In partnership with
AI-HOY, AInauts!
Welcome to a new edition of your favorite newsletter.
Broken weeks can be surprisingly productive.
Ours included 32 percent of an agent budget expiring, a clock that showed the wrong time, and concert tickets we suddenly could no longer access.
By the end, we had three solutions and a new issue for you: how to use leftover AI budget well, control real devices with ChatGPT Work, and write better assignments for more autonomous agents.
Here is what we have for you today:
🃏 The Joker List for Your Leftover Budget
📲 How Your Agent Controls Real Devices (...and Saved Our Night)
🤖 Agents Are Ready. We Often Are Not.
And if the word “agents” already makes you want to tune out: we mean programs such as ChatGPT Work or Claude Cowork that do not merely answer a task, but complete it for you across several steps.
Let’s get started.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
🃏 The Joker List for Your Leftover Budget
Thirty-two percent. That was how much weekly budget we still had when the reset arrived and the meter jumped back to full.
Almost a third. Not saved. Not carried over. Simply reset. That is annoying.
If you use ChatGPT Work/Codex or Claude Cowork/Code in the desktop app, you know the usage meter. Depending on your plan, time windows and weekly limits reset regularly.
You can run out before the work is finished. Or you finish early and discover that a surprising amount remains.
So we asked ourselves which useful jobs we could still hand off shortly before the reset.
An idea list currently circulating online mostly asks one question: How do I burn the rest? This is also called token-maxxing.
Half-built 3D games, image series, parallel agent swarms. Fun, but there is more value available.
We would rather finish work that would otherwise sit around for another week. That is why we created the Joker List.
Jokers Almost Everyone Can Use
The best Jokers are unspectacular. That is exactly why they remain undone.
Inspect the project graveyard: Search a selected work folder for half-finished ideas, contradictory versions, and projects with no next step. Deliverable:
project-graveyard.mdwith findings, safe suggestions, and a small cleanup plan. Safety rule: delete or move nothing.Collect meeting debt: Review one month of minutes and transcripts. Gather promises, open items, and decisions that were never recorded properly. Deliverable:
open-items-month.md. Safety rule: anything unclear remains explicitly marked unresolved.Build an agent pack: Find things in your files that you have solved or explained several times. Turn three of them into three small agents in
agent-pack/, as checklists, templates, or prompts. Each gets permitted inputs, stop rules, an output format, and test material.Create a second brain: Turn approved chats, notes, and project folders into
second-brain-YYYY-MM-DD/with an index, structure map, templates, and source links. As a test, the package must answer one real question from your own material.Give content a second life: Review existing articles, presentations, videos, or newsletters. Find strong material that is outdated, poorly distributed, or suitable for another format. Deliverable:
content-recycling-plan.md. Safety rule: publish or rewrite nothing.
The Prompt That Builds Your Own List
Of course, not every example fits you. The best entries are already on your drive, and an agent can find them faster than you can.
Language note: the full AInauten Joker List prompt is currently available in German.
Our detailed AInauten Joker List prompt helps adapt the system to your projects, folders, and working style.
If you want a quick taste first, use this shortened English version:
Build me a Joker List of jobs I can run unattended with leftover weekly budget.
Review what you know about me: projects, folders, open threads, recurring tasks, notes, and installed tools. The examples below are only models for scope and wording, not for the content.
Give me 15 suggestions tailored to me. For each, include the expected result, estimated runtime, required permissions (read only, write, network), and what could go wrong. Sort by value per runtime and write everything to joker.md.
[Insert the examples above here]Keep the List From Falling Asleep
A list you must maintain manually will be dead in two weeks. The trick is to let the agent collect candidates while the ideas are already emerging.
Our long Joker prompt helps set up three separate components: joker.md, an optional block for your Custom Instructions, and a recurring suggestion task. The agent creates the selected parts only after your explicit approval.
Custom Instructions can recognize new candidates during normal work, but they cannot wake up on a schedule. A Saturday 8:00 a.m. task can handle that instead. Claude Code calls them Routines; Codex calls them Automations. The task updates and sorts the list. You still decide which Joker to start.
For developers, Vox’s daily report is also interesting: find and verify safe overnight jobs, but start neither code nor agents automatically.
Our Take: Turn Leftover Budget Into Decisions
The goal is not to consume as many tokens as possible. Your leftover budget should turn neglected work into a file you can use the next morning. An agent that works for three hours and leaves nothing behind has failed.
Make longer jobs save intermediate results continuously. If a run hits the limit, the entire effort should not disappear.
After the reset, the best outcome is a tested package you can reuse the following week.
The Future Today: Get the AI & tech news that actually matters. With one email, and only five minutes, you can stay ahead of 99% of the world. The Future Today is read by teams at Anthropic, Microsoft, OpenAI, NVIDIA, and Salesforce. Join Free Today →
📲 How Your Agent Controls Real Devices (...and Saved Our Night)
A small device on our desk measures air quality. It works.
