NanoClaw: The playbook

These eight jobs fit what NanoClaw's docs show it doing well: working on files in its own workspace, reading the web, remembering rules, running on a schedule, answering from your phone, and handing work to a second agent. They follow Phone and schedules and run from the easiest to the most ambitious, so do them in order.

The jobs match the docs as of 14 November 2026. When one stalls or surprises you, use Fixes. For how to word a request so that a strong model or a weaker one gets it right, see Prompting strong and weak models.

How to send a job

Each job gives you words to paste. From the terminal, put them between the double quotes of a pnpm run chat command, like this:

cd ~/nanoclaw
pnpm run chat "Paste the words of the job here."

Leave out the characters !, $, and " inside your words, because the terminal treats them as part of the command rather than your message. Once you have paired a chat app on the previous page, you can paste the same words there instead.

The terminal command stops waiting a couple of seconds after the first reply, so a long job may finish after the command has ended. The agent remembers the conversation between commands, so ask "Did you finish? Tell me the file name." to get the result.

Every job below repeats the same limits inside its words: stay in your own workspace, install nothing, ask for none of my folders, and sign in to nothing. Those limits matter because the commands inside the container do not ask you first.

1. Tidy a messy list into a file

This job needs no web, so it works even when egress lockdown blocks a web tool. It is a good first test of how the agent handles files.

Turn the list below into a tidy table in a CSV file named expenses.csv in your workspace. Give each expense its own row with the item and the amount, read amounts written as words, and add a total row at the bottom. Stay in your own workspace, install nothing, and do not ask for any of my folders. The list: coffee 4.50, LUNCH 12, bus ticket 3.50 yesterday, groceries twenty dollars on Saturday.

Safety note: the agent only writes inside its own workspace, so the worst case is a wrong file you can delete.

What you should see a reply naming expenses.csv, and a total of 40 when you open it with cat after finding it the way you did on the files page.

2. Research a question and save the sources

The agent can read public pages with its built-in browser and search tools. This job keeps the answer short and makes the agent show where each fact came from, so you can check it.

Find three current facts about how long a 16 GB USB drive usually lasts. Use only public pages. Save a short note named usb-facts.txt in your workspace with each fact and the web address it came from. Do not sign in to anything, do not fill in any forms, do not buy anything, and install nothing.

Safety note: any page can carry a prompt injection, so the words keep the agent logged out and away from forms and carts.

What you should see a file named usb-facts.txt in the agent's workspace, with three facts and a web address beside each one. If lockdown blocks the web, the agent says so instead.

3. Give the agent a standing rule

NanoClaw keeps lasting rules about how the agent behaves in a file called instructions.prepend.md, and the agent can write that file itself when you ask. The docs' own example of a standing rule is to confirm before deleting anything, which is a useful one to start with.

From now on, always confirm with me before you delete any file, and keep your answers to five lines or fewer. Save this as a standing rule in your instructions file. Do not change anything else, install nothing, and do not ask for any of my folders.

Then read what it wrote, from your own computer:

cat ~/nanoclaw/groups/*/instructions.prepend.md

Safety note: a standing rule shapes what the agent tries to do, but it is not a wall, so keep relying on the container and the empty folder list.

What you should see the two rules in the file, and later replies that stay within five lines.

4. A daily briefing saved to a file

This is the first job that runs with nobody watching. The docs say a daily briefing is a good fit for a plain schedule, and each run is a paid call to the model, so check that your spend limit is set before you start.

Every day at 8am, read two or three public news pages about [a topic you choose] and save a five-line summary named briefing-[date].txt in your workspace, with the web address of each page. Only read public pages, do not sign in to anything, do not send any messages, install nothing, and do not ask for any of my folders. Run this once a day only.

Check that it exists, and stop it any time:

ncl tasks list
ncl tasks cancel --all

Safety note: an unattended job repeats any mistake every day, so this one only reads public pages and writes into its own workspace.

What you should see the task in ncl tasks list with its next run time, and a new dated file in the workspace after each run.

5. A reminder that reaches your phone

A terminal only shows replies while its command is running, so a reminder needs a chat app to reach you. Do this job only after you paired Telegram on the previous page and set it to strict. Send these words from that chat:

Remind me tomorrow at 2pm to review the budget. Send the reminder only to this chat, and do not do anything else when it fires.

Safety note: the bot answers anyone the chat policy lets through, so check that only your own account is paired before you rely on it.

What you should see the agent confirms the reminder, the task appears in ncl tasks list, and the reminder arrives in Telegram at about 2pm the next day.

6. Hand research to a second agent

NanoClaw can create a second agent with its own container, workspace, and memory, and the two pass messages to each other. The docs warn that there is no built-in limit on how long two agents keep messaging each other, so these words give the second agent a clear stopping point.

