Build an advisor

This page is how you add advisors. /chiron:distill turns a thinker’s published work into a new advisor (a seat) with sources, usable like the bundled ones. /chiron:distill-me turns you into a private seat that can chair your own councils. Both run the same depth bar, and both end at a passing lint.

In Codex and ChatGPT, ask in plain words: “Build a Chiron advisor for Brad Frost on design systems.” Both commands build from what the person actually published. See the advisor standard for the bar every advisor has to pass.

Command	Builds	Mode	Ships in the repo?
`/chiron:distill <desc>`	a public advisor seat from a subject’s corpus	`corpus` (real or fictional), or `original`	Yes, if you contribute it
`/chiron:distill-me`	your private `me` seat from your own history	`original`	Never

Part 1: `/chiron:distill`, a new seat from a corpus

/chiron:distill takes a one-line description of the advisor you want and builds the whole thing, checks it, and saves it where Chiron finds it. You don’t move any files by hand.

/chiron:distill "Brad Frost on design systems"
/chiron:distill "a negotiation expert, I'm thinking Chris Voss"
/chiron:distill "W. Edwards Deming on quality and management"

The flow

Four steps. It tells you what it’s doing as it goes, because the research takes several minutes.

1. A short interview (one round). Chiron infers what it can from your description and asks only what is missing:

Subject: a real person’s published corpus, a fictional character with a canon, or an original advisor you are inventing.
Living status: for real people, this sets the legal lane (see LEGAL.md).
Mode: corpus is recommended for any real person, and it is the only mode that ships for real people. There is a hard rule here (CH-E8): if you ask for persona mode (first-person voice) of a living, unlicensed subject, Chiron refuses, explains the rule in a sentence, and offers you either corpus mode or an original-mode archetype alias (“The Offer Architect” instead of the person). It will not build the persona under any rephrasing.
Domains: one to six registry tags, so /chiron:roster --domain <tag> and the auto-router can find the seat.
Anchor sources: if you already know the key books, talks, or essays, name three or more and Chiron starts there. If you do not, it discovers them in research.

If your request is vague, Chiron does not interrogate you. It offers two or three concrete candidate readings and lets you pick:

▸ By "a design expert" do you mean:
  (a) Brad Frost: design systems, atomic design
  (b) Don Norman: usability, affordances, human-centered design
  (c) someone specific?

2. Deep research (the part that takes real time). Chiron searches the subject’s published corpus in multiple rounds: survey the whole body of work first, then go deep per dimension. Books, talks, podcasts, essays, interviews. It is extracting a mind, not summarizing a Wikipedia page. The targets, straight from the depth standard:

every named framework, with when-to-use, steps, and source
every named mental model or concept, defined, with origin
the full belief catalog, each principle with its rationale
everything the subject warns against, each anti-pattern with the failure it produces
operational heuristics in IF/THEN form
verified short quotes, wording checked against a source, paraphrases marked as paraphrases
a complete dated bibliography, primary sources first

3. Author the seat. Chiron writes the seat as a real, auto-activating Agent Skill. The <id>/ directory it produces:

<id>/
├── SKILL.md              # the mind: frontmatter + Priors, Heuristics, Refusals, Voice (< ~6k tokens)
├── disagreements.md      # authored, cited conflicts with installed seats (omitted if none exist)
├── references/           # the depth: seven files at the depth bar
│   ├── principles.md     ├── mental-models.md   ├── frameworks.md
│   ├── anti-patterns.md  ├── heuristics.md       ├── quotes.md
│   └── sources.md

The consult log is not part of the seat. The first consult creates ~/.chiron/logs/<id>/log.md, outside the plugin install. It is not written under ~/.claude.

