Reference Enrichment

Reference Enrichment Skill

Enrich an agent or skill's reference files from Level 0-2 to Level 3+, or decompose bloated body files by extracting domain content into references. Enrichment adds knowledge; decomposition moves knowledge to where progressive disclosure says it belongs. The enrichment pipeline runs five phases with explicit gates because each phase feeds the next — starting Phase 3 without Phase 2 research produces filler, not depth.

Workflow

Phase 0: DECOMPOSE

Goal: Extract domain-heavy content from a bloated SKILL.md or agent body into reference files.

When to use: When a component's body exceeds ~500 lines and contains catalogs, code examples, specification tables, or agent rosters that should live in references/ per PHILOSOPHY.md's progressive disclosure architecture.

Trigger: Invoke with --decompose argument, or when the request matches "decompose", "extract references", "slim down", "too long", or "move to references".

1. Run the detection script to identify extractable content:

   python3 scripts/detect-decomposition-targets.py --skill {name}
   

   (or --agent {name})

2. If no extractable blocks found, report "nothing to decompose" and stop

3. Save a snapshot of the original file: cp {path} /tmp/decomp-before-{name}.md

4. For each extractable block identified by the detection script: a. Read the content block and its surrounding context b. Determine the best reference filename:

   • Use the detection script's suggestion as a starting point
   • If a reference file with related content already exists, MERGE into it
   • Follow naming convention: references/{topic}.md (lowercase, hyphens) c. Create or update the reference file following references/reference-file-template.md d. Remove the content from the body (MOVE, not copy) e. Add a loading table entry in the body that maps task signals to the new reference file

5. Ensure the body retains:

   • YAML frontmatter
   • Brief overview paragraph
   • Phase workflow (phases, gates, decision points)
   • Loading table with entries for all reference files
   • Error handling section
   • References section

6. Validate the decomposition:

   python3 scripts/validate-decomposition.py \
       --before /tmp/decomp-before-{name}.md \
       --after {path} \
       --refs {refs_dir}/
   

7. If validation FAILS: restore from snapshot and report the failure. Do not proceed.

8. If validation PASSES: run structural checks:

   python3 scripts/validate-references.py --skill {name}  # or --agent {name}
   python3 scripts/audit-reference-depth.py --skill {name} --verbose  # or --agent {name}
   

Gate: Validation passes. Body line count reduced. All extracted content exists in reference files. Loading table entries exist for all new references.

Phase 1: DISCOVER

Goal: Identify which sub-domains are missing reference coverage.

1. Run the gap analyzer: python3 skills/meta/reference-enrichment/scripts/gap-analyzer.py --agent {name} (or --skill {name})
2. Read the component's .md file to understand its stated purpose, triggers, and domain claims
3. Read any existing reference files to map current coverage
4. Compare stated domains against covered domains to identify gaps

Output format:

DISCOVER: {name}
  Current level: {0-3}
  Existing references: [{filenames}]
  Stated domains: [{domains from description and body}]
  Gaps: [{sub-domains with no reference coverage}]
  Recommended files: [{filename} → {why}]

Gate: Gap report exists with at least one identified gap. If no gaps exist (Level 3 already), report and stop — over-generating creates noise, not signal.

Phase 2: RESEARCH

Goal: Compile concrete, domain-specific content for each gap.

For each identified gap:

1. Read existing Level 3 reference files in this repo as exemplars — golang-general-engineer's references/ is the benchmark: version-specific patterns, grep commands, error-fix mappings
2. Identify: version-specific patterns (what changed in version X.Y), common failure modes with detection commands (grep -rn "pattern" --include="*.ext"), error-fix mappings (error message → root cause → fix), project-specific conventions visible in the codebase

Dispatch up to 5 parallel research agents — one per sub-domain gap — because sequential research bottlenecks the pipeline. Each agent receives: the sub-domain, the component's .md as context, and a path to an exemplar Level 3 reference file.

