rebuild-skill-index

Rebuild Skill Index

Generate one index file per platform, specific to the current Claude Code config. Each file is the extended list of every platform-relevant skill installed on this machine that is not already in the curated list inside the matching agent's markdown file.

Outputs are consumed by the matching subagents (ios-dev, android-dev, designer, marketer, neo4j-dev) and any caller that dispatches them (/specialist-review, /tech-vetting, /todos-review).

Step 0: Resolve base directory

Claude Code may be invoked with CLAUDE_CONFIG_DIR pointing at a non-default config. Resolve the current config's base directory first via Bash:

BASE="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
echo "$BASE"

Use the resolved absolute value in all subsequent read and write paths and in Path: metadata written into output files — readers should not have to re-resolve.

Step 1: Collect candidate skills

For each source below, read frontmatter only (the name: and description: fields). Do not read full SKILL.md bodies — that would burn unnecessary tokens. Use Read file_path=<path> limit=30 (30 lines covers any reasonable frontmatter).

Shared sources (used by every platform's classifier):

1. User-level skills: Glob pattern: $BASE/skills/*/SKILL.md

Android-specific sources:

• None beyond the shared user-level glob. All curated Android skills are vendored inside android-dev itself (under <plugin>/skills/) and excluded via the dedup list in Step 2.

iOS-specific sources:

2. Axiom plugin skills (optional):

   ls "$BASE/plugins/cache/axiom-marketplace/axiom/" | sort -V | tail -1
   

   • If the directory exists, use the highest semver version returned above and glob: $BASE/plugins/cache/axiom-marketplace/axiom/<version>/skills/*/SKILL.md
   • If it does not exist, record a "missing" warning for the iOS output file.

Design-specific sources:

3. frontend-design plugin (optional):

   ls -t "$BASE/plugins/cache/claude-plugins-official/frontend-design/" | head -1
   

   • This plugin uses hash directories, not semver; ls -t | head -1 picks the newest by mtime — matches "the one currently installed" even when multiple hash dirs coexist.
   • If present: $BASE/plugins/cache/claude-plugins-official/frontend-design/<hash>/skills/*/SKILL.md
   • If absent, record a "missing" warning for the design output file.

4. ui-ux-pro-max-skill plugin (optional):

   ls "$BASE/plugins/cache/ui-ux-pro-max-skill/ui-ux-pro-max/" | sort -V | tail -1
   

   • If present: $BASE/plugins/cache/ui-ux-pro-max-skill/ui-ux-pro-max/<version>/skills/*/SKILL.md
   • If absent, record a "missing" warning for the design output file.

Marketer-specific sources:

5. frontend-slides plugin (optional):

   ls "$BASE/plugins/cache/frontend-slides/frontend-slides/" | sort -V | tail -1
   

   • Slide-oriented skills — included in the marketer index because presentations are a content/brand output (the slide-design curated skill orchestrates them).
   • If present: $BASE/plugins/cache/frontend-slides/frontend-slides/<version>/skills/*/SKILL.md
   • If absent, skip silently (no warning — this is bonus, not required).

6. revealjs plugin (optional):

   ls "$BASE/plugins/cache/revealjs/revealjs/" | sort -V | tail -1
   

   • Same rationale as frontend-slides — alternative slide engine used by slide-design.
   • If present: $BASE/plugins/cache/revealjs/revealjs/<version>/skills/*/SKILL.md
   • If absent, skip silently.

Neo4j-specific sources:

• None beyond the shared user-level glob. All curated Neo4j skills are vendored inside neo4j-dev itself.

Step 2: Read curated dedup lists

Read each agent's markdown from the solopreneur marketplace plugins. Find the ## Curated Skills section. Extract skill names from every bullet line that looks like:

- `<skill-name>` — <description>

The section is subdivided by source (Plugin-bundled / Third-party / …). Collect names from all subsections — they form the per-agent dedup blacklist and should not be re-included in that agent's extended index. Names may be bare (asc-release-flow) or namespaced (ios-dev:ios-patterns, axiom:axiom-ios-ui) — strip any <plugin>: prefix before comparing against candidate skill names.

Agents to read (each lives in its own sub-plugin):

• ios-dev.md (from ios-dev) → iOS dedup list
• android-dev.md (from android-dev) → Android dedup list
• designer.md (from designer) → design dedup list
• marketer.md (from marketer) → marketer dedup list
• neo4j-dev.md (from neo4j-dev) → neo4j-dev dedup list

Locate each agent file via Glob (widens across any marketplace name the user chose at claude plugin marketplace add --name):

Glob: $BASE/plugins/cache/*/<plugin>/*/agents/<name>.md

If multiple versions coexist, take the highest semver match.

