newroaster

What this does

Adds one seat to the Roast Room. A seat is four things, and this skill makes all four:

A brain: .claude/skills/roast-room/brains/<slug>/, eight files compiled from the person's public work, every quote verbatim with a source id.
An agent: .claude/agents/council-<key>.md, the sub-agent that reasons and talks like them from that brain.
A portrait: four cartoon frames in the room (<key>-base, -open, -mid, -blink), made from a real photo of the person in exactly the style of the original seven.
A roster entry: one line in .claude/skills/roast-room/roster.json, so /roast-room and the room know the seat exists.

The raw sources stay in a local build folder that git ignores. Only the brain ships.

Build folder: build/<slug>/ at the project root.

Read before the first run: references/brain-spec.md (the brain layout and its rules), references/extract-prompt.md (what each extraction agent gets), references/writer-prompt.md (what each brain writer gets), references/agent-template.md (the seat agent).

Who can be a roaster

Public figures with a real body of public work: essays, letters, books as they teach them in talks, long interviews where they are the main speaker, keynotes, their own channel or podcast. Living or historical.
The user, or someone who gave consent: built from material the user provides (their posts, transcripts, notes). Use --add in step 3.
Not: fictional characters, private people without their consent, or anyone whose public work is too thin. Under about 15 usable sources or 60,000 words, say so and offer --quick with a warning, or a different person.

The seat is a simulation, not the person. The agent says so if asked, and only quotes what is in its brain.

Depth

Flag	Sources	Words	Extraction agents	Use it for
`--quick`	12 to 20	80k+	2 to 3	Trying someone out
default	30 to 50	250k+	4 to 6	A seat you will actually use
`--deep`	80+	500k+	8 to 12, in waves	Matching the original seven

Quality beats count: one 90-minute keynote is worth ten 6-minute clips.

Step 1: Who and what seat

From the user's message, get the person's full name. If anything below is missing, ask in one AskUserQuestion batch (at most four questions), offering your best guess as the first option:

The pillar and the question the seat owns. Propose one from what the person is known for, worded like the existing seats in roster.json (pillar: a few words; question: one plain question a founder would face). If an existing seat already owns that pillar, propose a sharper angle so the two seats argue instead of repeating each other.
Default board? Put them at the table for every council, or keep them on call (--seats and ask). If the default board already has nine seats, ask who they replace.
Depth: quick, standard, or deep (default standard).

Then fix the names: key (short lowercase, letters only, unique in the roster, usually the surname: naval, drucker; the first name when another seat already has the surname: leila), slug (first-last), short (how the room names them: "Naval", "Leila"), and the scorecard dimension (a few words for what this seat scores, like "Leverage and long-term wealth").

Check the roster first: if the person already has a seat, stop and offer to refresh it instead (step 3 onward on the same slug; fetch skips sources already in sources.json).

Step 2: Find sources

Use WebSearch to build a source list, best first. Look for:

Their own site, blog, essays, newsletter archive, or letters (the richest text, with dates).
Long talks and keynotes where they are the speaker.
Long interviews and podcasts where they are the main guest (their own podcast counts).
Their books as they teach them in public: author talks, interviews about the book, chapters they posted free. Never fetch pirated or paywalled book text.
Recent material (last two years) so positions.md reflects current views, and a spread of years so changes of mind show up.

Skip clips under about 10 minutes, fan compilations, reaction videos, and pieces where someone else does most of the talking. Write the URLs to <build>/urls.txt, one per line, with a # heading line per group. Aim for the source count of the chosen depth plus about 20 percent, since some will fail.

Step 3: Fetch

node .claude/skills/newroaster/scripts/fetch.mjs <build> --list <build>/urls.txt

Each source becomes <build>/corpus/S001-<title>.md, logged in <build>/sources.json. It prints ok, skip, thin (paywall or a JavaScript page), or fail per URL, then a manifest. YouTube uses yt-dlp when installed, otherwise YouTube's own captions. For local material the user gave you:

node .claude/skills/newroaster/scripts/fetch.mjs <build> --add path/to/file.md --title "..." --date 2024-05-01

If a source came back thin or failed, find a different copy (a transcript page, an archive.org copy) or replace it. A PDF without pdftotext is saved as pending-*.pdf: read it with the Read tool, save the text as a .md, then --add it. Re-run until the manifest meets the depth target, then open two or three corpus files to confirm they are clean text by the right person. Write the manifest line counts into your notes for the final report.

