The Engine Room

Claude OS — the operating system behind Hutchins Climate Capital. Eleven staff roles, fifteen skills, two regression-tested gates, one compass. This page is the map: what runs, what it's called, and where it lives.

Reference edition · 7 July 2026 · maintained in the claude-os repository

Five functions

Everything in the system does one of five jobs. The first three were built through spring 2026; the Adversary and the Compass were added 7 July 2026, when the architecture was re-rooted in the goal rather than the file tree.

What happened?

The Record

Calls, contacts, meetings, and the knowledge wiki. Captures everything so nothing is re-derived. Compounds weekly.

What ships?

The Factory

Skills + the worker queue. Briefs decompose into fresh-context runs; every deliverable is versioned, never overwritten.

Is it true?

The Gate

The Editor and claim-audit. Every number, name, and status verb verified against its source before anything ships.

Should we?

The Adversary

Attacks live decisions before commitment — premise, posture, kill condition. Challenge as architecture, not tone.

Toward what?

The Compass

The goal as a live object in the charter. Week and month reviews test attention and output against it.

The staff

One name per job. The depth lives in the skill or agent spec each role points to — the roster is kept honest by a daily script that checks every pointer resolves.

Scribe

Record

Calls become contact files, meeting notes, wiki entries. Disambiguators and name discipline enforced at capture.

transcriptpost-call

Chief of Staff

Record

The morning pulse — calendar, meeting prep, what needs attention today.

prepdaily-start

Researcher

Record

Background, connection mapping, deep research. Primary sources first; "doesn't exist" requires a direct search.

inteldeep-research

Planner

Factory

Frames the problem before anyone drafts — one-sentence problem, binding constraint, eigenquestion.

frame

Ghostwriter

Factory

Drafts in the house voice, by register — email, long-form primer, social. The voice guide is the single source.

emailprimervoice-guide

Quant

Factory

Carbon registry and market data; rebuilds the dashboards on a fortnightly clock.

data-pipelineexcel

New 260707

Counsel

Factory

Contract review against a stated negotiating position — reads the actual file, never trusts an extraction for absence; every deviation classified fight / trade / accept.

contract-review

New 260707

Banker

Factory

Deal and platform economics — scenario grids over point estimates, mandatory downside case, arithmetic shown inline.

deal-economicsexcel

Editor

Gate

Red-teams every deliverable: breaks the central claim first, then verifies numbers, names, status verbs, and the record-test on outward mail. PASS / FIX / DO-NOT-SEND.

review-deliverableclaim-audit

New 260707

Adversary

Attacks decisions before they're taken: premise → evidence age → incentives → record → cost of drift → kill condition. PROCEED / RESHAPE / STOP.

adversary

Coach

Compass

Weekly and monthly step-back — convergences, hypothesis testing, time allocation, the Compass check, frank assessment.

reviewcoach-playbook

Not staff, but on the payroll: the Librarian (wiki lint) tidies the asset; five guardrail hooks — Lock, Secrets-scan, Proofreader, source-check, Nudge — run deterministic checks in code; the heartbeat script watches the whole machine daily and goes RED when a gate is edited without being re-tested.

Four flows

How work actually moves. Each box names the skill or script that runs that step. Green steps are the Gate; coral steps are the Adversary; the black-bordered step is the one human in the loop — nothing publishes without that approval.

Capture — every call compounds

Call happensGranola records it

→

Processed same evening7:30pm pull, triaged by substancetranscript

→

Contact + meeting notetiered, names disambiguatednetwork/

→

Wiki synthesisone line on the topic pageknowledge/

→

Follow-up if warranteddrafted, never auto-sentpost-call

Produce — from ask to approved deliverable

The aska written job spec — "draft the article series", "review every skill"

→

Split into small jobsa queue of pieces, each doable in one sittingworker

→

One fresh run per jobclean context every time — quality doesn't decayworker frame

→

Editor passbreak the claim, then verify every factreview-deliverable

→

Fact auditif it could be publishedclaim-audit

→

Outputs checked on diska run's own say-so doesn't countgit

→

PF approvalthe only publish gate

Decide — challenge before commitment

Meeting prepstatus, context, talking pointsprep intel

→

Adversary blockon any live negotiation: their strongest case, what not to concedeadversary

→

The meetingconversations, decisions, deals

→

Follow-up draftedin-voice, delivered as a Gmail draftpost-call email

→

Record-testwhose case does this put on paper?review-deliverable

Steer — the cadence

Morning brief8am — system health, calendar, prepdaily-start

→

Week reviewFriday — every open decision gets a kill conditionreview adversary

→

Compass checkwhat moved the needle; what just burned attentionCLAUDE-OS.md

→

Month reviewrevenue reality, hypothesis pruningreview

→

Nightly snapshot9:30pm git commit — the safety netnightly_commit.sh

Tested, not trusted

Judgment doesn't survive a model swap; tests do. Both gates carry a regression suite built from real, documented failures. The heartbeat goes RED if either spec is edited without a re-run — and the day-one instruction after any model change is to run both suites and compare.

