AI Chief of Staff

v1.3.0 · Created by The NoW of Work · MIT licensed

An AI chief of staff for one leader. Morning briefs, prep before the meetings that matter, action items out of recorded calls, inbox triage, a ledger of open promises, a record of closed decisions, and a Monday review. Built out of plain text files you own.

Runs on Claude and on ChatGPT. Both versions are generated from one source, so they cannot drift apart.

Start here, on the platform you actually use

The leader works in	Read
Claude (Cowork for most people; Code if you are technical; ordinary chat via Path B)	docs/claude/READ-ME-FIRST.md
ChatGPT	docs/chatgpt/READ-ME-FIRST.md

Each set carries its own install guide, schedule prompts, and deployment runbook. Nothing in one applies to the other, so read only your own and ignore the rest.

---

The short version

/plugin marketplace add NoW-of-Work/ai-chief-of-staff
/plugin install ai-chief-of-staff@now-of-work

Then say "run onboarding" and answer six questions. Everything else is inferred from the calendar and mail you already have.

Those two lines are Claude Code. In Claude Cowork, add the marketplace through Customize > Plugins. Not using plugins, or using ChatGPT? Download the repo as a ZIP, unzip once, and follow the guide for your platform in the table above.

---

What it does

Thirteen behaviours. Nine of them can run on a schedule.

Behaviour	When	What comes out
onboarding	Once, at setup	A filled-in workspace, built from the calendar and mail rather than a questionnaire
morning-brief	Every morning	Under 400 words: today's priorities, the meetings that need something, open loops, drafts waiting on a yes
meeting-prep	Before a meeting that matters	Who, why, what happened last time, what is open between you both ways, three questions, the risks
transcript-to-actions	After a recorded call	Decisions, action items with owners, commitments for the ledger, what was too ambiguous to call
daily-transcript-sweep	Daily	Yesterday's meetings turned into actions without being asked
end-of-day-close	End of day	What moved, what came in, and a draft of tomorrow for you to approve
weekly-review	Monday	What slipped, where the time actually went, drift against what they said they are not doing
connection-check	Weekly	Notices tools that appeared, tools that quietly broke, and decides (rarely) whether a gap is worth one line
inbox-triage	Twice a working day	Mail cut to what needs a person: decisions waiting, drafted replies, promises both ways, a count of what it left alone
chase	Weekly	What you are owed and have not been given, a release call on what died, and at most three drafts sized to the relationship
decision-brief	When a decision is due, plus a Monday pass	The choice, the options and their costs, your own filters, a recommendation with two falsifiers
recall	On demand	An answer out of your own record, the file it came from, conflicts between sources, and what was searched and not found
health-check	Monthly	Whether the workspace is actually configured: blocking gaps, stale files, contradictions, folders that are not filling, one verdict

---

The three ideas it is built on

It drafts, it does not act. No sending, no booking, no replying. Twelve approval gates, listed in the operating manual, and none of them are negotiable.

It knows what it cannot see. A registry file tracks capabilities (Calendar, Email, Documents, Transcripts, Chat, Tasks, CRM) rather than vendors. That is the only reason the system can tell "nothing happened yesterday" apart from "I cannot see what happened yesterday." Those look the same in the output and they are completely different problems.

It stays quiet. A missing tool earns at most one line, at most once a fortnight, and only on a day where that tool would actually have paid for itself. Say "not now" twice and it never asks again. Three weeks out of four it says nothing at all about itself.

---

Repository layout

Path	What it is
docs/claude/, docs/chatgpt/	The install guide, the schedule prompts, and the client runbook, one set per platform. Read the set for the platform you use
QUICK-START.md	The ten-minute version, and the page to hand anyone else who will use it. The same on both platforms
src/	The only thing you edit. Behaviours, workspace templates, the worked example, docs
build/build.py	Generates every output below from src/
plugin/	The Claude plugin. skills/ inside it is generated
dist/claude/	Drop-in workspace folder, plus one upload ZIP per behaviour
dist/chatgpt/	Project instructions, thirteen prompt files, workspace files
.claude-plugin/marketplace.json	Lets anyone install this with /plugin marketplace add

Everything under dist/, docs/, and plugin/skills/ is generated and committed, so a client can download and use the repo without running anything. Every markdown file in the repo root is generated too, this README included. Edit the copy in src/docs/ and let the build write the output. A doc named in PLATFORM_DOCS in build/build.py is written once per platform into docs/, and everything else is written once to the root.

Inside src/

Path	What it is
src/behaviours/	Thirteen behaviour files, one per skill. YAML frontmatter carries name, order, and a description under 200 characters
src/workspace/	The seven files that make up a workspace: _manual.md, about-me.md, my-work.md, tomorrow.md, commitments.md, decisions.md, and connections.md, plus a README per output folder
src/example/	A filled-in workspace for a leader who does not exist. Copied into every build so you have something to compare against
src/docs/	Every markdown file that lands in the repo root. Rendered with platform=claude, so ChatGPT specifics go in as plain prose

decisions.md is your record of the calls you closed, and the system may open and re-status an entry without writing an outcome you have not stated. src/example/ is reference only. No behaviour reads it, and the build copies it in whole.

---

Maintaining a fork

Edit src/, then:

python3 build/build.py     # or: make build

One source, every output. The build:

• turns each behaviour in src/behaviours/ into a Claude skill and a ChatGPT prompt
• resolves {{#claude}} and {{#chatgpt}} blocks so each version says the right thing about its own platform
• copies the workspace templates into the onboarding skill, so setup can create files that do not exist yet
• copies src/example/ into the built workspace untouched, so the worked example ships beside the empty files
• renders src/docs/ to the repo root
• checks every skill description against the 200-character limit and fails the build if one is over
• keeps VERSION, plugin.json, and marketplace.json in step
• writes a single release ZIP

Then run python3 build/check.py. It verifies the generated output: no unresolved tokens, frontmatter and description limits, matching heading sets across the two platform renders of a behaviour, manifest versions in step with VERSION, a release ZIP that unzips to one folder, and a step 0 table in onboarding that names every file in src/workspace/. That last one exists because a template can ship in all three places and still never get created, since the only thing that copies it is a row in that table.

Bump VERSION before you publish. Users only get plugin updates when the version changes.

---

Credits and permission

Created by The NoW of Work.

MIT licensed. You may use it, edit it, adapt it for a client, rebrand it, and ship it commercially. Keep the licence file and the credit line. See CREDITS.md.

The Future is NoW.