Gate: Each gap has at least 10 concrete findings (version numbers, function names, grep patterns, code examples). Generic advice ("follow best practices") does not count toward this gate.

Phase 3: COMPILE

Goal: Assemble research into structured reference files.

For each gap, create one reference file following references/reference-file-template.md:

• One file per major sub-domain (not one monolithic file) because focused files are faster to load and easier to update as language versions change
• Max 500 lines per file (CLAUDE.md standard) — split into sub-topics if content exceeds this
• Include: overview paragraph, pattern table with version ranges, failure mode table with detection commands, error-fix mappings where applicable
• Match the tone of existing Level 3 references: direct, concrete, no hedging

Do-pairing rule (mandatory): Every failure mode block written during this phase must include a "Do instead" counterpart. If the retro learning or research does not carry enough information to write a concrete positive counterpart, omit the failure mode entirely rather than shipping it without the paired "Do instead". A bare negative block encodes no actionable knowledge and will fail structural validation. If a prohibition is a genuine absolute with no correct alternative, annotate it with <!-- no-pair-required: reason --> inline before the block.

Write files to: agents/{name}/references/ or skills/{name}/references/

Gate: Each generated file is between 80-500 lines. Run both checks:

python3 scripts/validate-references.py --agent {name}
python3 scripts/validate-references.py --check-do-framing

Both must exit 0 before proceeding to Phase 4.

After validation passes, process each generated reference file for density. Call the Skill tool with condense. Strip prose filler while preserving patterns, detection commands, and error-fix mappings.

Phase 4: VALIDATE

Goal: Confirm the reference files meet Level 3+ depth before integrating.

Tier 1 (Deterministic):

python3 scripts/audit-reference-depth.py --agent {name} --json

Verify the level field is 3 in the output. If still below Level 3, the files are too generic — return to Phase 2 for the weak sub-domain.

Tier 2 (LLM self-assessment): Read each generated file and apply the rubric from references/quality-rubric.md. Ask: would a reviewer using only this file produce Level 3 quality output? Concrete test: pick one failure mode from the file — does it include a grep command to detect it?

Gate: Both tiers pass. If Tier 2 fails for a specific sub-domain, loop back to Phase 2 for that gap only (not all gaps). Maximum 2 loops per gap before flagging for manual enrichment.

Phase 5: INTEGRATE

Goal: Wire the new references into the component so they are actually loaded.

1. Read the component's .md file
2. Add or update a loading table in the body — pattern from skills/meta/do/references/repo-architecture.md:

   | Task type | Load |
   |-----------|------|
   | {task}    | `references/{file}.md` |
   

3. Write the updated .md file
4. Run validation:

   python3 scripts/validate-references.py --agent {name}
   python3 -m pytest scripts/tests/test_reference_loading.py -k {name} -v
   

5. Stage all changes: git add agents/{name}/ skills/{name}/

Gate: Validation passes. Changes staged. Report the level change (was: N, now: M) and list each new file with its line count.

Reference Material

Load when the task type matches:

Reference Loading Table

Error Handling

Gap analyzer fails: The component may not exist in expected paths. Check both agents/ and skills/ directories, and ~/.claude/agents/ for deployed copies.

Phase 2 gate fails (fewer than 10 concrete findings): The domain may be too narrow or already well-documented upstream. Flag and suggest manual enrichment with project-specific production incidents rather than generic docs research.

Phase 4 Tier 1 still below Level 3 after compile: The files are too short or too generic. Read one file, apply the rubric directly, identify the weakest section, and target Phase 2 research at that section specifically.

validate-references.py not found: Script may not exist for this component. Skip that check, proceed with audit-reference-depth.py as the sole Tier 1 gate.

Decomposition validation fails: Content was lost during extraction. Restore from the snapshot at /tmp/decomp-before-{name}.md. Check that each extracted content block appears in a reference file. Common cause: a code block was partially extracted or a table was split across the body and a reference file.