Editor regression

11/12

Twelve synthetic drafts, one per documented failure class — fabricated numbers, wrong attributions, status verbs, stale figures, framing-with-citations, negative existence, compression inversion, source scoping, pin-cites, outward correspondence, extraction blindness, magnitude sanity. Last run 260707. Two known residuals, each with a compensating control at draft time.

Adversary regression

4/4

Four decision-situations where challenge-before-commitment would have paid: an untested frame, a stale premise, a self-harming record posture, and drift-by-default. Scored blind against what a passing attack must surface. Last run 260707 (3/4 first pass; spec patched where it proved thin, then 4/4).

The rhythm

When	What runs	Job
8:00am weekdays	`daily-start`	System heartbeat, calendar, prep briefs for today's meetings
7:30pm weekdays	`transcript-sync`	Pulls and processes the day's call transcripts
Friday 4pm	`week-review`	Decisions (with kill conditions), convergences, time allocation, Compass check
Last Friday	`month-review`	Revenue reality vs the Compass, hypothesis pruning, 90-day forward
1st + 15th	`dashboard-refresh`	Full market-data refresh and re-score
9:30pm nightly	`nightly_commit.sh`	Git snapshot of the whole system — the safety net everything else stands on
On demand	`worker`	Relay through the work queue, launched from a live session and reporting back into it — no cron, no background routine. Single-sitting briefs skip the queue and run interactively

The library — where the system lives

Paths are relative to the claude-os repository. The charter loads every session; everything else loads on demand.

Document	Role
`CLAUDE-OS.md`	The charter + the Compass. Mission, operating rules, 90-day priority. Loaded into every session; when reality moves, this changes first.
`system/reference/session-brief.md`	Start here. State, principles, open work — and the model-handover section: ten judgment behaviours a successor must hold, with the day-one test protocol.
`system/reference/system-map.md`	The system by role — staff, asset, guardrails, tools.
`system/reference/errors-log.md`	What went wrong, and the fix. Every entry stamped with where its rule now lives. The regression suites are built from this file.
`system/agents/STAFF.md`	The roster — one line per role, script-checked daily against the skills it points to.
`system/agents/review-deliverable.md`	The Editor's spec, plus `editor-regression/` (12 cases) and its score file.
`system/agents/adversary.md`	The Adversary's spec, plus `adversary-regression/` (4 cases) and its score file.
`system/reference/coach-playbook.md`	Live strategic hypotheses, presence baselines — feeds the reviews.
`system/reference/patrick-voice-guide.md`	The voice, by register. Skills point here; nothing duplicates it.
`system/reference/anti-slop-patterns.md`	Banned patterns and tell-words, loaded before content delivery.
`knowledge/INDEX.md`	The wiki's routing layer — one-line syntheses pointing at topic pages.
`system/reference/run-sheet-260704/QUEUE.md`	The work queue and the worker protocol — including deterministic outcome verification.

The fifteen skills

Skill	Owner	One line
`transcript`	Scribe	Calls → contact files, meeting notes, wiki. Disambiguators are load-bearing.
`post-call`	Scribe	Follow-up drafting in-voice, record-tested before anything outward.
`prep`	Chief of Staff	Meeting prep — a brief is a hypothesis, not a record; the transcript wins.
`intel`	Researcher	Background and connection mapping; negative-existence claims need a direct search.
`frame`	Planner	Problem → binding constraint → eigenquestion; refute your own frame first.
`email`	Ghostwriter	Correspondence in-voice; the record-test fires before drafting to any counterparty.
`primer`	Ghostwriter	Explanatory long-form — capable-but-uninformed reader, no manufactured jeopardy.
`contract-review`	Counsel	Position-first review; an extraction proves presence, never absence; fight / trade / accept.
`deal-economics`	Banker	Frame-test before the spreadsheet grows; scenario grids; downside mandatory.
`excel`	Banker / Quant	Workbook mechanics — numbers come from cells read this turn, never memory.
`data-pipeline`	Quant	Registry consolidation and dashboard injection; firm totals are cross-registry by definition.
`claim-audit`	Editor	Forensic fact-check of every claim in anything publication-bound; blocks on errors.
`review`	Coach	Week and month reviews — numbers trace to primary notes, never to last week's review.
`worker`	Factory	Queue relay for multi-sitting briefs, launched from a live session; outputs verified on disk, never self-reported.
`access`	Ops	Manages the allowlist on the gated data properties.

House rules that hold it together

A line earns its placeby catching a real error or saving real time — otherwise it's cut. Deleting is a success outcome. (The 260615 teardown bar; every element on this page passed it.)

Verify before you claimGo to the source that settles it. The wiki is the map, never the citation. Status verbs are sourced like numbers.

Capture it where it firesA correction becomes code if it's binary, a skill rule if it's craft, an errors-log entry if it's a fact — never a loose note.

Never overwriteVersions survive (-v2, -v3). The nightly snapshot means nothing is ever really lost.

Challenge by defaultThe hole first, not the strength. Now structural: the Adversary is mandatory on negotiation deliverables and week-review decisions.

One human gateNothing publishes, sends, or ships externally without explicit approval. The system prepares; Patrick decides.

Built 7 July 2026 with Claude Fable 5 · The structure survives a model swap; the regression suites prove whether its successor does.