If you can't find an agent file, ask the user where that sub-plugin is installed and continue with an empty dedup list for that platform (the extended file will have duplicates with curated, which is fine — costs one extra entry per dup).

Step 3: Classify each candidate

For each skill not in a given platform's dedup list, judge from name + description whether it is relevant to that platform. Each candidate can be classified into multiple platforms (e.g., a SwiftUI-specific design skill belongs in both ios.md and design.md).

Be inclusive but not sloppy.

iOS classification

• Yes: Swift / SwiftUI / UIKit / Xcode / iOS / macOS / iPadOS / watchOS / tvOS / visionOS / Apple frameworks (CoreData, AVFoundation, CoreLocation, etc.) / Apple toolchain (lldb, Instruments, asc CLI, App Store Connect)
• No: web (React, Vue, Next.js, CSS), Android (Kotlin, Compose, Room), Python, LLM/agent frameworks, generic design tools, infra/devops, prose helpers
• Borderline (e.g. cross-platform performance audit, generic git workflow): exclude. Curated catches must-haves; extended is best-effort.

Android classification

• Yes: Kotlin / Jetpack Compose / Android Gradle / Hilt / Room / Retrofit / Coroutines / Flow / WorkManager / KMP-on-Android / Material Design (Material 3, Material You) / AGP / R8 / ProGuard / Android Studio / adb / gradlew / Google Play Console (gplay-*, signing, release flows, in-app purchases) / Android-specific testing (Roborazzi, Espresso, Compose UI testing)
• No: web (React, Vue, Next.js, CSS), iOS (Swift, SwiftUI, Xcode, asc CLI), Python, LLM/agent frameworks, generic design tools, infra/devops, prose helpers
• Borderline (e.g. cross-platform performance audit, generic git workflow): exclude. Curated catches must-haves; extended is best-effort.

Design classification

• Yes: visual design / typography / color / spacing / layout / motion / component styling / design systems / design review / critique / UX copy / information architecture / mockup generation / responsive design / accessibility when it's about visual contrast or type
• No: pure frontend implementation (React hooks, Next.js routing) without design framing, QA testing, backend, iOS/Android development patterns that aren't design-focused, devops, presentation slides (these go to marketer)
• Borderline (polish / audit / harden / normalize and other "action verb" skills): include only when the description explicitly mentions UI / visual / design. When in doubt, skip — curated catches the must-haves.

Marketer classification

• Yes: brand / GTM / positioning / messaging / naming / copywriting / content strategy / social writing (X/Twitter, LinkedIn, Threads) / newsletter / email marketing / SEO copy / AI-pattern removal / presentation slides / pitch decks / community management / launch playbooks
• No: pure dev (any language), pure design (visual systems without brand voice), infra, QA testing
• Borderline: skills that touch both brand and design (e.g. brand-aware presentation builder) — include in marketer. Slide-design tooling specifically goes here, not design.

Neo4j-dev classification

• Yes: Cypher (read or write) / Neo4j drivers (Java, Python, Go, .NET, JS) / graph schema / graph data modelling / MERGE patterns / MATCH patterns / index & constraint design / vector indexes on graph / APOC / GDS (Graph Data Science) / Neo4j Aura / cypher-shell / neo4j-admin / knowledge graph / GraphRAG / text2cypher
• No: relational SQL without graph aspects, document DBs (Mongo, Firestore), vector DBs without a graph layer (Pinecone, Weaviate without Neo4j), generic prose
• Borderline: GraphRAG skills that span LLM + graph — include only in neo4j-dev. If the skill is explicitly about prompt orchestration rather than graph retrieval, it can also land in a future ai-engineer index.

Step 4: Write output files

mkdir -p "$BASE/solopreneur/skill-index"

Write one file per platform. Expand $BASE to its absolute resolved value in all Path: lines so that consumers (agents that Read this file) can open each path directly without re-resolving the env var.

