skill-injector

Skill Injector - Smart Skill Selection

Maps task signals to skills from the everything-claude-code catalog and injects only the relevant ones into each agent's prompt. Prevents context bloat by being selective.

When to Activate

• During task routing (called by task-router skill)
• When an agent needs additional skill guidance mid-task

Skill Selection Algorithm

Input

• Task classification (type + complexity) from task-router
• Agent role being dispatched
• Project stack from crew-profile.md
• Skill catalog from crew-skills.json

Selection Rules

Rule 1: Always-Active Skills (injected into every agent) These are universal best practices (kept minimal for token efficiency):

• context-budget — monitor token consumption (subsumes strategic-compact)
• verification-loop — verify before claiming done
• coding-standards — follow project conventions

Note: iterative-retrieval moved to bug-fix task-type skills. search-first moved to feature task-type skills. strategic-compact dropped (redundant with context-budget).

Rule 2: Stack-Specific Skills (based on detected languages)

| Stack | Skills | |-------|--------| | JavaScript/TypeScript | typescript-patterns, backend-patterns | | React | react-patterns, react-testing | | Next.js | nextjs-patterns | | Vue | vue-patterns, nuxt4-patterns | | Python | python-patterns | | Django | django-patterns, django-security, django-tdd | | Flask | python-patterns | | Go | golang-patterns | | Rust | rust-patterns | | Java | java-patterns, springboot-patterns | | Kotlin | kotlin-patterns, kotlin-testing | | C++ | cpp-coding-standards, cpp-testing | | C# | csharp-testing, dotnet-patterns | | Flutter/Dart | dart-flutter-patterns | | Swift | swiftui-patterns, swift-concurrency-6-2 | | PHP/Laravel | php-patterns, laravel-patterns |

Rule 3: Task-Type Skills (based on classified task type)

| Task Type | Skills | |-----------|--------| | bug-fix | iterative-retrieval, deep-research, codebase-onboarding | | feature | architecture-decision-records, agentic-engineering, blueprint, search-first | | refactor | architecture-decision-records, backend-patterns | | test | tdd-workflow, e2e-testing, ai-regression-testing | | security | security-scanning | | docs | article-writing, documentation-lookup | | devops | deployment-patterns, docker-patterns, database-migrations, git-workflow | | review | verification-loop | | performance | cost-aware-llm-pipeline |

Rule 4: Agent-Role Skills (specific to the dispatched agent)

| Agent | Additional Skills | |-------|------------------| | ceo | agentic-engineering, architecture-decision-records | | vp-engineering | architecture-decision-records, hexagonal-architecture | | vp-quality | verification-loop, tdd-workflow | | senior-dev | (stack skills + task skills) | | junior-dev | coding-standards only (keep prompt small for Haiku) | | code-reviewer | verification-loop, security-scanning | | debugger | iterative-retrieval, deep-research | | security-analyst | security-scanning | | devops-engineer | deployment-patterns, docker-patterns | | doc-writer | article-writing | | test-engineer | tdd-workflow, e2e-testing |

Output: Skill Injection Payload

For each agent being dispatched, generate a skills section:

## Injected Skills for this Task

### From: {skill-name}
{Condensed key guidance from the skill - NOT the full SKILL.md, just the most relevant rules and patterns for this specific subtask}

### From: {skill-name-2}
{Condensed guidance}

Token Budget per Agent

To prevent context bloat from too many skills:

| Agent Model | Max Skill Tokens | Max Skills | |-------------|-----------------|------------| | Opus (CEO) | 3000 tokens | 6 skills | | Sonnet (leads/senior) | 2000 tokens | 4 skills | | Sonnet (planner in lite mode) | 2500 tokens | 5 skills | | Haiku (junior/devops/docs) | 800 tokens | 2 skills |

Lite mode note: When --lite flag is used, no Opus agents are dispatched. The vp-engineering agent takes over planning duties and gets an expanded budget (2500 tokens, 5 skills) to compensate for the extra architectural context it needs.

If selected skills exceed the budget:

1. Prioritize: always-active > task-type > stack > agent-role
2. Truncate to key rules only (skip examples, anti-patterns)
3. Drop lowest-priority skills

Skill Content Condensation

When injecting a skill, don't paste the entire SKILL.md. Extract only:

1. The core rules/principles (2-5 bullet points)
2. The most relevant pattern for this specific task
3. Any critical anti-patterns to avoid

Example condensation:

Full SKILL.md: 200 lines, ~2000 tokens
Condensed injection: 15 lines, ~150 tokens

This keeps agent prompts focused and within token budget.