Skills

Progressive Disclosure

Skills are selected from request intent. Manual skill selection is usually unnecessary.

Each skill uses a token-optimized two-layer design:

Layer	Content	Size
`SKILL.md`	Identity, routing conditions, core rules	~40 lines (~800B)
`resources/`	Execution protocols, examples, checklists, playbooks, snippets, tech stack	Loaded on-demand

This achieves ~75% token savings on initial skill loading (3-7KB → ~800B per skill).

Common resources deduplicated across all skills:

Resource	Purpose
`reasoning-templates.md`	Structured fill-in-the-blank templates for multi-step reasoning
`clarification-protocol.md`	When to ask vs. assume, ambiguity levels
`context-budget.md`	Token-efficient file reading strategies per model tier
`context-loading.md`	Task-type to resource mapping for orchestrator prompt construction
`skill-routing.md`	Keyword-to-skill mapping and parallel execution rules
`difficulty-guide.md`	Simple/Medium/Complex assessment with protocol branching
`lessons-learned.md`	Cross-session accumulated domain gotchas
`verify.sh`	Automated verification script run after agent completion
`api-contracts/`	PM creates contracts, backend implements, frontend/mobile consumes
`serena-memory-protocol.md`	CLI mode memory read/write protocol
`common-checklist.md`	Universal code quality checks

Each skill provides domain-specific resources:

Resource	Purpose
`execution-protocol.md`	4-step chain-of-thought workflow (Analyze → Plan → Implement → Verify)
`examples.md`	2-3 few-shot input/output examples
`checklist.md`	Domain-specific self-verification checklist
`error-playbook.md`	Failure recovery with "3 strikes" escalation rule
`tech-stack.md`	Detailed technology specifications
`snippets.md`	Copy-paste ready code patterns

This keeps initial context lean while still supporting deep execution when required.