QwenPaw

QwenPaw: Playbook of eight jobs

This is page 5 of 8. It lists eight jobs in order from easy to ambitious, each with a prompt you can paste that already has its safety limits written in.

Every job assumes the setup from the earlier pages: STRICT on your agent, all seven shell evasion checks on, and execute_shell_command off unless a job says otherwise. Read each approval request before you click, and reject anything the prompt did not ask for. The abilities used here are the ones the QwenPaw docs describe as of 14 November 2026, and a small local model may struggle with the later jobs. For how to word a request so that a strong model or a weaker one gets it right, see Prompting strong and weak models.

1. Turn a messy list into a spreadsheet file

This is the gentlest job, because it uses only the file tools in the agent's own workspace. Paste a rough list straight into the chat, such as your monthly bills with amounts and due dates in any order. The agent tidies it into a CSV file that any spreadsheet program can open.

Here is a rough list of my monthly bills: [paste your list]. Turn it into a CSV file called bills.csv in your own workspace folder, with the columns name, amount, and due day, sorted by due day. Use only the file tools. Do not run commands, open a browser, or touch any other folder.

Safety note: leave out account numbers and passwords, because everything you paste goes to your model's provider unless you use QwenPaw Local.

2. Summarize a document you attach

QwenPaw has built-in skills for PDFs and Office files, plus file_reader for plain text, so it can read an attached copy and write a short summary. Pick something long but not private, like a product manual or a public report, and attach a copy rather than the only one you have.

I attached a document. Read it and write a summary of no more than ten bullet points, plus a list of any dates or deadlines it mentions. Save the summary as summary.md in your own workspace folder. Do not change the attached file, do not run commands, and do not open a browser.

Safety note: this guide could not confirm where QwenPaw stores attachments, so do not attach anything with private details in it.

3. Compare three products from public pages

This job adds the web. With web_search and web_fetch turned on in Workspace → Tools, the agent can look up three products and put what the pages say side by side. Keep the browser tool off, so nothing can sign in or click a button.

Compare three popular rice cookers under 100 dollars. Use web search and fetch only public product or review pages, and list each page address you used. Put the price, capacity, and one common complaint for each in a table in a new file called rice-cookers.md in your own workspace folder. Do not sign in anywhere, do not buy anything, do not fill in any form, and do not run commands.

Safety note: any page can carry a prompt injection, so reject a step that wants to visit a site you did not expect or write outside the workspace.

4. Get a morning news digest in your Inbox

QwenPaw's built-in news skill, in the docs' words, fetches and summarizes "latest news from configured sites," and a scheduled task can run it every morning. First run the prompt below by hand in a chat and check the result. Then, under Control → Cron Jobs, click Create Task, give it a name, pick a daily time, choose the agent task type, paste the same prompt, turn on Save result to Inbox, and click Save. Click Execute Now once to test it before you leave it alone.

Use the news skill to find the five most important technology news stories from the last day. For each one, give a two-sentence summary and the page address. Do not save or change any file, do not run commands, and do not sign in to any site.

Safety note: before you schedule it, add execute_shell_command to the denied tools list under Settings → Security → Tool Guard, because a run with nobody watching may log a danger and keep going.

5. Turn a job that worked into a skill

When a job from this list has gone well, /make-skill can save the steps as a reusable skill. The docs say the agent first proposes a plan with a name, the steps, and a testing choice. After you approve, it saves the new skill to your workspace "enabled by default." Run it in the same chat where the job succeeded, and pick Smoke test so the skill is tried once on a small task.

/make-skill monthly bills CSV from a pasted list

When the plan appears, read it and reply with something like this:

Use Smoke test and keep Batch disabled. The skill must only read what I paste and write files in your own workspace folder. It must not run commands, open a browser, or contact any website.

Safety note: set the Skill Scanner to Block under Settings → Security first, and if the plan lists a step you did not ask for, reply that you want it removed before you approve.

6. Give research its own agent

QwenPaw can run several agents, each with its own workspace, tools, skills, and chat history. A separate research agent keeps web jobs away from the files your main agent works on. Go to Settings → Agent Management, click Create Agent, and fill in a name such as Research, a short description, and an ID such as research. Then pick it from the agent selector in the top-left corner and give it this first job.

You are a research agent. Find three recent articles that explain how heat pumps work, using web search and fetch on public pages only. Save a one-paragraph summary of each, with its address, in a file called heat-pumps.md in your own workspace folder. Do not sign in anywhere, do not run commands, and do not touch any folder outside your workspace.

Safety note: this guide could not confirm that a new agent copies your STRICT setting. Before the first job, set the new agent's approval level to STRICT and turn off its command and browser tools.

7. Split a comparison across sub-agents

QwenPaw's spawn_subagent tool starts sub-agents inside your agent's own workspace. Each one runs as the same agent, with the same settings and tools, in a fresh session that is thrown away when it finishes. A plain call makes the main agent wait for each sub-agent, so plain calls run one after another. The docs also describe a background mode, where the main agent hands off a task, keeps working, and checks back later. The v2.3.0 code starts each background sub-agent as its own task and can send up to 10 in one batch, so several can work at the same time.

Because sub-agents share your agent's tools, the setup from the earlier pages applies to them too. The command sandbox starts switched off in v2.3.0, so this job does not count on it and uses no commands at all. Turn on web_search and web_fetch in Workspace → Tools as in job 3, and keep execute_shell_command and browser off.

Compare three cities for a weekend trip from [your city]: [first city], [second city], and [third city]. Use spawn_subagent in background mode to start three sub-agents at the same time, one for each city, and give each one only the web_search and web_fetch tools. Each sub-agent should find two sights, one free activity, and the usual train or bus fare from [your city], with the page address for each fact, using public pages only. When all three have reported back, write a file called weekend-trip.md in your own workspace folder, with a section for each city and a table that puts the three side by side. Do not run commands, do not open a browser, do not sign in anywhere, do not buy or book anything, and do not touch any folder outside your workspace.

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

Safety note: the code sends each sub-agent's approval requests back to your chat, but this guide could not confirm how the Console shows them. Expect many requests under STRICT, and reject any that are not a search or a fetch of a public page.

8. Write a deep-research report

This guide could not find a built-in deep-research mode in the QwenPaw docs, so this job asks for one in plain words: many sources, a source for every claim, and a clear note wherever the sources disagree. It uses the same web_search and web_fetch tools as job 3, with one agent and no sub-agents.

Research this question: [your question]. Use web search and fetch on at least eight public pages, and prefer the original source over pages that only repeat it. Save a report called research-report.md in your own workspace folder. Start with a summary of no more than one page, and put the page 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 do not answer and a numbered list of every source with its address. If you cannot find a source for a claim, leave the claim out. Do not run commands, do not open a browser, do not sign in anywhere, do not buy or send anything, and do not touch any folder outside your workspace.

When it finishes, open two or three of the listed pages yourself and check that they say what the report says they say. Treat the report as a well-sourced first draft, because a model can still misread a page.

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, so reject a step that wants to visit a site you did not expect, sign in, or write outside the workspace.

Sources

Checked 14 November 2026.

Next

Continue to Fixes.

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