The AI Org Chart

    Paste one prompt into Claude Code (or Codex, Copilot, etc.) and get back an org chart of your AI setup — every agent you've built, who reports where, what's quietly broken, and which department has nobody in it.

    BEFORE YOU RUN IT

    Before you run it

    Three things to do first, then let it work.

    1. Open Claude Code (or Codex or Copilot, etc.) — the terminal, the desktop app, or claude.ai/code in your browser.
    2. Switch to the biggest model — type /model and pick Fable. This audit reads your whole setup, so give it the good brain.
    3. Give it a few minutes. It works quietly and comes back with one page.

    Nothing to prep. If you've barely built anything yet, run it anyway — a mostly-empty chart is a real finding, and it tells you exactly what your first hire should be.

    PROMPT 01

    The org chart prompt

    Copy the whole thing, paste it into Claude Code, hit enter. You get one page back.

    Prompt
    Build me an org chart of my AI setup. Do the whole thing in one pass — no
    check-in, no questions.
    
    STEP 1 — Inventory, quietly. Don't print any of this in chat. Find what I
    have running, and for each note what triggers it, what it does, and what
    context it can read:
    - Skills: ~/.claude/skills/*/SKILL.md and .claude/skills/ in my projects
    - Slash commands: ~/.claude/commands/
    - Hooks + config: ~/.claude/settings.json
    - Context files: every CLAUDE.md, with size and last-modified date
    - Cloud routines: RemoteTrigger tool, action "list" (skip if unavailable)
    - Scheduled jobs: macOS — launchctl list | grep -v com.apple, crontab -l.
      Windows — Task Scheduler. Linux — systemctl list-timers.
    - MCP connectors: claude mcp list
    - What I actually do: the FIRST user message only from recent sessions in
      ~/.claude/projects/*/ (head/jq — never read whole transcripts). Group
      into themes and count them as a percentage of my sessions. This is where
      my time really goes, as opposed to where my tooling points.
    If something doesn't exist, skip it silently. Never invent entries.
    
    Also note anything wrong: duplicates, references to files not on disk,
    anything untouched in 60+ days, anything switched off but still sitting
    there, any scheduled job whose output I've probably never seen.
    
    STEP 2 — Sort every item by ONE question: who decides what happens next,
    me or the model?
    L1 Automation — I wrote every step, no model involved
    L2 AI + Automation — I wrote the steps, the model fills one slot
    L3 Agent — the model decides the steps, working in one context
    L4 Multi-Agent — the model decides AND delegates to fresh sub-agents
    A skill is not a worker — it's a job description on a shelf, and it only
    runs when something picks it up. Keep skills separate from workers.
    
    STEP 3 — Build ONE self-contained HTML page and publish it as an artifact.
    Put these five sections on it, in exactly this order. The chart comes
    first; do not lead with the explainer.
    
    1. VERDICT — one sentence at the top, in plain counts, naming the thing
       that's actually true about my org. Then one sentence on what it costs
       me. Make this the boldest thing on the page.
       Directly beneath it, a single row of 4-6 bare numbers with labels —
       workers running, how many sit at each level, empty departments, jobs
       failing silently. Numbers only, no commentary.
    
    2. THE ORG CHART — draw it as a real org chart, a tree with connector
       lines: my box alone at the top, one row of department boxes directly
       below me, and each worker card hanging in a vertical stack beneath its
       department. Name departments the way a company would, choosing from:
       Marketing Analytics, Content Production, Sales, Client Delivery,
       Operations & IT, Finance, Research. Use only the ones that apply, and
       keep a department in the row even at zero workers — an empty Sales box
       hanging off the tree is the finding.
       Each card: name, an L1–L4 badge, a status flag (healthy / degraded /
       over-provisioned / unused), its trigger, and ONE line on what it does.
       Nothing else on the card.
       Dashed cards for open roles — jobs nobody is doing.
       A department with no workers and no open roles gets no cards at all —
       just one factual sentence beneath its box saying what happened to it or
       why it never existed. State it as a fact, not as a plan.
       The tree is wider than body text: let it break out of the text column
       to the full page width, and on a phone collapse it to departments
       stacked vertically with the cards indented beneath each one. The page
       itself must never scroll sideways.
       Every number in the row of numbers above must match the cards exactly.
       Count them twice.
    
    3. ON THE SHELF — my skills as a compact table: name, last touched, and
       state (in use / rarely picked up / N days cold / blocked / superseded /
       client gone). Visually distinct from the worker cards. If I have no
       skills, say so in one line; that's the finding, not an omission.
    
    4. THE LADDER — now explain L1–L4, using MY examples from the chart above
       on each rung. Open the section with: the rung is not a quality score,
       it's where judgment lives, and higher is not better — running a job one
       rung above what it needs is the most expensive mistake on the page.
       40 words max per level. A small inline-SVG diagram per level is good.
    
    5. FIX THESE THREE FIRST — ranked by what each costs me, not by how hard
       it is. For each: what's broken in one sentence, the evidence you
       actually found (a date, a file, a count), and what it's costing me.
       60 words max each.
       Diagnosis only — no commands, no config snippets, no terminal output,
       no step-by-step repair. Tell me what's wrong and what it costs me, and
       stop there.
    
    Hard limits:
    - Max 14 worker cards. Merge the small ones if I have more.
    - No paragraphs anywhere except the ladder. No glossary, no methodology
      section, no context-layer diagram, no file-path tables, no code blocks.
    - Every claim traces to something you actually found. If a section would
      be thin, say "not enough data" rather than padding it.
    - Inline all CSS, works in light and dark mode, plain absolute file paths.
    
    In chat: the artifact link, plus 3 bullets and nothing else — the biggest
    mismatch between where my tooling points and where my time actually goes,
    the most over-engineered thing I'm running, and the one role to hire next.
    Be blunt. A flattering chart is useless.