Inside NanoClaw

Alright, this is where it gets technical.

When you asked Louis to talk like Phua Chu Kang, or to say hello every 5 minutes, something had to be written down somewhere. Otherwise they would have forgotten it the moment they went to sleep. This page shows you where.

You don't need to understand the code. The point is to see that every instruction you gave became an ordinary file, or an ordinary row in a list, that you can open and read.

TIP

You don't have to do this by hand. In the nanoclaw folder, start codex and ask it in plain words, for example: "Show me where Louis's personality is stored, and how it reaches them when they wake up." Codex reads the same files and code this page points to, and explains them. Use this page to check what it tells you.

Louis's folder

Every agent has one folder of its own, inside the nanoclaw folder. Louis's is groups/louis. NanoClaw puts this folder inside their Docker container. Apart from a few of NanoClaw's own files, it is the only part of your computer they can see.

In your terminal, go there and list what is in it:

cd ~/nanoclaw/groups/louis
ls

You see a list of names. These are the ones that matter:

Name	What it is	Who writes it
instructions.prepend.md	Their persona: who they are, their job, and their ground rules. It was copied from the template when Louis was made.	The template, then you. Louis doesn't change this one for style requests.
personality.md	Their personality notes: their voice, tone and language. It only appears once you ask them to change how they talk.	Louis
memory/	Their notes to themselves, like the link to your Notion page.	Louis
tasks/	A diary for each scheduled task: one file per task, with a line for every time it ran.	NanoClaw
plugins/	A copy of the community-assistant template: their skills and their three starting tasks.	Setup, when Louis was made
AGENTS.md	The instruction file Codex hands to the model: the instruction half of Louis's harness. NanoClaw writes it fresh every time Louis wakes up.	NanoClaw. Never edit it: your change would be gone the next time they wake up.

Where your personality change went

Read Louis's personality notes:

cat personality.md

You see what Louis wrote when you asked for Phua Chu Kang: a few lines about their voice, written in their own words. Compare these words with your coursemates, everyone's should be different depending on the model in use and/or how you phrased the request.

cat shows a file's contents in the terminal. If it says No such file or directory, Louis hasn't written any personality notes yet.

Now see how those notes reach them. Look at the top of the instruction file:

head -40 AGENTS.md

It starts with the line Composed at spawn - do not edit. Under it is a heading # Persona, with your instructions.prepend.md, and then # Personality, with your personality.md. After those come NanoClaw's own rules for every agent.

That is the whole trick. The model doesn't remember being told to talk like Phua Chu Kang. Every time Louis wakes up, NanoClaw glues the files together into AGENTS.md, and the model reads it again from the top. This is also why you ran make reload-community-agent: a running Louis has already read their file, and only a fresh start makes them read it again.

In the code, if you want to follow it:

• src/project-doc-compose.ts builds AGENTS.md. Search it for Persona and Personality to find where each file goes in.
• src/group-personality.ts reads personality.md. It skips the file if it grows past 4 KB, so a runaway personality can't push the ground rules out.
• container/CLAUDE.md is NanoClaw's rulebook for every agent. It is what tells Louis that "talk like…" requests go in personality.md, and never in their ground rules.

Where your scheduled tasks went

Scheduled tasks aren't kept in Louis's folder. NanoClaw keeps them itself, because NanoClaw is the one with the clock: it has to wake Louis up, and a sleeping Louis can't watch the time.

List them, from the nanoclaw folder:

cd ~/nanoclaw
pnpm ncl tasks list

You have seen this table before, in Schedule a message. To see one task in detail, put a name from the SERIES column in place of <name>:

pnpm ncl tasks get <name>

It shows the task's schedule, how many times it has run or failed, and the last few lines of its diary. The full diary is the file with the same name in groups/louis/tasks/.

The three tasks Louis started with

The daily brief, the week ahead and the memory clean-up came from the template. Each one is a small file: a schedule at the top, and the instructions under it. Open them on GitHub:

Task	Schedule	File
Daily brief	0 7 * * *, every day at 7am	daily-brief.md
Week ahead	0 18 * * 0, Sundays at 6pm	weekly-week-ahead.md
Memory clean-up	0 10 * * 0, Sundays at 10am	weekly-memory-hygiene.md

The schedules are cron codes. Their life has three steps:

1. When Louis was made, NanoClaw read the three files and created a task for each one, all paused. That way nothing posts into a community Louis doesn't know yet. The code is src/templates/tasks.ts.
2. The first time you talked to Louis, their welcome instructions told them to turn the tasks on. Read them in the template's welcome skill. Until you sent them your Notion link, each run checked, found nothing to report on, and posted nothing.
3. When you answered their questions about when you want summaries, they changed the tasks' times to match.

Your 5-minute hello

Your hello went through the same door. Louis turned "every 5 minutes" into */5 * * * * and ran NanoClaw's ncl tasks create command themselves. They know how, because NanoClaw puts the instructions for it in their AGENTS.md. Search that file for ncl tasks. The source is scheduling.instructions.md.

The limit that made Louis check with you first is in src/modules/scheduling/create.ts. Search for MAX_DAILY_FIRES = 4. Right under it is the warning NanoClaw gave Louis, word for word.

How a task wakes Louis up

Once a minute, NanoClaw looks for tasks that are due (src/host-sweep.ts). When one is, it starts a fresh Louis just for that task, and hands them the task's instructions instead of a message from a person. Louis does the job, posts in the group through a message tool, and goes back to sleep. Then NanoClaw writes a line in the task's diary, and works out when to run it next.

That is why your hellos could come a minute or two late. NanoClaw only looks once a minute, and Louis needs a few seconds to wake up.

Next

You know where everything lives. Some of it, like personality.md, exists only because we customised NanoClaw for this course. See what we changed: Our modified NanoClaw.