How it works

This page is for the curious. You don’t need any of it to use Chiron.

Inside Chiron, an advisor is called a seat, and a saved council is called a bench. Those are the words you’ll see in the files and commands.

An advisor (seat) is one real thinker’s published record

A seat is one advisor, distilled from what a real person actually published, cited claim by claim. Not a vibe, not a name on a card, not a prompt that says “act like Munger.” A seat carries the person’s frameworks, mental models, heuristics, anti-patterns, and refusals, each traceable to a book, a talk, or an episode.

A seat is a standard Agent Skill: a folder at skills/<id>/ with a SKILL.md. Claude, Codex and ChatGPT can load it on their own when a question fits. Ask “what would Munger say about this incentive?” and the munger skill loads without a command.

What makes a skill a seat, and not just any skill, is that it passes the seat standard: cited sources, explicit refusals, a consistent voice, and a full set of reference files. The frontmatter carries an x-chiron: block that flags the skill as a seat:

---
name: munger
description: Applies the reasoning, mental models, and temperament of Charlie Munger...
x-chiron:
  display_name: Munger Lens
  mode: corpus
  domains: [investing, decision-making, incentives]
  provenance:
    subject: Charlie Munger
    living: false
    sources:
      - "Poor Charlie's Almanack (2005)"
      - "The Psychology of Human Misjudgment (1995)"
      - "Wesco Financial annual meeting transcripts (1999-2011)"
---

That x-chiron marker is load-bearing. The registry and linter find seats by scanning for it, which means Chiron can share a skills directory with unrelated skills and the chiron router itself: anything without an x-chiron block is ignored, so a folder full of your other skills never gets mistaken for a seat.

The anatomy of a seat, in one glance:

skills/munger/
├── SKILL.md            # how it thinks: priors, heuristics, refusals, voice
├── disagreements.md    # recorded disagreements with other seats, both sides cited
├── references/         # the depth: full extraction of the published thinking
│   ├── principles.md   ├── mental-models.md   ├── frameworks.md
│   ├── anti-patterns.md ├── heuristics.md     ├── quotes.md   └── sources.md

Memory is not in the seat folder. Notes live in ~/.chiron/logs/<id>/log.md (see Your data), so a plugin update can’t wipe them. Chiron never writes to the plugin install or to ~/.claude.

Everything is files. No database, no server, no API keys. The full contract is SEAT_SPEC.md, summarised in the advisor standard.

Three modes

Every seat declares a mode. The mode decides how it may speak.

Mode	For	Voice	Constraints
corpus	Real people, built from published work	Third person about the corpus, a source per claim	The default, and the only mode that ships for a real person
original	Your own operating system, or an invented advisor	First person in its own voice	No provenance constraints (nobody to cite but the author)
persona	First-person emulation of a subject	Speaks as the subject	Never a default. Forbidden for a living subject without a granted license

Corpus mode is the backbone of the roster. A corpus seat speaks about the record, never as the person: “the Munger corpus would flag the incentive here (The Psychology of Human Misjudgment, 1995),” not “I, Charlie, would flag this.” Eighteen of the nineteen bundled seats are corpus mode. Rayo OS is the one original-mode seat: it’s the project author’s own framework.

Original mode is for a mind that is its own source. Rayo OS is one: the author’s own framework for finishing and shipping work (FATE, the Ship Gate, MAX THREE), with no outside subject to cite. Your own me seat is another. When you run /chiron:distill-me, Chiron reads how you actually operate and writes an original-mode seat holding your mission, values, goals, frameworks, and the experts you already reach for, so your own operating system can chair a council and be argued with. That me seat is private: it lives in ~/.chiron/seats/ or a project’s .chiron/seats/, both gitignored, never committed, and never shipped from the plugin repo.

Why living people are corpus-only

A living person’s seat is always corpus mode. This is both a legal line and a quality line.

The legal line: first-person emulation of a living, unlicensed person runs into right-of-publicity and false-endorsement risk. The linter enforces it mechanically. Rule L1 hard-fails any seat that combines mode: persona, living: true, and an ungranted license. It is a load-blocking error, not a warning, and /chiron:distill will refuse to build such a seat under any phrasing. The full policy is in LEGAL.md.

The quality line: sourced rules of thumb beat an imitation. A corpus seat that has to cite a page for every claim cannot drift into a plausible-sounding paraphrase of nothing. The constraint that makes the seat legally safe is the same one that makes it trustworthy.

