent work items with coordination needs |
TeamCreate with shared task list + SendMessage |
| Fire-and-forget research or verification |
Task() subagents (cheaper, simpler) |
| Agents need to share findings mid-work |
TeamCreate with peer-to-peer SendMessage |
Default to Task(). Escalate to TeamCreate when:
- Agents need to share intermediate findings mid-research (e.g., /heavy Deep cross-pollination)
- 3+ execution items need coordination on shared resources (e.g., /melt worker teams)
Trust skill-specific triage. If a skill's complexity assessment selects TeamCreate, follow it — the skill has determined that peer communication will produce higher quality results. Do not override with cost concerns.
Skill Fluidity
Skills are capabilities, not cages. If the task evolves, adapt:
- Building (/melt) and discover a critical bug? Fix it inline using /repair techniques. No formal mode switch needed.
- Debugging (/repair) and find the root cause is tech debt? Apply /burndown patterns to the area.
- Any task turns out to be architecturally complex? Use /heavy analysis for the sub-problem, then continue.
- Task decomposes into 3+ independent work items? Spawn a team (
TeamCreate) inline. No mode switch needed.
The autonomous-state.json mode field drives auto-approval and checkpoint enforcement. It does not constrain your cognitive approach. Use the best technique for each sub-problem regardless of which skill activated the session.
When to formally re-invoke a skill (via Skill tool):
- The ENTIRE task has shifted (not just a sub-problem)
- You need the full activation ceremony of another skill
When to just adapt inline:
- A sub-problem needs a different approach
- You discovered something that changes the next step but not the overall goal
1---2name: skill-routing-23description: The autonomous-state.json mode field drives auto-approval and checkpoint enforcement. It does not constrain your cognitive approach.4---5# Skill Routing67You do not need explicit `/command` invocation. Auto-select based on task signals:89| Signal | Skill | Rationale |10|--------|-------|-----------|11| "fix", "broken", "debug", error context | /repair | Debugging router (auto-detects web vs mobile) |12| "build", "implement", complex task | /melt | Autonomous execution with verification |13| "clean up", "tech debt", "slop" | /burndown | Debt elimination |14| "improve design/UX/perf" | /improve | Recursive improvement loop |15| "analyze", "think deeply", "evaluate" | /heavy | Multi-perspective analysis |16| "create video", "promo video", "hero video", "AI cinema" | /video | AI-cinematic video generation |17| "good morning", "briefing", "catch me up", "what did I miss" | /briefing | Morning digest across all platforms |18| "check discord", "discord messages", "discord DMs" | /discord-digest | Discord activity summary |19| "check X/twitter", "timeline", "bookmarks", "what's on X" | /x-digest | X timeline + Grok analysis |20| "reply on discord", "respond in channel", "send discord" | /discord-reply | Context-aware Discord reply |21| "tweet", "post on X", "draft thread" | /x-post | Voice-consistent tweet drafting |22| "engage on X", "find conversations", "grow audience" | /x-engage | Strategic X engagement |23| No clear task / research only | No skill | Just answer directly |2425## Parallelization Strategy2627| Condition | Strategy |28|-----------|----------|29| Single focused task | Single-agent execution |30| 2 independent work items | Parallel `Task()` calls in a single message |31| 3+ independent work items with coordination needs | `TeamCreate` with shared task list + `SendMessage` |32| Fire-and-forget research or verification | `Task()` subagents (cheaper, simpler) |33| Agents need to share findings mid-work | `TeamCreate` with peer-to-peer `SendMessage` |3435**Default to `Task()`.** Escalate to `TeamCreate` when:36- Agents need to share intermediate findings mid-research (e.g., /heavy Deep cross-pollination)37- 3+ execution items need coordination on shared resources (e.g., /melt worker teams)3839**Trust skill-specific triage.** If a skill's complexity assessment selects TeamCreate, follow it — the skill has determined that peer communication will produce higher quality results. Do not override with cost concerns.4041## Skill Fluidity4243Skills are capabilities, not cages. If the task evolves, adapt:4445- Building (/melt) and discover a critical bug? Fix it inline using /repair techniques. No formal mode switch needed.46- Debugging (/repair) and find the root cause is tech debt? Apply /burndown patterns to the area.47- Any task turns out to be architecturally complex? Use /heavy analysis for the sub-problem, then continue.48- Task decomposes into 3+ independent work items? Spawn a team (`TeamCreate`) inline. No mode switch needed.4950The `autonomous-state.json` mode field drives auto-approval and checkpoint enforcement. It does not constrain your cognitive approach. Use the best technique for each sub-problem regardless of which skill activated the session.5152When to formally re-invoke a skill (via Skill tool):53- The ENTIRE task has shifted (not just a sub-problem)54- You need the full activation ceremony of another skill5556When to just adapt inline:57- A sub-problem needs a different approach58- You discovered something that changes the next step but not the overall goal
tools-only