Auto-Invoke

Apply the right skill automatically. Do not wait to be asked — recognize the situation and act.

Skill Routing Table

Situation	Skill to apply	When exactly
Building document/semantic search	rag	At task start, before writing any retrieval code
Changing a prompt or model	eval	Before the change ships; write eval first
Tool call or phase fails	self-heal	Immediately — diagnose before any retry
Any error, test failure, broken build	debug	Before touching code
Context ≥ 60% of limit	context	Proactively — don't wait for degradation
Session ending or key decision made	memory	At session end; also after any surprising gotcha
Before PR touching auth/payments/data	secure	Before git push, not after
Before claiming any phase done	verify	Every time, no exceptions
Picking model/agent for a task	route	At task assignment, before delegating
Long autonomous run	loop	At the start; defines phases and exit conditions
Coordinating multiple specialized agents	multi-agent	When task requires >1 agent with dependencies
Choosing or configuring an inference server	model-serving	Before selecting vLLM/TGI/Ollama or quantization
Proactive context cleanup at 45% capacity	context-hygiene	After completing a large phase or at 45-50% context

Decision Protocol

When you encounter a new task or event, ask in this order:

1. Is there an error? → self-heal (recovery) or debug (root cause)
2. Does it involve LLM output or prompts? → eval (before shipping)
3. Does it involve document retrieval or vector search? → rag (pipeline)
4. Is context growing large? → context (compress now)
5. Is the session ending? → memory (persist decisions)
6. Is security-sensitive code changing? → secure (checklist)
7. Is a phase or PR about to be marked done? → verify (evidence)
8. Coordinating multi-agent work? → multi-agent (DAG + failure handling)
9. Choosing an inference server? → model-serving (vLLM/TGI/Ollama decision tree)

Rules

• Skills are automatic responses to situations, not optional enhancements
• Apply verify before EVERY PR — no exceptions, no shortcuts
• Apply memory at EVERY session end — do not let decisions live only in conversation
• Apply self-heal BEFORE retrying any failed tool call — diagnose first
• Apply eval BEFORE any LLM feature ships — write the measurement first
• Never apply multiple skills simultaneously; sequence them (heal first, then verify, then ship)

What NOT to auto-invoke

• plan — only when starting a new multi-phase task, not on every subtask
• research — only when genuinely uncertain about external state (library, API, tool)
• orchestrate — only when spawning subagents for parallel work
• rag — only when the task involves external document retrieval, not general coding

Usage

This skill itself is rarely invoked directly. It is loaded by the orchestrator at session start to calibrate routing decisions. Individual skills should be read when the routing table above points to them.