Expert Mind Skill

A rotating circle of 21 named experts living inside your Claude Code. Mention an expert by name, by domain, or by group — Claude answers through their frameworks, principles, and voice. Auto-refreshes weekly from LinkedIn, RSS, YouTube, podcasts.

flowchart LR
    U["Your question<br/>in Claude Code"] --> R{How you frame it}
    R -->|"Use Eric Seufert..."| N["1 expert<br/>(by name)"]
    R -->|"mobile UA experts..."| D["2-4 experts<br/>(by domain)"]
    R -->|"growth experts lens..."| G["2-3 experts<br/>(by group)"]
    N --> C["Claude responds<br/>through that lens"]
    D --> C
    G --> C
    style C fill:#d4f1d4

Why this exists

You follow great operators on LinkedIn, podcasts, Substack. Their frameworks shape how you think. But when you actually need one of those frameworks — at midnight, mid-decision, in a meeting — you don't have their voices in your head. You have their posts in a tab you never reopened.

The information is accessible. The recall is not. This plugin fixes the recall layer: it summons distilled expert thinking on demand, in the expert's actual voice, with their frameworks intact.

What you can do with it

1. Decision support before commitment

"Use Rory O'Driscoll — at our ARR, is now the right time to raise Series B, or should we focus on Rule of 40 for two more quarters?"

Get a structured answer through a SaaS-VC's framework before the board meeting, not after.

2. Borrow expert framing for your own writing

"I'm drafting a memo on cohort financing for mobile apps. Use Eric Seufert's framing — economic-historical, with concrete dollar examples."

The skill includes voice samples — characteristic phrasing, not just topics.

3. Multi-perspective without scheduling four calls

"Give me 3 expert perspectives on whether to launch a referral program — growth lens, brand lens, monetization lens."

Activates three different personas in parallel. Each answers in their own section, no synthesis.

4. Stay current without doomscrolling

The weekly cron pulls fresh content automatically. You don't need to remember to read. New frameworks from your tracked experts land in the skill, ready to invoke.

5. Use as a research starter

"Before I write my position paper on AI in advertising — what do Eric Seufert, Andrew Chen, and Andrej Karpathy think about this space?"

Three lenses in one prompt. You write the paper; the experts surface relevant prior thinking.

Who it's for

• Founders & investors making strategic decisions in mobile apps, B2B SaaS, growth, monetization
• Operators & PMs at growth-stage companies who want senior frameworks without senior salaries
• Builders who follow named thinkers but suffer from "I read this and I know it's relevant but I can't remember the specifics"
• Forkers who want to bundle their own expert circle (different domain, different language, different sources)

Two install paths

Path A — Just use the 21 bundled experts (zero config)

# Inside Claude Code
/plugin marketplace add github:mooreslaws/expert-mind-skill
/plugin install expert-mind-skill@expert-mind-skill

Done. All 21 experts are now active. No API keys, no setup, no coding. Mention any of them and they activate.

Path B — Add your own experts (interactive wizard)

/expert-mind-skill:init

A guided dialog walks you through everything. No coding required. The wizard asks plain questions:

✅ LLM provider configured? (if not, paste an Anthropic API key)
✅ Which experts to add? (pick from 21 presets, or define your own)
✅ Which sources to pull from? (LinkedIn, RSS, X, YouTube, podcasts, …)
✅ Setup auto-refresh? (weekly cron, manual, or "just ping me in chat")
✅ Run the pipeline now?

You can exit anytime — every step is optional. Re-run init later to continue where you left off. Filesystem-driven: the wizard inspects what you already have configured and only asks what's missing.

How to invoke an expert (3 patterns)

Available groups: marketing, product, strategy, ai, finance, leadership, design, sales, growth, saas, legal. Each expert is tagged with 1-3 groups.

For parallel side-by-side answers (no synthesis), say so explicitly:

"Give me 3 separate expert perspectives on X — don't synthesize, let each answer in their own section."

What's in the box (21 bundled experts)

Each skill: ~2-3k tokens of frameworks + principles + opinions + voice samples, distilled from 80-120 LinkedIn/RSS/YouTube items per author.

Quality report: output/EVALUATION_REPORT.md (21/21 with usable applicability, 0 fair-applicability).

Technical description

Everything below is for people who want to fork, add their own experts via config files, or understand how the pipeline produces a skill. Skip if you just want to use the bundled set.