SKILL.md carries the frontmatter contract (a name matching the folder, a triggering description, and an x-chiron block with display_name, mode, domains, an alias for living subjects, and provenance with three or more sources). Its body carries the required ## Priors (five or more), ## Heuristics (five or more, in IF/THEN shape), ## Refusals (two or more), and ## Voice, plus the narrative sections that give a seat its texture. Real-person seats end with the disclaimer block. The subject’s two or three signature mental models get named inline in the body, so the mind is visible without loading the reference files.

disagreements.md only appears when there is a real, citable, published conflict between this subject and a seat you already have installed, with both positions sourced. If none exists on the record, Chiron skips the file rather than inventing conflict. A chairman that manufactures disagreement is the exact failure mode Chiron exists to kill.

4. Lint until it passes. Chiron runs the linter and fixes findings until the seat is clean:

python3 scripts/lint_seat.py <path-to-new-seat>
# 0 = pass · 1 = warnings · 2 = errors (load-blocking)

Then it reports the seat card (id, display name, mode, domains, source count), the per-file word counts, and how to use it. You can re-run the check yourself any time with /chiron:lint.

Here is roughly what a run looks like end to end:

/chiron:distill "Chris Voss on negotiation"

▸ interview
  Subject: Chris Voss (living). Mode: corpus (recommended). Domains:
  [negotiation, persuasion, communication]. Anchor: Never Split the
  Difference (2016). Discovering talks and interviews for the rest.

▸ researching · surveying the corpus
  Never Split the Difference; The Black Swan Group masterclass; TED-style
  talks; podcast appearances. Going deep per dimension now.
  found: tactical empathy · calibrated questions · the accusation audit ·
  mirroring · labeling · "That's right" vs "you're right" · the 7-38-55 rule

▸ authoring ~/.chiron/seats/voss/  (writing SKILL.md + 7 references + disagreements.md)

▸ lint
  PASS [voss] ~/.chiron/seats/voss
  references: principles 1,180w · mental-models 1,020w · frameworks 1,340w ·
  anti-patterns 900w · heuristics 1,110w · quotes 640w · sources 720w

  Seat ready. Try:  /chiron:ask voss "counter a lowball on the retainer"
                    add voss to a bench in ~/.chiron/benches.yaml

Where the finished seat lands

Where it’s saved depends on your app. Either way you don’t move anything by hand.

Claude Code, Codex and ChatGPT write to ~/.chiron/seats/<id>/, so every Chiron command finds it from then on. It doesn’t write to ~/.claude/skills/. If you want it scoped to one repo instead, it writes <project>/.chiron/seats/<id>/. It never writes inside the plugin install. If Codex’s sandbox blocks the home folder, it saves to .chiron/seats/<id>/ in the project and tells you.
Cowork cannot write the home directory, and cannot write ~/.claude. It writes <project>/.chiron/seats/<id>/ in the open workspace (the registry scans that path, so councils find it with no home-directory write). If there is no writable project, it emits a one-advisor .plugin into outputs/ for you to accept in the app. More in Cowork and Other agents.

The registry merges seats by precedence: project shadows global shadows bundled. A seat you distill with the same id as a bundled one (say, a richer munger of your own) overrides the shipped seat everywhere you use Chiron.

Fictional and original advisors

You can distill any mind, not only living authors.

Fictional characters distill in corpus mode, cited to their canon. The bundled detectives prove it: Sherlock Holmes (corpus), Dr. House (corpus), and Columbo (corpus) are each sourced to the stories and episodes and pass the same lint as the real-person seats. Point /chiron:distill at the canon and treat it as the corpus.
Original advisors you invent from scratch use original mode. There is no provenance constraint and no disclaimer, because nothing is being attributed to a real person. Rayo OS, the author’s own framework, is the bundled example. For an original seat, Chiron substitutes your own materials and an interview for web research: the frameworks come from you, not from a published body of work.

One at a time, never a shallow batch

If you ask for several advisors at once, Chiron distills them sequentially, each with its own full deep-research pass. It will not batch them into a single shallow sweep. Thinness and repetition across seats is the tell of a rushed job, and it is exactly what the depth bar exists to prevent. Expect each real-person seat to carry roughly 6,000 to 8,500 words of cited reference material across the seven files. That takes minutes per seat, and that is the point.

