OpenClaw

OpenClaw: Eight jobs to try

This is the fifth of eight OpenClaw pages, and it assumes you have finished Files, the browser, and memory. The eight jobs below go from easy to ambitious, and each one uses something OpenClaw's docs show it can do: files in its workspace, PDFs, its own browser, memory files, schedules, and sub-agents that work at the same time.

These jobs match the docs as of 14 November 2026. You paste the same words on Windows and on a Mac.

To see why these prompts are worded the way they are, and how to adjust them for a strong cloud model or a small local one, read Prompting strong and weak models.

How to ask

Each request below says what the finished file is, where to save it, and what not to touch. These are the safety rules every job repeats:

The approval switch stays on ask and the sandbox stays off, so the agent is working on your real computer. The words in the request, together with your answers to its approval requests, are what keep each job small.

When a job stalls or loops, press Stop in the chat page and see Fixes.

1. Turn a messy list into a tidy table

This job needs no web and no commands, so it is a good first test of how well the model follows instructions. A CSV file opens in Excel, Numbers, or Google Sheets.

Turn the list below into a table and save it as expenses.csv in your workspace. Give each expense a row with the item and the amount, read amounts written as words, and add a total row at the bottom. Only work inside your workspace. Do not run commands, do not open any website, and do not delete or move anything.

coffee $4.50 LUNCH 12 bus ticket 3.50 yesterday groceries - twenty dollars on Saturday

For the sample list, the four amounts add to $40. If you swap in your own list, add the amounts yourself to check the total.

Safety note: use made-up or low-stakes numbers, because everything you paste goes to the model provider.

2. Summarize a PDF you attach

OpenClaw has a PDF tool that reads up to 10 PDFs at once, so this job makes a short brief from a document. Pick something public, such as the manual for a kitchen gadget, not a bank statement or a medical letter.

Attach the PDF in the chat page, then send:

Read the PDF I attached. Save a one-page brief as manual-brief.txt in your workspace. Start with what the product does, then the three steps a new owner most needs, then any safety warnings the PDF gives. Only work inside your workspace. Do not run commands and do not open any website.

If the agent says it cannot reach the attachment, copy the PDF into your workspace folder and name it in the request instead.

Safety note: a PDF can hide a prompt injection, so if the agent suddenly asks to run a command or open a website, deny it.

3. Compare two products with the browser

This job uses the agent's own browser profile, which is separate from yours and signed in to nothing.

Use the browser to open public product pages for a 16 GB USB drive and a 32 GB USB drive. Compare the price and one difference that matters, and save the notes as usb-compare.txt in your workspace with the web address of each page you used. Do not sign in, do not add anything to a cart, do not fill in any form, and do not buy anything.

When it finishes, open the page addresses in your own browser to check the prices.

Safety note: never switch the agent to the user browser profile for a job like this, because that one is signed in as you.

4. Teach it how you like summaries

OpenClaw keeps preferences in plain files, so this job saves one and then proves it worked. Send these as two separate messages.

First message:

Remember this preference. When I ask for a summary, use four short lines, put the most useful fact first, and skip any greeting.

Second message, after it says the preference is saved:

Summarize the note below, using the summary preference you saved.

The community garden opens at 8 on Saturday. Plots are free for neighbors who live within ten blocks. Tools are in the shed, and the hose was broken last week.

The second reply should be four short lines with the useful fact first. To see where it saved the preference, open USER.md or MEMORY.md the same way as on page 3.

Safety note: memory is a note, not a lock, so a saved rule never replaces the approval switch.

5. A short note every morning

Schedules are built into OpenClaw, and this job uses one that only writes a file. Change the time to suit you before you send it.

Every morning at 7:30, while this computer is on, write a file in your workspace named with today's date, such as note-2026-11-15.txt. Put one short gardening tip in it. Only write in your workspace. Do not run commands, do not open any website, do not send messages to anyone, and do not spend money.