How the pipeline works

flowchart TD
    S1[LinkedIn]:::src --> A[Adapters<br/>scripts/scrape_*.py]
    S2[RSS / Substack]:::src --> A
    S3[YouTube]:::src --> A
    S4[Podcasts]:::src --> A
    S5[X / Telegram / Readwise]:::src --> A
    S6[IMAP / Manual text]:::src --> A

    A --> B[(Raw content<br/>staging/raw/)]
    B --> J{LLM Judge<br/>scripts/judge.py}

    J -->|score < 0.7<br/>noise / news / promo| X[Discarded]
    J -->|score ≥ 0.7<br/>framework / principle / opinion| E[(Accepted excerpts<br/>staging/scored/)]

    E --> D[Distill<br/>scripts/distill.py]
    D -->|dedup + voice extraction| O[(SKILL.md ~3k tokens<br/>skills/&lt;id&gt;/SKILL.md)]

    O --> CC[Claude Code<br/>skill loader]
    CC --> R[Activates when you<br/>mention the expert]

    classDef src fill:#e3f2fd,stroke:#1976d2
    style J fill:#fff3cd
    style D fill:#fff3cd
    style R fill:#d4f1d4

Every cadence cycle (default: weekly):

1. Ingest — each enabled source adapter pulls latest content. RSS gets new posts since last sync; LinkedIn pulls last 100 posts via Apify; etc.
2. Judge — Claude scores each excerpt 0-1. Calibrated against calibration_seed.md (50 labeled examples). Items ≥0.7 are kept. Errors surface loudly as ⚠️ ERRORS=N in the run summary.
3. Distill — accepted items bucketed by type (framework / principle / opinion / voice sample), deduplicated semantically, written to a compact SKILL.md (target: 2-4k tokens, hard cap: 8k).
4. Auto-refresh — skills update in place. Manual edits between <!-- BEGIN AUTO --> markers are preserved.

Pre-flight checks before any expensive call: Apify and Anthropic are pinged for auth + balance. Aborts in <1 second if 402/credit-low. No wasted spend on misconfigured runs.

Source adapters

All 12 adapters are implemented. Each is a single Python file in scripts/.

Get API keys at:

• Anthropic: https://console.anthropic.com/settings/keys
• Apify: https://apify.com/sign-up → Settings → Integrations → API token
• Readwise: https://readwise.io/access_token

Planned for v1.2: reddit_user, hackernews, arxiv, readwise_highlights, podcast RSS with Whisper transcription, Farcaster.

Cadence and scheduling

Three options for auto-refresh:

Option 1: "Ask me in chat" (recommended for non-developers)

No cron at all. Just say "refresh my skills" or "обнови Эрика" to Claude Code, the agent runs /expert-mind-skill:run. Zero setup.

Option 2: GitHub Actions

For users with the plugin checked into a GitHub repo. Copy deploy/github-actions.yml.template to .github/workflows/. Auto-commits refreshed skills back. Free for public repos.

Option 3: Local launchd (macOS) / systemd (Linux)

For users who want refresh to keep happening when their machine stays on. /expert-mind-skill:init will generate the unit file and install it.

Cost

Default config (Sonnet judge, 21 personas, ~1.5 items/day each, weekly cadence):

Bootstrap (initial pull, 100 posts/persona): ~$15 one-time.

Switch to Haiku for ~$0.30/week:

