My Claude Code and Codex Workflow: The Problems I Had to Solve Myself

AigoPilot founder built AigoPilot, a studio and mobile app for managing AI coding sessions across projects, after running roughly 100 Claude Code sessions and hitting problems locating the right conversation and keeping agents from interfering with his own desktop use. The studio adds a voice assistant that opens a session directly in a few seconds instead of returning a session ID, and a remote-machine workflow that moves a project and its sessions to a spare laptop, including environment preparation such as installing Unity for a Unity game project. The account describes the fixes as self-built responses to context reconstruction and shared-desktop conflicts that also arise when automation shares one interactive environment, including a browser.

I have sent a prompt to the wrong project. I was working across five projects, with a separate VS Code window for each and many AI coding sessions inside them. The windows looked similar. Finding the right conversation was already taking more effort than I expected, and sometimes I got it wrong. Then desktop testing introduced a different problem. An agent needed to launch the application, inspect its interface, move the mouse, and type. While it did that, using my own computer could interfere with the test. I have had days when this kept me from managing other work on the same machine. After more than a year of intensive AI-assisted development, these were the problems occupying me: where the work was happening, what it needed from me, and how to keep it moving without repeatedly reconstructing its context. Prefer listening? Here’s an AI-generated audio discussion based on this article. The fixes gradually became AigoPilot https://aigopilot.com , the studio I now develop and use. I am its founder. I can select a project, see its sessions, and check which ones are working or waiting in one place. That helps even when I use a single AI provider. My Claude Code workflow, together with Codex, changed as I worked through the situations below. Even in the correct window, I could spend time looking through tabs to find the conversation I needed. My Claude Code work grew to around a hundred sessions. Some were finished; others were waiting for a decision or another instruction. A bug made this particularly frustrating. I wanted the session that had built the affected feature, because it might still have the relevant design decisions in context. A fresh session had to investigate the code and reconstruct that background first. Searching titles rarely worked for me. A conversation could begin with one task and accumulate many others. Its title was a poor record of everything inside it. I could ask an agent to search the session history. Sometimes that took two or three minutes and substantial reading, then returned a session ID. I still had to work out how to reopen it, sometimes through a terminal instead of the interface I was using. So I added a voice assistant to the studio and mobile app. I can describe the project and the feature I am looking for. In my use, finding it usually takes a few seconds, and the assistant opens the session directly. If there are several candidates, I get a short description of each and choose. I use the same assistant to find and open an email. Both requests end at the item I need to work on, instead of leaving me with another identifier to look up. Two of my projects are desktop applications: AigoPilot and Cilevi Screen Recorder. Testing their interfaces introduced a problem I had not felt as strongly in my background development tasks. After changing an application, an agent may need to launch it, inspect the screen, click controls, and type. When that happens on my main desktop, my own input can interfere with the test. I have had days when I could barely use my computer because an agent was working through the application. Managing a second project was difficult. Running another UI test made things worse: one application would come to the foreground while another agent was trying to interact with something else. This was about the desktop available to the agents. Similar conflicts can happen whenever automation shares the same interactive environment, including a browser. I started looking at the computers I already had. A spare laptop could take a project and its sessions, leaving my main workstation free. A friend also had an unused laptop at home. Connecting to a remote machine over the internet made that useful too. Moving files was only part of the work. One project was a Unity game, and the destination laptop did not have Unity or all the prerequisites installed. That led me to add environment preparation to the transfer workflow. I have worked through Unity in this setup. I have not automated installation for every application someone might use. Other dependencies may still need an AI-assisted or manual installation. But a project that can run and test on another machine leaves me able to use my own computer while it works. I wrote more about testing desktop apps with AI on spare computers https://medium.com/@ohadi.anoush/my-ai-coding-agent-needed-the-mouse-i-needed-my-computer-b4fefe10dd4d , including the environment setup problem I encountered when moving a Unity project. My AI Coding Agent Needed the Mouse. I Needed My Computer. https://medium.com/@ohadi.anoush/my-ai-coding-agent-needed-the-mouse-i-needed-my-computer-b4fefe10dd4d Usage limits kept interrupting me. Waiting, buying additional access, or switching subscriptions all had a cost, so I spent time investigating token consumption. Some experiments ran for three or four days, comparing an approach across projects with and without a proposed technique. A few methods made consumption worse. Reading a persuasive article was no substitute for trying it on my own work. One experiment began with a tiny request, such as generating a Python Hello World program. I wanted to understand why the input could be so much larger than the sentence I had typed. The request could carry a substantial amount of surrounding information. That made skill loading worth examining: which instructions needed to be present, and which could wait until the