To see the job, paste this into your terminal:

openclaw automations list

You may also see a job you did not create: OpenClaw's memory dreaming sets up one repeating job of its own. To remove your morning note later, find its ID in openclaw automations list, then run openclaw automations remove <jobId>, and run the list again to check that it is gone.

The computer and the Gateway both have to be running at 7:30, or that day's run does not happen.

Safety note: nobody is watching a scheduled job, so keep it to writing files until you have seen it run several times.

6. Check a price once a day

This is the riskiest job, because it combines a schedule with the browser, so it runs unattended on pages you do not control. Do it only after job 5 has run cleanly for a few days.

Once a day at noon, while this computer is on, use the browser to open this page: [page address]. Add one line to price-log.txt in your workspace with the date, the price, and the page address. Only write in your workspace. Do not sign in, do not add anything to a cart, do not buy anything, do not run commands, and do not send messages to anyone.

To see how the runs went, find the job's ID in openclaw automations list, then paste this with that ID in place of <jobId>:

openclaw automations runs <jobId> --limit 20

Safety note: a page can hide a prompt injection, and with no one watching, a command the job asks for gets no answer, which OpenClaw's documented fallback treats as a refusal. Keep the approval switch on ask so that refusal stays in place.

7. A deep-research report with a source for every claim

OpenClaw ships a Deep Research plugin that is on by default. Its deep_research tool searches, reads pages, and returns a report in which each finding rests on an exact quote from a source. A separate check looks for claims the sources contradict, and anything it could not find is listed as a gap.

The tool uses your web tools, so choose a search provider first, as on page 3, or put the page addresses you want it to read into the request.

Use your deep_research tool to answer this question: [your question, such as whether a heat pump would save money in a cold climate]. Ask it for a short answer, each finding with the web address of its source, the points where the sources disagree, and any gaps. Save the report as research-report.md in your workspace. Only work inside your workspace. Do not run commands, do not sign in to anything, do not send messages to anyone, and do not buy anything.

By default one research job makes at most 8 model calls, 6 searches, and 12 page reads, reads at most 2 pages at the same time, and stops after 180 seconds. A source it runs out of budget for shows up as a gap rather than as a longer run. If the agent says it has no such tool, type /tools in the chat to see its tool list.

When it finishes, research-report.md is in your workspace with findings, page addresses, disagreements, and gaps. Open a few of the addresses to check the quotes.

Safety note: the docs say the tool's source filters are "evidence filters, not a network firewall," and that its checks "do not establish truth or completeness," so treat the report as a starting point.

Cost note: set a spend limit at your provider before you start, because a research job can use many times the tokens of a normal job. The docs say its token totals "are observed, not a hard token or dollar cap."

8. Split a research job across sub-agents

OpenClaw can spawn sub-agents, which are background runs that each work in their own session and report back to your chat. They run at the same time, not one after another, and by default one chat can have up to 5 of them active at once.

Sub-agents start without your memory files, such as USER.md, so put everything they need into the request.

Compare [first product], [second product], and [third product]. Spawn one sub-agent for each product, working at the same time. Each sub-agent should read public pages to find the price, one strength, one drawback, and the web address of each fact, and report back without writing any file. When all three have reported, write compare.csv in your workspace with one row per product. Only work inside your workspace. Do not run commands, do not sign in to anything, do not add anything to a cart, do not send messages to anyone, and do not buy anything.

To watch them, open Subagents from your conversation in the chat page, where each run is listed under Running or Finished. Press Stop, or type /stop in the chat, to end the main job and every sub-agent under it.

When it finishes, compare.csv is in your workspace with three rows, each with its page addresses.

Safety note: sub-agents get the same tools as the main agent, and with the sandbox off they work on your real computer. If one asks to run a command, deny it.

Cost note: set a spend limit at your provider before you start. The docs say "each sub-agent has its own context and token usage," so this job can use many times the tokens of a normal one.

Sources

All checked 14 November 2026.

Next

Continue to Fixes.

Checked 14 November 2026 · prices, limits, and versions change