Create an agent called researcher. Its job is to take one research question from you, look it up on public pages, report its findings back to you once with sources, and then wait. It must not sign in to anything, install anything, or message anyone except you. Then ask researcher for three facts about growing basil indoors, and give me a five-line summary of its answer.

When you are done with it, list your agents and remove the extra one, putting its id where it says <id>:

ncl groups list
ncl groups delete --id <id>

Safety note: the first agent that setup made is trusted, so it creates the second one without asking you, and each extra agent can spend through your key.

What you should see a five-line summary in the reply, and a new researcher entry in ncl groups list until you delete it. The delete command does not remove the agent's folder under ~/nanoclaw/groups/, so delete that folder yourself if you want it gone.

7. Split one research job between two agents

Job 6 handed work to one extra agent, and this job splits one question between two. Each agent runs in its own container, and the docs describe such an agent as a parallel specialist that works on its own and reports back, so the two can work at the same time. NanoClaw cannot send one message to several agents at once, so the main agent sends each one its part in turn.

The docs say there is no limit on how many agent containers run at once, and no built-in limit on how long agents keep messaging each other. That is why these words give each new agent a clear stopping point.

Create two agents called heating-researcher and appliance-researcher. Give each one this role: take one research question from you, look it up on public pages, report back once with the web address of each fact, and then wait. Neither may sign in, install anything, create other agents, or message anyone but you. Ask heating-researcher for three ways to cut the cost of heating and cooling a home. Ask appliance-researcher for three ways to cut the cost of running appliances and lights. When both have replied, save energy-plan.md in your workspace with a section for each agent and the web address beside each fact. Stay in your own workspace, install nothing, ask for none of my folders, sign in to nothing, and do not buy or send anything.

The two agents reply after the terminal command has ended. Wait a few minutes, then ask "Did you finish? Tell me the file name." When you are done, remove both agents and their folders as you removed researcher in job 6.

Cost note: set a spend limit at your provider before you start. Each agent holds its own conversation with the model, so this job can use many times the tokens of a normal job.

Safety note: the trusted first agent creates both agents without asking you. The docs say any other agent that tries to create one needs an admin to approve it, so reject any such request you did not expect.

What you should see a file named energy-plan.md in the workspace with two sections and a web address beside each fact, and two new entries in ncl groups list until you delete them.

8. Write a deep-research report

NanoClaw's docs show web search and page reading, but no deep-research mode, so this job asks for one in plain words: many sources, a source for every claim, and a clear note wherever the sources disagree. One agent does all the reading, so nothing new needs cleaning up afterward.

Research this question: [your question]. Use at least eight public sources, and prefer the original source over pages that only repeat it. Save a report named research-report.md in your workspace. Start with a summary of one page or less, and put the web address of a source after every claim. Then list each point where the sources disagree, and say which sources say what. End with the open questions the sources leave and a numbered list of every source with its web address. If no source backs a claim, leave the claim out. Use only public pages and stay in your own workspace. Install nothing, ask for none of my folders, sign in to nothing, and do not buy or send anything.

A job this long finishes well after the terminal command ends. After several minutes, ask "Did you finish? Tell me the file name." Then open two or three of the listed pages and check that they say what the report says.

Cost note: set a spend limit at your provider before you start. A deep-research job reads many pages, and every page it reads goes into the prompt, so this job can use many times the tokens of a normal job.

Safety note: more pages mean more chances of prompt injection. Commands inside the container do not ask you first, so the words keep the agent logged out and inside its workspace.

What you should see a file named research-report.md in the workspace, with a summary, a section on disagreements, open questions, and a numbered source list. If lockdown blocks the web, the agent says so instead.

Sources

Opened 14 November 2026.

NanoClaw:

• CLI channel, for one-shot pnpm run chat, the short wait after the first reply, and replies that are not shown later.
• Quick start, for terminal replies only showing while the command runs, and a chat app for reminders.
• Give agents tools, for the built-in browser and search.
• Customize an agent, for instructions.prepend.md and the confirm-before-deleting example.
• Schedule recurring work, for daily briefings, paid runs, the one-shot reminder example, and ncl tasks.
• Build a multi-agent team, for create_agent, the trusted first agent, the missing loop limit, and cleanup.
• The same multi-agent page, for one container per active agent, no limit on containers running at once, no way to message several agents at once, and approval for agents other than the first.
• Agent tool instructions at v2026.10.0, for a created agent working as "a parallel specialist that works independently and reports back."
• README, for web access to "search and fetch content from the web."
• ncl CLI reference, for ncl groups delete leaving the folder behind.
• Security model, for what commands inside the container can do.

This guide's own notes:

• Safety versus capability.

Next

Continue to Fixes.