skill-generation

name	skill-generation
description	LLM-powered generation pipeline for Claude Code skills and CLAUDE.md — doc-init command, prompt system, context building, and output parsing
triggers	{"files":["src/commands/doc-init.js","src/lib/context-builder.js","src/prompts/**/*"],"keywords":["doc-init","generate skills","discovery agents","chunked generation","recommended"]}

You are working on aspens' skill generation pipeline — the system that scans repos and uses Claude/Codex CLI to generate skills, hooks, and instructions files.

Domain purpose

aspens doc init orchestrates a multi-step LLM pipeline that turns a scanned repo + import graph into a base skill, per-domain skills, and an instructions file (CLAUDE.md or AGENTS.md). Generation is always done in Claude-canonical format and transformed for other targets afterwards. The end product is what other coding agents (and aspens' own hooks) consume to stay grounded in the repo.

Critical files (purpose, not inventory)

• src/commands/doc-init.js — the pipeline orchestrator (backend → target → scan → graph → discovery → strategy → mode → generate → validate → transform → write → hooks → recommended extras → config)
• src/lib/runner.js — runLLM(), loadPrompt(), parseFileOutput(), validateSkillFiles() shared across all LLM-driven commands
• src/lib/skill-writer.js — writes parsed files, generates skill-rules.json, injects domain bash patterns, merges settings.json
• src/lib/skill-reader.js — parses skill frontmatter, activation patterns, keywords (consumed by skill-writer)
• src/lib/git-hook.js — installGitHook() / removeGitHook() for post-commit auto-sync (monorepo-aware)
• src/lib/timeout.js — resolveTimeout() for auto-scaled + user-override timeouts
• src/lib/target.js / src/lib/backend.js / src/lib/target-transform.js — target/backend resolution and Claude→other-target transform
• src/prompts/ — doc-init.md, doc-init-domain.md, doc-init-claudemd.md, discover-domains.md, discover-architecture.md, plus partials/ (skill-format, preservation-contract, examples)

