compound - SKILL.md Agent Skill

name: compound description: Capture and apply knowledge from course development to improve future runs. license: MIT metadata: author: Andamio version: 1.0.0

Skill: Compound Knowledge

Description

Extracts patterns, heuristics, and calibration data from course development artifacts. Feeds knowledge back into /draft-slts, /assess-slts, /self-assess-readiness, and /classify-lesson-types to make each run smarter than the last.

Invocation Modes

/compound                          # Interactive: asks what to compound
/compound quality-review           # Compound from specific phase
/compound readiness                # Compound from readiness assessment
/compound classification           # Compound from lesson type classification
/compound --course=go-pbl --rollup # Full course retrospective

Instructions

Path Resolution

Resolve file paths based on your execution context:

Plugin context (${CLAUDE_PLUGIN_ROOT} is set): Read knowledge from ${CLAUDE_PLUGIN_DATA}/knowledge/ (user data), falling back to ${CLAUDE_PLUGIN_ROOT}/knowledge/ (seed data). Write all knowledge updates to ${CLAUDE_PLUGIN_DATA}/knowledge/ — never modify the plugin's bundled seed data.
Clone/symlink context (default): Read and write knowledge at knowledge/ relative to the project root.

All knowledge/ paths referenced below follow this resolution. In plugin context, substitute the appropriate prefix.

Phase Selection

If invoked without arguments, present phase options:

## What would you like to compound?

| # | Phase | Source Artifact | Extracts |
|---|-------|-----------------|----------|
| 1 | quality-review | 02-slts-quality-review.md, 01-slts.md | Successful rewrites, quality issues |
| 2 | readiness | 05-readiness-assessment.md | Tier distribution, context shopping list |
| 3 | classification | 04-lesson-type-classification.md | Verb patterns, edge cases, heuristics |
| 4 | lesson-build | lessons/*.md | Actual vs self-assessed confidence |
| 5 | context-add | assets/ + re-run readiness | Which resources unlocked which SLTs |
| 6 | rollup | All artifacts | Full course retrospective |

Which phase? (Or specify course: --course=slug)

Course Selection

If no course specified, scan courses-in-progress/ and ask which course to compound from:

## Select Course

| # | Course | Status | Artifacts Available |
|---|--------|--------|---------------------|
| 1 | andamio-for-contributors | building | 01, 02, 03, 04, 05 |
| 2 | andamio-for-api-developers | building | 01, 02, 03, 04, 05 |

Which course?

Also check examples/ for seeding data (like go-slts-readiness-assessment.md).

Extraction Logic by Phase

Phase: quality-review

Source files:

02-slts-quality-review.md (assessment output)
01-slts.md (revised SLTs, if exists)

Extract:

Successful rewrites: Compare SLTs between quality review suggestions and revised SLTs. For each rewrite:

- before: "original SLT text"
  after: "improved SLT text"
  issue_type: unmeasurable_verb | task_focused | too_broad | etc.
  key_change: "what made the difference"
  course: "course-slug"
  date: "YYYY-MM-DD"

Append to knowledge/slt-patterns/successful-rewrites.yaml

Quality issues: Extract patterns from "Needs Work" SLTs:

- pattern: "how to detect"
  description: "what the problem is"
  impact: ["Student-Facing Language", "Specificity"]
  frequency: 1
  example_bad: "I can understand blockchain"
  example_fix: "I can explain how a blockchain maintains data integrity by identifying three mechanisms"
  courses_seen_in: ["course-slug"]

Append to knowledge/slt-patterns/quality-issues.yaml

Verb effectiveness: Extract verbs from "Strong" SLTs and add to verb bank:

- verb: "compare"
  bloom_level: analyze
  success_count: 1
  example_slts: ["I can compare X to Y by identifying..."]

Update knowledge/slt-patterns/verb-bank.yaml

Phase: readiness

Source files:

05-readiness-assessment.md
examples/go-slts-readiness-assessment.md (for seeding)

Extract:

Context leverage: Parse the Context Shopping List and update rankings:

- resource: "Apollo API reference + transaction building examples"
  type: "Docs + Example Code"
  slts_unlocked: ["102.2", "102.3", "102.5", "102.6", ...]
  priority: High
  obtained: false
  effectiveness: null

Update knowledge/readiness/context-leverage.yaml

Calibration baseline: Record self-assessed tiers for later comparison:

- slt_id: "go-pbl:099.1"
  self_assessed: Ready
  actual_outcome: null  # filled in after lesson-build
  dimensions_off: null
  notes: null
  date: "YYYY-MM-DD"

Append to knowledge/readiness/calibration.yaml

Phase: classification

Source files:

04-lesson-type-classification.md

Extract:

Verb patterns: From the Heuristics Developed section:

- verb: "explain"
  suggests: exploration
  confidence: high
  count: 1
  examples: ["I can explain why Bursa was built..."]