export EXPERT_MIND_JUDGE_MODEL=claude-haiku-4-5

Plugin layout

expert-mind-skill/
├── .claude-plugin/
│   ├── marketplace.json         # marketplace manifest (1 plugin)
│   └── plugin.json              # plugin manifest
├── commands/                    # slash commands
│   ├── init.md                  # /expert-mind-skill:init
│   ├── add-persona.md           # /expert-mind-skill:add-persona
│   ├── status.md                # /expert-mind-skill:status
│   └── run.md                   # /expert-mind-skill:run
├── skills/                      # 21 bundled expert skills (read-only)
│   ├── eric-seufert/SKILL.md
│   └── …
├── scripts/                     # pipeline (Python, stdlib only + PyYAML)
│   ├── orchestrator.py          # main entry — ingest → judge → distill
│   ├── persona_loader.py        # reads personas/*.yaml
│   ├── judge.py                 # LLM-as-judge scoring
│   ├── distill.py               # bucket + dedup + voice extraction
│   ├── wizard_state.py          # filesystem-driven setup state
│   ├── scrape_linkedin.py       # Apify-driven LinkedIn ingester
│   ├── scrape_x.py              # Apify-driven X/Twitter ingester
│   ├── pull_rss.py              # RSS/Atom ingester
│   ├── pull_youtube.py          # YouTube channel + transcript ingester
│   ├── pull_telegram.py         # public Telegram channel scrape
│   ├── pull_readwise.py         # Readwise Reader API
│   ├── pull_mastodon.py         # public Mastodon timelines
│   ├── pull_bluesky.py          # public Bluesky feeds
│   ├── pull_webfeed.py          # generic URL → plain text
│   ├── pull_email.py            # IMAP folder
│   ├── scrape_x_official.py     # X API v2 (alternative to x_apify)
│   └── import_manual.py         # drop-in .txt files
├── cold-start/registry.yaml     # 21 pre-curated persona templates
├── hooks/hooks.json             # SessionStart welcome
├── deploy/
│   ├── github-actions.yml.template
│   └── test-workflow.yml.template
├── calibration_seed.md          # 50 labeled judge examples
├── SPEC.md                      # full system specification
└── README.md                    # this file

After installation, you create these in your working directory (gitignored by default):

your-project/
├── personas/<id>.yaml           # one file per expert YOU added
├── .env                         # API keys (gitignored)
├── output/skills/<id>/SKILL.md  # YOUR custom skills (alongside bundled)
├── logs/<id>.jsonl              # attribution log per persona
└── staging/                     # ingest cache (gitignored)

Adding your own expert (advanced — config files)

The wizard (/expert-mind-skill:init) is the recommended path. But if you prefer to edit YAML directly:

1. Create personas/<your-id>.yaml:

   id: your-id
   name: Person Name
   role: One-line description
   expertise: [tag1, tag2, tag3]
   voice: |
     2-3 sentences describing distinctive speech patterns
   groups: [marketing, growth]
   sources:
     - type: linkedin_apify
       handle: linkedin-handle
     - type: rss
       url: https://their-blog.com/feed.xml
   cadence: weekly
   

2. Add API keys to .env (the wizard does this for you):

   ANTHROPIC_API_KEY=sk-ant-...
   APIFY_TOKEN=apify_api_...
   

3. Run the pipeline:

   python3 scripts/orchestrator.py
   

Three rules for good personas

Rule 1: Voice description is the judge's filter. It's what the judge uses to distinguish "this excerpt sounds like our author" vs "generic noise."

• ❌ Generic: "Speaks about marketing topics."
• ✅ Distinctive: "Brand-first contrarian. Argues against pure-performance ROAS optimization. Uses concrete examples (Chubbies, Liquid Death) to anchor frameworks."

Rule 2: Multiple sources beat single source. LinkedIn-only personas produce thin skills (LinkedIn is full of reposts). LinkedIn + newsletter RSS + podcast appearances = 3x signal. Recommended minimum: 2 sources.

Rule 3: Don't bloat the skill. Hard cap 8k tokens. If hitting 8k:

• Tighten expertise tags (too broad → judge accepts too much)
• Sharpen voice description (too generic → false positives)
• Reduce max_per_section in judge/config.yaml

Quality & validation

• output/EVALUATION_REPORT.md — per-persona stats (accepted ratio, tokens, fullness, applicability) across all 21 bundled skills
• calibration_seed.md — 50 labeled examples (35 YES + 10 NO + 5 BORDERLINE) used to anchor the judge prompt
• SPEC.md — full system architecture

Contributing a new expert to the cold-start registry

PRs welcome to expand the 16-persona default set.

1. Fork this repo
2. Add an entry to cold-start/registry.yaml:

   <persona-id>:
     name: ...
     role: ...
     expertise: [...]
     voice: |
       2-3 sentences describing distinctive speech patterns
     groups: [...]
     recommended_sources:
       - type: rss
         url: ...
       - type: linkedin_apify
         handle: ...
   

3. (Optional) Generate the SKILL.md by running the pipeline against your own sources, commit the resulting file to skills/<persona-id>/SKILL.md
4. Open a PR

See CONTRIBUTING.md for details on adding new source adapters.

License

MIT. Forks welcome. PRs to expand cold-start/registry.yaml very welcome.