Claude Code After Hours

Where this is heading

The steps of AI adoption

My current north star.

• Boris Cherny — creator of Claude Code — published the ladder in July 2026
• Section 9 is sometimes step 1, sometimes step 2
• Our top priority is getting to step 3

Claude Code · the daily cockpit

What I use every day

All Claude services — open all day, every day.

• Anything browser-related runs through the Claude Code desktop app
• The CLI is local background-agent management
• Cloud is for the phone and the automated agents

Surfaces · the design rule

Harness Calculus

Conceptually, we think about all surfaces.

• In practice, this is hard
• Never give up thinking this way
• Always be learning

The meta-repo · Ghost

One repo runs the company

Managing second-brain data is extremely hard.

• Optimizing for security, speed, and flexibility
• Context at my fingertips
• Security is managed through submodules

The meta-repo · what runs from here

Company Stack

Judging criteria:

• Security: looking for enterprise certs + subprocessors list
• Speed: agent-first — can be used without opening a browser
• Purpose-specific systems that we don’t want to roll our own

@ghost · the transport

Every chat app, one inbox

Centralized comms are important for throughput.

• By default, we never poll — push-based only
• Claude does the reading and writing
• All outgoing messages must be approved by me

The prerequisite

Auto mode is mandatory

None of this is possible without auto mode.

• Use auto mode
• Use auto mode
• Use auto mode

@ghost · the tag

Tag the company into any chat

Type “@ghost” in any Beeper chat. Ghost picks up the task.

• Multiplayer AI — but budget
• Certain security settings allow collaborators to dispatch work in the chat
• Same idea as Anthropic’s @Claude tag

@ghost · trust

Everyone gets the agent

Security is a big deal.

• Allowed chats are whitelisted
• Further security permissions exist for allowed chats
• Scary to operate

@ghost · under the hood

Ghost tag architecture

High-level view.

• A tiny bun process, fired by launchd every minute
• One dispatched session per chat — it stays warm and absorbs follow-ups
• Everything runs on the workstation — no server anywhere

The backlog · the belief

Backlog System

All work is tracked. If it isn’t in the backlog, it doesn’t exist.

• One system, one ingestion route. From any surface.
• Items carry their own state: what’s next, what’s blocked, what’s due
• All of this captured in the file system

The backlog · dispatch

Capture → dispatch

An item can be handed to a background agent with one command.

• ”Work on GHOST-231” spins up a session that already knows the context
• The item is the memory: every session it touched, logged on the record
• Guardrails for operational context is hard

The backlog · the impetus

Fable said: keep a daily error log

”Spend ten minutes every day deliberately writing down what you got wrong.”

• Not journaling, not gratitude lists — an error log
• I’d known I needed this for a long time — this answer got me moving
• The next slide is that habit, built into the system

The backlog · sessions

Every work session is a record

Sessions open, checkpoint, and close on their own.

• Begin: the session files itself and locks onto the backlog items it serves
• Close: outcomes written back to the items — nothing evaporates when the window closes
• My calendar shows what the COMPANY did, not what I remembered to log

Surfaces · the worked example

One command, every social platform

”/post” — one line from ANY surface becomes a draft for X, LinkedIn, Threads, and the rest.

• Every draft is checked against our publishing rules before it can ship — automatically
• The queue lives server-side — start a post at the desk, approve it from the street
• No platform tabs, no copy-paste — the system holds the keys, not the laptop

Routines · while you sleep

It works while you sleep

Configure once — it runs on a schedule, a webhook, or an event.

• The daily brief, the inbox triage, the new-ticket triage — no human needed to kick it off
• You wake up to finished work and a summary of what happened
• This is the employee model, not the chatbot model

Routines · one real run

The Friday-night deck routine

Every Friday at midnight, an agent builds Saturday’s AI-news deck — while I sleep.

• Pulls the week’s saved bookmarks + researches the news window
• Writes every source into the database, authors the slides, checks them against our publishing rules
• Opens a pull request; Saturday morning the deck is live on the site

Routines · philosophy

Philosophy

Build your agent. Let your agent work for you.

The meta-repo · the mental model

TIP: Catch yourself reaching for the mouse

Every time you click through an app by hand, pause: why am I doing this?

• If you can describe the task, you can delegate the task — Claude should be doing it
• That one habit — noticing the clicks — is how a business becomes Claude-native

Start here

Your move

• Start a session with Fable
• Tell Fable the thing you wish were automated — but don’t have guardrails for, because you don’t know how to build them
• Execute