Update knowledge/lesson-types/heuristics.yaml

Subject patterns: From topic clusters:

- keywords: ["API", "endpoint", "library"]
  suggests: developer_documentation
  confidence: high
  count: 1
  examples: ["I can build a web API using Fiber..."]

Update knowledge/lesson-types/heuristics.yaml

Edge cases: From ambiguous classifications:

- slt: "I can set up my development environment..."
  candidates: ["how_to_guide", "organization_onboarding"]
  chosen: how_to_guide
  deciding_factor: "Generic procedure, not org-specific"
  question_that_helped: "Would this SLT exist in a generic course?"
  course: "course-slug"
  date: "YYYY-MM-DD"

Append to knowledge/lesson-types/edge-cases.yaml

Phase: lesson-build

Source files:

lessons/*.md
05-readiness-assessment.md (for comparison)

Extract:

Calibration updates: Compare actual lesson-building experience to self-assessed readiness:

- slt_id: "course:module.slt"
  self_assessed: Ready
  actual_outcome: success | partial | failure
  dimensions_off: ["Code Demo was actually Weak"]
  notes: "Apollo API changed since training"
  date: "YYYY-MM-DD"

Update existing entries in knowledge/readiness/calibration.yaml

Compute calibration stats: After updating entries:
- Calculate accuracy_rate
- Identify common_overconfidence patterns
- Identify common_underconfidence patterns
- Generate adjustment rules

Phase: context-add

Source files:

assets/ (newly added context)
Re-run /self-assess-readiness (or compare to previous)

Extract:

Context effectiveness: For resources that were obtained:

- resource: "gOuroboros README"
  obtained: true
  effectiveness: confirmed | partial | unhelpful
  notes: "Unlocked 4/5 expected SLTs, one still needs examples"

Update knowledge/readiness/context-leverage.yaml

Phase: rollup

Run all extraction phases for a single course. Produce a summary report:

## Compound Report: [Course Name]

### Knowledge Captured

| Category | Count | Files Updated |
|----------|-------|---------------|
| Successful Rewrites | 3 | successful-rewrites.yaml |
| Quality Issues | 2 | quality-issues.yaml |
| Verb Bank Entries | 5 | verb-bank.yaml |
| Context Resources | 8 | context-leverage.yaml |
| Calibration Entries | 12 | calibration.yaml |
| Lesson Type Heuristics | 4 | heuristics.yaml |
| Edge Cases | 2 | edge-cases.yaml |

### Aggregate Stats Update

- Courses processed: [n]
- Total SLTs analyzed: [n]
- Successful rewrites captured: [n]
- Calibration accuracy: [%]

### Top Insights

1. [Most impactful pattern discovered]
2. [Second most impactful]
3. [Third most impactful]

Output Format

After extraction, always report:

## Compound Complete

**Phase:** [phase name]
**Course:** [course name]

### Extracted

| Knowledge Type | Count | Status |
|----------------|-------|--------|
| [type] | [n] | Added / Updated / Unchanged |

### Files Modified

- `knowledge/slt-patterns/successful-rewrites.yaml` - Added 2 entries
- `knowledge/readiness/context-leverage.yaml` - Updated 3 entries

### Index Updated

- `last_updated`: [timestamp]
- `slts_analyzed`: [new total]

Knowledge Consumption Check

Before modifying knowledge files, read the current state. When updating:

Increment counts (don't reset)
Append to lists (don't overwrite)
Merge patterns (combine evidence from multiple courses)
Deduplicate (same pattern from different courses = one entry with multiple course references)

Integration Points

This skill produces knowledge that other skills consume:

Skill	Reads From	Uses For
`/draft-slts`	verb-bank.yaml, quality-issues.yaml	Prefer effective verbs, avoid problematic patterns
`/assess-slts`	quality-issues.yaml, successful-rewrites.yaml	Flag known issues, suggest proven fixes
`/self-assess-readiness`	calibration.yaml, context-leverage.yaml	Adjust confidence, prioritize shopping list
`/classify-lesson-types`	heuristics.yaml, edge-cases.yaml	Improve initial guesses, handle known ambiguities

Guidelines

Always read before writing. Load current YAML state before appending.
Preserve existing data. Never overwrite — merge and increment.
Be specific in patterns. Vague patterns don't compound.
Update the index. Always update knowledge/index.yaml stats after any extraction.
Report what changed. The user should see exactly what knowledge was captured.
Seed from examples. Use examples/go-slts-readiness-assessment.md to prime the knowledge base.