One problem: the clock was wrong.
After tapping around on the display, one thing became clear: the time could not be changed there.
So we connected the device to the Mac and asked ChatGPT Work.
The result: USB only charges this model. The actual control path runs through the Temtop app and Bluetooth. After we connected the cable to a Google Pixel phone and approved access, ChatGPT Work installed the app. Two clicks later, the clock was correct.
Very nice.
Visual note: the interface in the following screenshot is from our German setup.
This is exactly where an agent can help. It looks for the bridge through which a device or program can be controlled: a manufacturer app, Bluetooth, command line, ADB, API, MCP, or whatever the interface happens to be called.
We now keep a permanent thread for this in ChatGPT Work: clean up the drive, adjust a VLC setting in a video app, configure monitor profiles. That thread definitely knows the machine better than we do. 😁
You can even have it analyze and optimize your Wi-Fi coverage. In this example, the agent uses the /goal command so the desired result and acceptance criteria remain available throughout the run.
Computer Use in ChatGPT Work/Codex can operate a screen you share, work across several apps, and ask before app access or other sensitive steps.
If you do not have a stubborn device nearby, you can let ChatGPT Work inspect your computer instead.
Language note: our detailed System Caretaker prompt is currently in German.
That is why we built our System Caretaker prompt.
For a specific device problem, this English starting prompt is often enough:
I want to solve this problem:
[Device/software and the specific problem].
Find the controllable bridge through which the solution can be reached: manufacturer app, Bluetooth, USB data connection, driver, ADB, command line, API, MCP, and so on.
Check what can be done through the connected device. Operate only apps and interfaces I explicitly approve.
Stop and ask before changing firmware, resets, accounts, permissions, or data. If no controllable bridge exists, say so directly and name the next manual step.The next case arrived the following day, and it was the real night-saver. The AXS app holding all our event tickets simply stopped letting us sign in on the phone. 😱 Neither the app nor the mobile browser worked. Desktop login did.
Two interacting problems with passkey and password login were behind it.
Honestly, we would never have found that on our own.
ChatGPT Work found and fixed the problem. Tickets restored. Night saved. Party on.
Visual note: the following screenshot comes from our German device setup.
Our Take: Find the Control Path First
In both cases, the real problem was not the device. It was identifying how to reach it and defining the desired end state.
That is where an agent is strong. It saves you from hunting for the right app, setting, or connection before you lose hours in support forums.
🤖 Agents Are Ready. We Often Are Not.
The models are strong. ChatGPT Work and Claude Code can operate for hours. Agent platforms are appearing everywhere.
This week, three ads from three different companies appeared in the same New York subway car. All three promoted agents.
The topic has reached the mainstream. Quiz question: why is the big breakthrough in daily work still taking so long?
We do not find that surprising.
For a long time, we were comfortable on the chatbot treadmill: task in, draft out, corrections afterward. It works remarkably well because we humans hold the missing context and repair every wrong turn immediately.
That works less well with an agent. It works alone for longer, uses tools, and makes more small decisions. Prompting suddenly becomes work organization.
The most important reason agents have not spread widely despite mature technology is unexciting, and it is not a technical weakness:
Leading agents resembles delegating to employees. And quite a few people are only so-so at that.
Microsoft’s Work Trend Index 2026 distinguishes four ways of working with AI: asking, exploring, collaborating, and delegating. Advanced users know which mode fits which task.
In Microsoft’s analysis, company culture, manager support, and suitable working practices are more than twice as strongly associated with reported AI value as individual factors.
Good agents need more than good prompts. They need an environment where the goal, permissions, and quality bar are clear.
At the same time, agents are gaining more power to act.
In Australia, an agent was asked to move its user forward on a gym-class waitlist. It found a gap in the booking system and canceled another person’s reservation. ABC News reconstructed the case.
The agent was not malicious. It pursued its goal creatively and took a shortcut that would have been obviously off-limits to a person.
Cloudflare shows how much automated traffic is already moving across the web: more than 50 percent of the internet traffic observed by its systems is now nonhuman.
Not every bot is an agent. But once software acts on its own, a goal alone is no longer enough.
The machine is leaving the chat. That is exactly why prompting alone no longer works.
Our Take: Good Agents Need Good Managers
We can wait another ten model generations. A vague assignment will remain a vague assignment.
The biggest agent advantage therefore lies with us: clear direction, useful context, and a measurable definition of done.
More autonomy is the wrong demand right now. An agent with an unclear goal does not suddenly produce better work. It simply travels in the wrong direction more independently.
That’s it. See you in the next issue.
Reto & Fabian from AInauten