Tips for a good distill

Give it anchor sources. If you know the two or three foundational works, name them. Chiron starts from primary sources instead of spending a round finding them, and you get a better seat.
Let it offer options when you are vague. “A pricing expert” is a fine way to start. Chiron will hand you a short menu of who it thinks you mean. Pick one, and it confirms before it commits.
Name the conflicts you care about. If you already run a council and want the new seat to argue with a specific existing one, say so during the interview so its disagreements.md covers that advisor.
Contribute it back. A seat you author is a standard Agent Skill directory. If it is a real thinker distilled to the bar, it can go in the public roster. Your private me seat never does (below).

Your own advisor

/chiron:distill-me puts yourself in the room.

/chiron:distill-me distills you the way /chiron:distill distills a thinker. It writes an original-mode seat with id me: your own operating system, so your priorities, values, and frameworks can sit in the room, chair your councils, and be argued with. It also reads out the experts you already reach for and turns them into a recommended roster.

/chiron:distill-me
/chiron:distill-me ~/notes/how-i-work.md
/chiron:distill-me ~/obsidian/vault

The optional argument points it at notes, a vault, a folder, or a document to read as source material.

The me seat holds personal data: your mission, your values, your goals, your decisions. Chiron states the privacy rules up front and then honors them. It reads personal sources only with your consent, names each source out loud before reading it, and never transmits any of it anywhere.

What it gathers

You choose which of these it draws on. It reads only what you approve:

This conversation and recent chat history: the frameworks, phrases, and decisions you actually reach for.
Instruction and memory files: your CLAUDE.md or AGENTS.md files and any memory files your app keeps. These already encode how you operate.
Anything you point it at: a notes vault, a “how I work” doc, a resume, past writing. It reads these as your corpus.
An interview to fill the gaps, run as a conversation rather than an interrogation. It offers concrete suggestions (“a lot of people channel someone like Munger for money calls, who is yours?”) instead of a wall of open questions.

What it extracts

It synthesizes, in your own terms and cited to the source where possible:

Mission and what you are optimizing for, plus your current active priorities.
Values and non-negotiables: the lines you will not cross.
Goals: concrete and time-bound, with relative dates converted to absolute ones.
Frameworks and mental models you operate by, named, each with when it applies.
Operating principles and heuristics in IF/THEN form.
Anti-patterns and traps: how you characteristically fail, and the counter.
Voice: how you think and want to be addressed.
Go-to experts: the thinkers you channel, tagged by domain.

Where it lives, and why it is private

The me seat is written to ~/.chiron/seats/me/ (global) or <project>/.chiron/seats/me/ (project). In Cowork, which cannot write the home directory, it uses the project path, which the registry scans. It is never written under ~/.claude, never written into the Chiron plugin repo, never committed, and never shipped. If it is written into a project, Chiron adds .chiron/seats/ to that project’s .gitignore, and offers to do it for you.

It uses the full seat structure: SKILL.md with name: me, x-chiron set to display_name: "<your name> (you)", mode: original, and your domains. Original mode needs no provenance and no disclaimer, because it is self-authored. The body carries the same required sections as any seat, where ## Refusals becomes your own guardrails: the decisions and moves you refuse to make. The seven reference files hold the depth so SKILL.md stays lean, and Chiron lints it until it passes.

How it chairs your councils

After it writes the seat, /chiron:distill-me maps your named go-to experts against the installed roster and reads out three groups: experts you already have installed, experts that are available and worth distilling next (it offers to /chiron:distill each), and proposed starter councils built from your experts by domain. It suggests the me seat as the chairman for personal-decision benches, so your own operating system arbitrates the council instead of a neutral chair:

benches:
  personal:
    seats: [munger, naval, taleb]
    chairman: me            # your operating system runs the synthesis
    default_depth: full

From there, put the new bench to work with a real decision:

/chiron:council personal "take the retainer at 15k/mo or walk?"