Key Concepts

• Pipeline steps: (1) detect backends (2) backend selection (3) target selection (4) scan + graph (5) existing docs discovery check (6) parallel discovery agents (7) strategy (8) mode (9) generate (10) validate (11) transform for non-Claude targets (12) show files + dry-run (13) write (14) install hooks (Claude-only) (15) recommended extras (save-tokens, agents, git hook) (16) persist config to .aspens.json
• Early config persistence: Target/backend config is written to .aspens.json before generation starts (after step 4), so a failed generation run still records the user's explicit target/backend choice. saveTokens from existing config is preserved. Final writeConfig at step 16 adds saveTokens from recommended install.
• --recommended flag: Skips interactive prompts with smart defaults. Reuses existing target config from .aspens.json. Auto-selects backend from target. Defaults strategy to improve when existing docs found. Auto-picks discovery skip when docs exist. Auto-selects generation mode based on repo size. Also installs save-tokens, bundled Claude agents, dev/ gitignore entry, and doc-sync git hook (step 15).
• Recommended extras (step 15): When --recommended and not --dry-run: calls installSaveTokensRecommended() from save-tokens.js (if Claude target), copies all bundled agent templates to .claude/agents/ (skips existing) via installRecommendedClaudeAgents(), adds dev/ to .gitignore, installs doc-sync git hook if not present. Summary lines printed after.
• Backend before target: Backend selection (step 2) happens before target selection (step 3). If both CLIs available, user picks backend first, then targets. Pre-selects matching target in the multiselect. With --recommended, backend is inferred from existing target config.
• Canonical generation: All prompts receive CANONICAL_VARS (hardcoded Claude paths: .claude/skills, skill.md, CLAUDE.md, .claude). Generation always produces Claude-canonical format regardless of target. Non-Claude targets are produced by post-generation transform via transformForTarget().
• Incremental writing (chunked mode): When mode === 'chunked' and not dry-run, generated files are written to disk as each chunk completes instead of waiting until the end. User is prompted to confirm incremental writes before generation starts. Helper functions: validateGeneratedChunk() validates and strips truncated files per chunk; buildOutputFilesForTargets() handles multi-target transform; writeIncrementalOutputs() deduplicates and writes changed files. Tracks written content via incrementalWriteState (contentsByPath + resultsByPath Maps). When incremental mode is active, post-generation validation/transform/confirm/write steps are skipped (already done per-chunk).
• parseLLMOutput with strict single-file fallback: Codex often returns plain markdown without <file> tags. parseLLMOutput(text, allowedPaths, expectedPath) only wraps tagless text as the expected file for true single-file prompts (exactly one exactFile in allowedPaths, no dirPrefixes). Multi-file prompts require proper <file> tags.
• Existing docs reuse: When existing Claude docs are found and strategy is improve, loadExistingDocsContext() inlines them as ## Existing Docs (improve these — preserve hand-written rules...) into the prompt. chooseReuseSourceTarget() decides whether Claude or Codex docs are the source. Supports cross-target reuse (e.g. Claude docs → Codex output).
• Domain reuse helpers: loadReusableDomains() tries loadReusableDomainsFromRules() first (reads skill-rules.json), falls back to findSkillFiles() with extractKeyFilePatterns() parsing ## Key Files blocks.
• Config persistence with target merging: Uses mergeConfiguredTargets() to avoid dropping previously configured targets. writeConfig now also persists saveTokens config from the recommended install.
• Hook installation: Only for targets with supportsHooks: true (Claude). installHooks() generates skill-rules.json, copies hook scripts, injects generated domain patterns into post-tool-use-tracker.sh via # BEGIN/END detect_skill_domain markers, merges settings.json (backs up existing to .bak).
• Git hook offer: With --recommended, git hook is auto-installed (no prompt). Without --recommended, interactive prompt offered. Detection looks for the marker string aspens doc-sync hook (<rel>) in .git/hooks/post-commit.
• Discovery agents: Two LLM calls run in parallel — discover-domains (hub files + domain clusters) and discover-architecture (hub files + ranked + hotspots). Findings are merged into discoveryFindings and parsed; domain-specific slices are injected into each domain prompt as ## Discovery Findings for {domain}.

Critical Rules

• Base skill + instructions file are essential — pipeline retries up to 2× with format-correction prompts when <file> tags are missing. Domain skill failures are acceptable (user retries with --domains).
• improve strategy preserves hand-written content — LLM must read existing skills first and not discard human-authored rules. The preservation-contract partial enforces this in every prompt.
• Discovery runs before user prompt — domain picker shows discovered domains, not scanner directory names. Falls back to scanner domains if discovery fails.
• PARALLEL_LIMIT = 3 — domain skills generate in batches of 3 concurrent calls. Base skill always sequential first. Instructions file always sequential last.
• CliError, not process.exit() — all error exits throw CliError; cancellations return early.
• --hooks-only is Claude-only — hardcoded to TARGETS.claude regardless of config.
• Incremental write deduplication — writeIncrementalOutputs() skips files whose content hasn't changed since last write, using contentsByPath Map for tracking. Directory-scoped AGENTS.md files (path ends with /AGENTS.md but not the root AGENTS.md) go through writeTransformedFiles(), all others through writeSkillFiles().
• Read-only LLM tools — generation calls always pass allowedTools: ['Read', 'Glob', 'Grep']. The LLM explores the repo itself; aspens never lets it write.
• CLAUDE.md post-processing — generated instructions files are run through syncSkillsSection() and syncBehaviorSection() (and ensureRootKeyFilesSection() as a backwards-compat stripper of legacy ## Key Files blocks) so aspens owns the Skills list and Behavior block deterministically; prompts explicitly forbid the LLM from emitting these sections.

References

• Prompts: src/prompts/doc-init*.md, src/prompts/discover-*.md
• Partials: src/prompts/partials/skill-format.md, src/prompts/partials/preservation-contract.md, src/prompts/partials/examples.md

---

Last Updated: 2026-05-11