Step 4: Extract (map)

Split the sources into batches of about 50,000 words each (a 90-minute talk is about 13,000; a long essay 5,000). Spawn one extraction agent per batch, in one message, in parallel, foreground: subagent_type: general-purpose, model: sonnet, with the prompt from references/extract-prompt.md (fill in the name, the seat's pillar and question, the batch's source ids and file paths, and the notes file to write). Run at most six at a time; for --deep, run them in waves.

Each agent writes <build>/notes/batch-NN.md. Wait until every agent has reported. When all are back, check every note quote:

node .claude/skills/newroaster/scripts/check_quotes.mjs <build> <build>/notes

Fix what it reports: correct a wrong source id, delete a quote it cannot find. Notes with no verifiable quote can stay as paraphrase, marked "inferred from S0xx".

Step 5: Write the brain (reduce)

Create .claude/skills/roast-room/brains/<slug>/. Write sources.md with the script:

node .claude/skills/newroaster/scripts/fetch.mjs <build> --sources-md .claude/skills/roast-room/brains/<slug>/sources.md --name "Full Name"

Then spawn three writer agents in one message, in parallel (general-purpose, default model), with the prompt from references/writer-prompt.md. Each reads every notes file plus references/brain-spec.md and writes its files into the brain folder, following the spec's budgets and rules:

Writer A: frameworks.md, principles.md, and books.md.
Writer B: positions.md and stories.md.
Writer C: voice.md.

Tell each writer the seat's pillar and question, the build folder (so it can open a corpus file to check a quote's context), and that every quote must be copied exactly from the notes with its source id. Writers write only their own files.

When they are back, write _hot.md yourself from the six finished files (the spec lists what goes in it, including "where this lens misfires"). Then check the whole brain:

node .claude/skills/newroaster/scripts/check_quotes.mjs <build> .claude/skills/roast-room/brains/<slug>

Fix until it passes: every quote verbatim, every id right. Read _hot.md once more as the agent will: it should tell you how this person decides, in their words, in under 800 words.

Step 6: Write the agent

Write .claude/agents/council-<key>.md from references/agent-template.md. Fill every bracket from the brain: the lens (five or six questions this person asks of any decision), how they talk (from voice.md), how they debate (name their natural opponents among the seats in roster.json and what they concede, from "where this lens misfires"), and the never-list. The description names the pillar, the source count and what the brain is built from, and the trigger phrases ("what would [name] say").

Step 7: Register the seat

node .claude/skills/newroaster/scripts/add_seat.mjs --key <key> --name "<Full Name>" --short "<Short>" --slug <slug> --pillar "<pillar>" --question "<question>" --dimension "<dimension>" [--default] [--swap <key>]

It refuses if the agent file or a brain file is missing. --default adds the seat to the default board; --swap <key> takes another seat off it in the same move. Then:

Add one or two "natural tension" lines for the new seat to .claude/skills/roast-room/references/seats.md (who they will clash with, and over what).

Step 8: Portrait

Every seat gets a real portrait. Never leave a new seat on the drawn placeholder bust and never draw a real person from a written description alone: the likeness comes from a photo. This is the same pipeline the original seven went through, with the same prompts and the same style reference (assets/style-reference.jpg, the Buffett cartoon). Portrait folder: <build>/portrait/ (ignored by git). Frames folder: app/public/portraits/.

