Cascade Brain — discovery, editorial, and tuning

The companion to cascade-discover. Where discover is fast/breadth (gh search → ranked list), brain is depth + editorial — pulls in curated awesome-lists, cross-references against installed inventory, and produces a thoughtful HIGH/MED/SKIP table with rationale.

Primary command — analyze

python3 ~/.claude/skills/cascade-brain/brain.py analyze [--top 30] [--min-stars 30]

Outputs a research brief with three sections:

1. GitHub search candidates (via cascade-discover --json) — top N repos ranked by stars × recency, already cross-checked against installed inventory, with closest_installed annotation per candidate (heuristic redundancy hint)
2. Curated awesome-lists to fetch — URLs to hesreallyhim/awesome-claude-code, wshobson/agents, ComposioHQ/awesome-claude-skills, etc.
3. Editorial framing for Claude — explicit instructions to bucket every candidate into HIGH / MED / SKIP with one-line rationale, scoped to your RevOps/sales workflow

Workflow when invoked

When you (Claude) run cascade-brain analyze:

1. The script outputs the structured brief to stdout
2. You then WebFetch each curated list URL listed in section 2
3. Parse those READMEs for repos NOT in the inventory list
4. Merge: gh-search candidates + curated-list net-new entries
5. For each candidate, decide:
   • HIGH = net-new gap-filler, recommend installing
   • MED = complement to existing skill X, worth watching
   • SKIP = redundant / niche / stale / low-quality
6. Save the final markdown report to ~/.claude/discover-reports/analysis-<date>.md

The script is data infrastructure; the editorial reasoning is yours.

Secondary command — sharpen

Improve a SKILL.md description using the TRIGGER when: / SKIP: pattern.

Inspect mode (default)

python3 ~/.claude/skills/cascade-brain/brain.py sharpen <skill-name>

Outputs the current description, score, body excerpt, and a template. You read it, draft a sharpened version.

Apply mode

python3 ~/.claude/skills/cascade-brain/brain.py sharpen <skill-name> --apply "description: New text. TRIGGER when..."

Writes the new description to the SKILL.md, prints before/after score. Run cascade-brain rebuild afterward to refresh the index.

Workflow

1. cascade-brain audit --weak-only to find candidates
2. For each weak skill: cascade-brain sharpen <name> to read context
3. Draft new description following the template (one line, includes TRIGGER when quoted phrases and SKIP: alternatives)
4. cascade-brain sharpen <name> --apply "<new desc>" to write
5. cascade-brain rebuild once batch is done

Utility subcommands

When to invoke

Use this skill when:

• "what should I install next" / "do a deep cascade analysis" / "research new claude tools" → analyze
• "sharpen X" / "fix the description for X" / "audit weak skills" → sharpen / audit
• "why didn't X fire" / "trace routing for Y" → trace
• "cascade health" / "rebuild index" / "lurking skills" → health / rebuild / stats

Skip when:

• Quick gh-only search without editorial → use cascade-discover directly
• One-off code edit to hooks → just edit
• Installing a specific named skill → use the existing install flow

Schedule weekly analysis

schedule "weekly cascade analysis" "python3 ~/.claude/skills/cascade-brain/brain.py analyze --top 50 > ~/.claude/discover-reports/weekly-$(date +%F).md"

Then on Friday morning Claude reviews the file and produces the editorial report.

Files this skill touches

sharpen --apply is the only command that writes to skill files. Everything else is read-only.