I pay >2300$ a month for my AI coding assistants

Here is how I replace a 5-person dev team.

Author

Martin Musiol
July 24, 2025 • Estimated Reading Time: 9 minutes

AI enthusiasts,

After a couple of weeks of a break (vacation + 3 weeks of heavy workload), I am back!

I was busy building AI solutions for clients e.g. a global NGO, but also others.

Now, my monthly bill for Cursor is just over 2300 $.

This screenshot shows a German-language account statement or billing list. Each row has a small blank checkbox on the far left, a date in German, a description of the charge, and an amount in euros. For instance, the first entry reads “22. Juli” followed by “CURSOR USAGE MID JUL NEW YORK” with an amount of “87,84 €” and a right-pointing arrow on the far right. The next rows on 18. Juli and 16. Juli have the same description with amounts of “88,32 €” and “87,86 €.” Subsequent dates—15. Juli, 10. Juli, 8. Juli (appearing twice), 7. Juli, 6. Juli (three times), 4. Juli (twice), and 29. Juni (twice)—repeat similar descriptions such as “CURSOR USAGE MID JUL NEW YORK” or “CURSOR USAGE JUN NEW YORK.” The amounts range from €82,88 to €87,77, except for the two entries on 29. Juni labeled “CURSOR, AI POWERED IDE NEW YORK” with amounts of €0,15 and €174,38. The right-facing arrow at the end of each row implies that details can be viewed by selecting the entry. The overall layout is tabular, with the text aligned left and amounts aligned right on a white background.

And I pay another couple hundreds for other AI products and agents.

Why? Because I'm soloing projects that consultancies like IBM or Infosys would staff with five people. These aren't vibe-coding side quests; they're enterprise-grade platforms on the cloud with proper backends, frontends, DBs, and AI services via Bedrock.

The result is a forecasted six-figure revenue and, more importantly, no burnout. I have enough time to take care of multiple projects.

You need to be a hands-on coder or have a deep technical understanding. Use ChatGPT, Grok, and Perplexity for your research beforehand.

This is my playbook.

General model guidance:

  • o3 for research and planning and Opus 4 for implementation is the best combo right now. It’s not even close. If the bill is too high, and with Opus 4 it is rising fast, switch to Sonnet 4.

  • Grok 4 is exceptionally intelligent (raw intelligence), but not pro-active enough. It wouldn’t try to solve a problem from different angles. It also doesn’t communicate well of what it does. Use it carefully.

  • Have your own library of reusable prompts (separated into your projects and a general part)

Be Surgical

End every key prompt with this: Be surgical, precise, thoughtful, and concise.

"Surgical" is the operative word. It forces targeted, minimal changes. To scale this, I keep a library of reusable prompts, split by project and a general folder.

When planning or implementing, i am always using it.

The Workflow for Existing Code

A developer-assistant interface focuses on a test script. At the top of the card, a grey bar shows the shell script name test-agent-setup-sse.sh and indicates it spans Lines 1–161. Just below, a highlighted tag, @test-agent-setup-sse.sh, is followed by a sentence from the user: “The test doesn't finish 100%. Why?” They instruct the assistant to analyze the codebase in depth and suggest an implementation plan to fix the error, stressing that the plan should be surgical, precise, thoughtful and concise. A “Restore checkpoint” button sits in the bottom right of the card. Under the conversation, a list logs the assistant’s internal actions: a line marked “Thought for 22s” indicates the thinking time, and subsequent lines with icons show what was searched or read. For example, it lists searches like “How does the agent setup process work, including SSE progress…” and “What are the steps in the agent setup process and how are the…” and file operations such as grepping setup-progress/stream and reading agents.ts. Each entry uses grey text and icons such as a magnifying glass or document. The interface uses muted colours and simple typography.

When tackling a bug or feature, the first move is a concise context dump. Give the agent everything relevant (but be concise): logs, docs, relevant files, error messages. Then, the directive:

Please analyze the codebase in depth and suggest an implementation plan to fix this error. Be surgical, precise, thoughtful, and concise.

This makes the AI think before it acts. You get a plan to approve, not a blind code dump.

Set the Stage, Every Time

This screenshot captures a conversational interface between a user and an assistant. The user’s message appears as a paragraph in a bordered card. They write that they want to remove a notification box labeled “Limited Access” — it explains that the user currently has access to six training paths and instructs them to request more paths through their manager — and ask that it be removed for all accounts locally before pushing code to CodeCommit. They add background information: code is pushed to AWS CodeCommit, mirrored to GitLab, built via a runner and deployed to Amazon ECS Fargate. They note that the assistant cannot interact with a remote terminal or git, so they want commands to run in an interactive dialog. A grey “Restore checkpoint” button sits at the lower right corner of the card. Beneath it, a faint status line reads “Thought for 4s,” indicating how long the assistant considered the request. The assistant’s reply follows in a separate card: “I’ll help you remove that ‘Limited Access’ notification box. Let me first search for where this text appears in your codebase.” The overall interface has a minimalist white-and-grey design with subtle borders and buttons.

A new thread means a blank slate. Always define the environment.

Background: I push to AWS CodeCommit, it mirrors to GitLab, and a runner deploys it to ECS Fargate.

Or set boundaries:

INFO: You can't use the terminal or git remote. Give me the commands to run. We'll do this interactively.

Hot Take: Don’t work in SCRUM or with a PM in general

In this new way of implementing with AI Agents, you develop best either alone or in a small, highly technical team. Nothing else is accepted.

  • I recently worked on the backend while a separate team handled the frontend, led by a non-technical scrum master/project manager. The project was a disaster: I had to create unnecessary Cucumber tests, scripts, and excessive documentation, which muddled the backend's clarity. The frontend made little progress, and I believe I could have built it faster alone.

Look at this horror (I needed 1 hour just to understand what is needed; image just shows top 40% of the ticket):

A documentation page outlines a user story titled “Core flow – Generate a phone number on test screen.” At the very top there are two small icon buttons — a plus symbol and a square with a diagonal cross — with a tooltip reading “Collapse” hovering nearby. Under the title, a section labeled “Description” contains a user story: as a user, I want to generate a phone number for my agent so that every inactive agent has a corresponding phone number. A scenario spells out that when the user provides a store location, they should receive a phone number linked to their agent. Below, a line labeled “Related API endpoint” links to the AIBPO API Documentation. The subsequent paragraph explains that when a location and agent ID arrive from the frontend, the backend should pull an unused number from a pool, link it to the agent, restrict it to 10 calls and respond with that number. A section titled “Follow the steps below” lists seven numbered backend tasks such as receiving the agent ID and location, pulling an unused number, validating that the agent is inactive, validating the location, marking the number as assigned, limiting it to 10 calls and returning it to the frontend. The last section, “Note to think about for the production,” questions whether the 10-call limit persists indefinitely or resets monthly and notes potential issues with agents switching between inactive and active states. A side panel on the right with the heading “Ready for…” has empty fields for Assigned to, Reporter, Labels, Due date, Sprint, Priority and Parent, and below it are collapsed sections labeled More, Automation, and Check…. The design is clean with black text on a white background and blue links.

I hope you enjoyed it.

Happy weekend!
Martin 🙇

Our webpage