Find a photo. One recent, clear, front-facing photo of the person alone, from an official source: their own YouTube channel avatar (yt-dlp -J --flat-playlist --playlist-items 0 https://www.youtube.com/@handle lists avatar_uncropped; swap its =s0 for =s900), their own site, or their Wikipedia or Wikimedia Commons photo. Skip logos, group shots, and side profiles. For the user or a mentor built from their own material, use a photo they give you.
Save it and look at it: node .claude/skills/newroaster/scripts/portrait.mjs photo <build>/portrait <url>. Open photo.jpg with Read and confirm it is the right person and the face is clear.
Two cartoon candidates: node .claude/skills/newroaster/scripts/portrait.mjs base <build>/portrait "<Full Name>. <hair, facial hair, eye color, skin, build, and the outfit they are known for>". Describe what the photo shows; name the signature outfit or accessory (a cap, glasses, a chain). The description steers the details; the photo carries the likeness.
Pick by likeness: python .claude/skills/newroaster/scripts/build_portraits.py sheet <build>/portrait, then open candidates.jpg (photo first, then each candidate). Pick the one closest to the photo (face shape, eyes, hair, signature details). If neither is close, run base again with a sharper description (eye color and hairline are the usual misses) and --n 2, then pick: portrait.mjs pick <build>/portrait <n>.
Talking frames: node .claude/skills/newroaster/scripts/portrait.mjs frames <build>/portrait (mouth open, mouth "oh", eyes closed, each an edit of the chosen base).
Build: python .claude/skills/newroaster/scripts/build_portraits.py build <build>/portrait <key> --out app/public/portraits. It aligns each edit onto the base, patches only the mouth or eyes, keys the green, and crops every head to the same scale as the seven. Open <build>/portrait/review.jpg: four frames, same face, no green fringe, the mouth and eyes the only change. If a head sits too small or too large next to the others, rebuild with --zoom 0.9 or --zoom 1.1.

Needs KIE_AI_API_KEY (kie.ai, about 5 image jobs a seat) in .env, Node, and Python with opencv-python and numpy (pip install opencv-python numpy if the build step says they are missing). If there is no kie.ai key, finish the seat anyway (the room shows a drawn placeholder), tell the user in one line how to add a key (https://kie.ai/api-key, then KIE_AI_API_KEY= in .env), and that /newroaster portrait <name> makes the portrait afterward.

Portrait only (/newroaster portrait <name>, or a seat that shows the placeholder): run this step alone for a seat already in roster.json, using its key and slug, then tell the user to reload the room.

Step 9: Test the seat

New agent files load when Claude Code starts, so test with the fallback the council itself uses: spawn general-purpose with Read .claude/agents/council-<key>.md and follow it as your instructions for this whole task. Only read files; do not write anything. followed by the single-seat consult prompt from .claude/skills/roast-room/references/round-prompts.md, on a short sample decision in the seat's pillar. Then check its answer:

It sounds like the person (compare with voice.md), speaks in the first person, and never claims to be them.
It names frameworks with source ids, and every quote passes check_quotes.mjs (save the answer to <build>/test-consult.md and run the checker on it).
It owns its pillar instead of giving generic advice.

If it fails, fix the agent or _hot.md and test again, once.

Step 10: Report

Tell the user, in under 150 words:

Who was added, their pillar and question, and whether they are on the default board.
The brain: sources, words, years covered, and the files written.
How to use them now: /roast-room ask <key> ... for one answer, /roast-room --seats <key>,buffett,... ... to seat them, and /roast-room board to change the default board.
One line: restart Claude Code (or open a new session) so the seat loads as its own agent type; until then the council uses the fallback automatically.
One line: their portrait is in the room (reload the page), or, when no kie.ai key was available, that it shows a placeholder until /newroaster portrait <name> runs.

Without sub-agents (Codex or a runtime that cannot spawn agents)

Run steps 4 and 5 yourself, one batch at a time: extract a batch into its notes file, then move to the next. Write the brain files one at a time from the notes. Everything else is the same.

Rules

Public and lawful sources only. The corpus stays in the build folder and is never committed or shared; the brain holds short quotes (under 45 words) and paraphrase, never book text.
Every quote in the brain is verbatim and passes check_quotes.mjs. A line with no source is inference and says so.
Do not touch the other seats' brains or agents. Only roster.json's "seats" and "default" change, through add_seat.mjs.
Plain English in everything you write yourself: short sentences, no em dashes, no filler.
A portrait is part of every seat, made from a real photo with the original seven's prompts and style reference. A written description alone never sets the likeness.
Cost: a standard seat is about 8 to 10 sub-agent runs plus about 5 kie.ai image jobs. Say which depth you used.