$BASE/solopreneur/skill-index/ios.md

# iOS Skills Index — Extended
Generated: <ISO 8601 timestamp with timezone>
Config: <resolved absolute $BASE>
Classified by: <model name running this skill>

## Auto-classified user skills
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Auto-classified Axiom plugin skills (axiom-marketplace v<version>)
- `<name>` — <description>
  Path: <$BASE resolved>/plugins/cache/axiom-marketplace/axiom/<version>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Missing
<empty if all good; otherwise warnings, e.g.:>
- Axiom plugin not installed. Install with: `claude plugin install axiom`
  (https://github.com/CharlesWiltgen/Axiom)

$BASE/solopreneur/skill-index/android.md

# Android Skills Index — Extended
Generated: <ISO 8601 timestamp with timezone>
Config: <resolved absolute $BASE>
Classified by: <model name running this skill>

## Auto-classified user skills
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Missing
<empty if all good; otherwise warnings>

$BASE/solopreneur/skill-index/design.md

# Design Skills Index — Extended
Generated: <ISO 8601 timestamp with timezone>
Config: <resolved absolute $BASE>
Classified by: <model name running this skill>

## Auto-classified user skills
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Auto-classified frontend-design plugin skills (v<hash>)
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/plugins/cache/claude-plugins-official/frontend-design/<hash>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Auto-classified ui-ux-pro-max plugin skills (v<version>)
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/plugins/cache/ui-ux-pro-max-skill/ui-ux-pro-max/<version>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Missing
<warnings for any optional source not found>

$BASE/solopreneur/skill-index/marketer.md

# Marketer Skills Index — Extended
Generated: <ISO 8601 timestamp with timezone>
Config: <resolved absolute $BASE>
Classified by: <model name running this skill>

## Auto-classified user skills
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Auto-classified frontend-slides plugin skills (v<version>)
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/plugins/cache/frontend-slides/frontend-slides/<version>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Auto-classified revealjs plugin skills (v<version>)
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/plugins/cache/revealjs/revealjs/<version>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Missing
<empty if all good; otherwise warnings>

$BASE/solopreneur/skill-index/neo4j-dev.md

# Neo4j Skills Index — Extended
Generated: <ISO 8601 timestamp with timezone>
Config: <resolved absolute $BASE>
Classified by: <model name running this skill>

## Auto-classified user skills
- `<name>` — <one-line description trimmed to ~120 chars>
  Path: <$BASE resolved>/skills/<name>/SKILL.md
- ... (alphabetical by name)

## Missing
<empty if all good; otherwise warnings>

Skip a section entirely if it would be empty (except ## Missing — always emit it for transparency).

Step 5: Report to the user

Print a short summary (one Wrote line per file actually written — skip platforms that had no sources available):

Wrote <$BASE resolved>/solopreneur/skill-index/ios.md
Wrote <$BASE resolved>/solopreneur/skill-index/android.md
Wrote <$BASE resolved>/solopreneur/skill-index/design.md
Wrote <$BASE resolved>/solopreneur/skill-index/marketer.md
Wrote <$BASE resolved>/solopreneur/skill-index/neo4j-dev.md

iOS:
- N user skills classified as iOS-relevant
- M Axiom skills classified as iOS-relevant (axiom v<version>)
- K curated skills excluded from extended index
- Warnings: <list, or "none">

Android:
- N user skills classified as Android-relevant
- K curated skills excluded from extended index
- Warnings: <list, or "none">

Design:
- N user skills classified as design-relevant
- M frontend-design plugin skills classified as design-relevant (v<hash>)
- L ui-ux-pro-max plugin skills classified as design-relevant (v<version>)
- K curated skills excluded from extended index
- Warnings: <list, or "none">

Marketer:
- N user skills classified as marketer-relevant
- S frontend-slides plugin skills classified as marketer-relevant (v<version>)
- R revealjs plugin skills classified as marketer-relevant (v<version>)
- K curated skills excluded from extended index
- Warnings: <list, or "none">

Neo4j:
- N user skills classified as Neo4j-relevant
- K curated skills excluded from extended index
- Warnings: <list, or "none">

Notes

• This skill is manual. There is no auto-rebuild on subagent invocation or on plugin update.
• Output is per-machine and per-config. Each Claude Code config (CLAUDE_CONFIG_DIR) gets its own index because each config has its own set of installed plugins and user skills.
• The plugin git repo only ships the curated lists inside each agents/<name>.md.
• A single candidate skill can land in multiple platform indexes (design + iOS overlap is common for SwiftUI-focused visual skills; brand + design overlap exists too, but slide tooling specifically routes to marketer).
• Classification is LLM judgment, not regex. Output may shift slightly between runs at the margin; that's acceptable because the curated list is the deterministic top-priority layer.
• The ai-engineer agent does not yet read an extended index file. If its curated list grows beyond what's manageable, add an ai-engineer.md output here following the same pattern.