The depth bar

Seats are not stubs. A name and a one-paragraph description is exactly what Chiron refuses to ship. Every real-person seat carries the seven standard reference files (principles, mental-models, frameworks, anti-patterns, heuristics, quotes, sources), each a full extraction of that dimension of the subject’s thinking with a named source per claim. In practice that is roughly 6,000 to 8,500 words of cited material per seat.

The linter defends the bar so “a seat” always means the same thing:

L4 fails a corpus seat with fewer than three provenance sources.
L12 fails a seat that is missing any of the seven reference files.

If a seat skimps on provenance, the build fails. That is what separates a seat from a name, and it is why depth is baked into the hire rather than sold as an upgrade. See the-standard.md for the whole rule set.

Memory

Each seat keeps a short note per question in ~/.chiron/logs/<id>/log.md: the date, the question, its one-line take, and your decision if you chose to share it. Recording your decision is opt-in: answers end with “Want me to remember what you decide…?” and nothing more is recorded unless you answer. When a seat loads, it reads its last 10 notes, so it can say “in June I advised X; you decided Y.”

The notes stay on your computer. /chiron:log <seat> --open-loops lists decisions you never closed. Full details in Your data.

The router

You don’t have to know which advisor to ask. Ask without naming anyone (or use /chiron:consult) and Chiron decides who should answer. It weighs three things: - Stakes. Cheap and reversible, or expensive and irreversible? Reversibility is the biggest single input. A call you can undo next week does not need a council.

Breadth. One seat’s domain, or does it straddle several (a pricing move that is part strategy, part psychology, part negotiation)?
Type. A judgment call a seat should weigh in on, or something factual, technical, or trivial? Seats opine on decisions, not on syntax errors or lookups.

Then it picks the cheapest verdict that fully serves the decision:

SKIP. No seat has standing. Factual, trivial, or outside every seat’s competence. Answering “you do not need a council for this” is a feature, not a dodge.
ASK one seat. One clear domain owner and low-to-moderate stakes. Routes to /chiron:ask.
COUNCIL. Convened only when the call is expensive or irreversible, genuinely cross-domain, or sits on a recorded disagreement between installed seats. Capped at five for a routed council.
CLARIFY. If the stakes truly cannot be read, one sharp question, then route. Never convene to cover for ambiguity.

The router states its verdict before it acts, so you can override it:

This is a pricing call that's hard to undo and touches strategy, persuasion and
leverage, so I'll ask Munger, Naval and Rumelt separately. Munger and Naval
disagree on record about concentrating your bets, which matters here.

You can always say “just ask Voss” or “make it a full council” to overrule the route.

A council

A council is the opposite of one model wearing five hats. It runs in three steps.

Independent takes, in isolation. Chiron spawns one subagent per seat, in parallel. Each subagent sees only that seat’s SKILL.md, that seat’s last 10 log entries, and the question. It does not see the other seats, their takes, the bench, or your conversation. The disagreement is structural, produced by giving each mind a different corpus, not by telling one model to “be the contrarian.”
Recorded disagreement. Chiron reads the disagreements.md of every seat in the room and surfaces only conflicts that are on the record: entries where both counterparts are present and the topic matches the question, quoted with the citations for each side. If there is no recorded disagreement on this topic, the council says so. It will not generate, infer, or dramatize a disagreement that is not in the files. Made-up conflict is exactly what Chiron is built to avoid.
One recommendation, disagreement kept. A chairman returns a single recommendation, not a menu, and names the dissent instead of averaging it away: which seats disagree, why, what the recommended path costs you, and what would change the call. The chairman is a neutral analyst by default, or a seat you name (your me seat can chair its own councils).

Councils never invent a seat, and never invent a fight. Both guardrails are what make the debate worth reading.

A bench

A bench is a saved line-up of seats with a chairman. It is plain YAML you can edit by hand, and Chiron reads the same file:

benches:
  money:
    seats: [munger, naval, hormozi]
    chairman: neutral      # a seat id, or "neutral"
    default_depth: quick   # quick | full

Benches live globally at ~/.chiron/benches.yaml and per project at .chiron/benches.yaml, and project shadows global on a name collision. Until you have either file (or an older ~/.claude/chiron/benches.yaml, which is still read and not written), the example benches bundled with the plugin are used. Name a bench where a council expects a seat list and it convenes that line-up. See Councils for the details.

Why sourced, not generated

Many AI advisor tools generate a “mind” from any input in a weekend, then fake disagreement by handing each generated persona an opposing archetype. Chiron takes the other bet: seats are hand-authored and curated, every substantive claim points at a source, a linter fails the build on seats that skimp on provenance, and a council surfaces only conflicts the real people actually argued, both positions cited. That is the reason a Chiron seat can tell you something you do not want to hear and you have a reason to believe it.