Skill Integration

Skill-Creator Integration

Manages skill loading, observation, and bounded learning for the GSD ecosystem. Ensures learned knowledge persists across sessions and subagent contexts.

Skill Loading Before GSD Phases

Before executing any GSD phase, load relevant generated skills:

1. Check .claude/commands/ for project-level skills
2. Check ~/.claude/commands/ for user-level skills
3. Project-level skills take precedence over user-level on conflict
4. Load only skills relevant to the current phase and task
5. Respect the token budget: 2-5% of context window maximum

Critical: When forking subagent contexts for GSD phases (execute-phase, verify-work), include relevant skills in the subagent's context. Clean context means free of stale conversation history -- not free of learned knowledge. A subagent that can't access learned skills is throwing away everything skill-creator has captured.

For full 6-stage pipeline details, see references/loading-protocol.md.

Token Budget Management

The 6-stage skill loading pipeline (Score, Resolve, ModelFilter, CacheOrder, Budget, Load) manages what gets loaded. When multiple skills compete for budget:

1. Phase-relevant skills get priority (e.g., testing skills during verify-work)
2. Recently activated skills rank higher than dormant ones
3. Project-level skills override user-level duplicates
4. If budget is exceeded, queue overflow skills and note what was deferred

Session Observation Protocol

During all work sessions, maintain awareness of patterns worth capturing:

• Tool sequences that repeat across sessions (e.g., git log -> read file -> edit -> test -> commit)
• File patterns consistently touched during specific GSD phases
• Commands run after test failures or verification issues
• Corrections the user makes to agent output -- these are the highest-signal observations
• Phase outcomes: success, failure, partial, and what was different

If .planning/patterns/ exists, record observations to sessions.jsonl in that directory. This data feeds skill-creator's pattern detection pipeline.

For observation taxonomy, signal strength ranking, and JSONL schema, see references/observation-patterns.md.

Bounded Learning Guardrails

These constraints are NON-NEGOTIABLE:

• Maximum 20% content change per skill refinement
• Minimum 3 corrections before a refinement is proposed
• 7-day cooldown between refinements of the same skill
• All refinements require user confirmation
• Skill generation never skips permission checks
• Agent composition requires 5+ co-activations over 7+ days

For rationale and enforcement details, see references/bounded-guardrails.md.

Phase Transition Hooks

After completing any GSD phase transition, check for skills that should trigger:

Run matching skills after the phase completes, not during -- avoid adding overhead to the critical path.

Skill Suggestions

At the start of each new session, check skill-creator for pending suggestions. If suggestions exist with high confidence (3+ pattern occurrences), briefly notify the user:

"skill-creator detected a repeating pattern: [brief description]. Run skill-creator suggest to review."

Never auto-apply suggestions. Always require explicit user confirmation. This is a core safety principle of the bounded learning system.

When to Suggest skill-creator

Suggest skill-creator when:

• You notice the same sequence of steps has occurred 3+ times
• The user corrects the same kind of output repeatedly
• A workflow is complex enough to benefit from codification
• The user asks "why do I keep having to tell you this?"

Suggest naturally:

I've noticed we run the same lint -> test -> fix cycle after every code change.
Want me to capture this as a skill so it happens automatically?

Auditing a Skill (counterfactual)

Pass-rate is blind to most skill effects — a skill can change behavior in hundreds of ways while pass-rate barely moves. When a skill is created, modified, or proposed for retirement, run the skill-counterfactual-audit skill before finalizing. It runs a paired probe (the same task once with the skill loaded and once without), segments and aligns both traces into goal-directed phases, and emits a SIP report (surface anchoring, template copy, excess planning, task recovery, off-task artifact). Treat its findings as the behavioral-impact signal pass-rate cannot give you: a low-impact skill is a retirement candidate; a high-negative-impact skill needs revision, not just a green test. This is the audit gate for the skill lifecycle this skill manages.

Anti-Patterns

• Do not execute GSD phases without checking for relevant learned skills first
• Do not auto-apply skill suggestions -- always require user confirmation
• Do not load skills that exceed the token budget -- defer and note what was skipped
• Do not ignore user corrections -- they are the primary signal for skill refinement