task required them? I worked on keeping short skill descriptions available and loading the detailed material when needed. In one of my comparisons, the input-token count fell by about 60%. That figure describes that comparison, not every prompt or a guaranteed reduction in a subscription bill. Selective loading is also documented practice today. Anthropic’s skill authoring guidance https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices describes loading skill metadata first and reading the instructions and references as needed. My experience concerns the setup I examined and the overhead I removed; it would be inaccurate to say every coding tool always sends every installed skill in full. Prepared file-processing tools helped too. For tasks that need extracted text or tables, the studio can convert PDFs, Word documents, and spreadsheets into Markdown before handing the content to the model, instead of repeatedly asking an agent to develop a converter. The model still has to read the relevant content. A task involving a chart, formula, or visual layout may also require the original file. And token counts need interpretation. Cached input can be accounted for differently https://platform.claude.com/docs/en/build-with-claude/prompt-caching from newly processed input. I keep the size of a request separate from what it ultimately costs. As a conversation grew, I also needed a way to continue with less accumulated context. A handoff summary helped, but the next session could lose a detail that mattered immediately. On one occasion, an earlier session had worked with an API key. The new one did not know where it was stored. I had to return to the old conversation, find the location, and relay it. A short path can matter more to the next task than pages of discussion. My handoff process now pays attention to operational details: relevant files, URLs, completed work, decisions, and the next step. For credentials, the useful record is the storage location and access method; copying secrets into more summaries creates its own problems. I also removed the manual copying. From the prompt box, I can choose to continue in a new session, select the provider, model, and effort, and have the handoff prepared and passed to that session. The studio can warn me when a conversation has become heavy so I remember to consider moving on. I wrote a practical guide to starting a new AI coding session without losing the details https://medium.com/@ohadi.anoush/starting-a-new-ai-coding-session-without-losing-the-details-760216e10781 , with a handoff template for file paths, verification status, access locations, and the next action. Starting a New AI Coding Session Without Losing the Details https://medium.com/@ohadi.anoush/starting-a-new-ai-coding-session-without-losing-the-details-760216e10781 Switching providers introduced another gap in the setup I was using. My Codex work used AGENTS.md; my Claude Code sessions used CLAUDE.md. I added a small reference between the project instructions so I did not have to maintain separate copies. Current Claude Code documentation also describes direct AGENTS.md support in supported versions and configurations https://code.claude.com/docs/en/memory agentsmd . A shared-file import also remains an option. The practical lesson is to check what a session actually loads, rather than infer it from the tool’s name. I wrote a practical guide to using CLAUDE.md and AGENTS.md in the same project https://medium.com/@ohadi.anoush/claude-md-agents-md-same-project-f618c8fe9054 , including the optional import and a check for what each session loads. That gives a new session access to the recorded project guidance. It does not transfer every thought from the previous conversation. A useful handoff still has to say what is happening now. I also wrote about how I’m designing project memory for AI coding sessions https://medium.com/@ohadi.anoush/my-ai-coding-sessions-had-the-code-they-were-missing-the-decisions-9cdebd252698 to preserve decisions, technical evidence, and the conditions under which they apply. My AI Coding Sessions Had the Code. They Were Missing the Decisions. https://medium.com/@ohadi.anoush/my-ai-coding-sessions-had-the-code-they-were-missing-the-decisions-9cdebd252698 In the mobile workflows I tried, opening a session sometimes showed me its previous state while the latest messages loaded. I have waited five to ten seconds, or hit a Retry message and eventually returned to my computer. I built the mobile side of the studio around seeing the latest state quickly and acting on it. I can create a project on a connected machine, choose its drive and folder, and start sessions from my phone. Notifications needed to explain why I should open the app. A session requesting permission needs attention. A finished task may only need a glance. A result that ends by asking whether to perform another step is waiting on me again. The notifications now identify the project and session and take me straight there. Customer support messages and requests for a human chat operator get the same treatment. Then there was the interruption I could miss entirely. I would leave home expecting a task to finish, return an hour later, and discover it had stopped shortly after I left because of a usage limit. Even after the allowance became available again, the session could remain idle until I intervened. I added continuation after the limit resets so it can resume without waiting for me to return. The waiting period still exists; the avoidable delay afterward is what I wanted to remove. I wrote more about managing Claude Code and Codex from my phone https://medium.com/@ohadi.anoush/managing-claude-code-and-codex-from-my-phone-what-i-actually-needed-02da5e805c2d , including the notifications and small decisions that help work continue while I’m away from my desk. Managing Claude Code and Codex from My Phone: What I Actually Needed https://medium.com/@ohadi.anoush/managing-claude-code-and-codex-from-my-phone-what-i-actually-needed-02da5e805c2d I kept recognizing another habit in myself and developer friends: finish a product, then start building the next idea. Finding customers and handling their questions required work we were less comfortable doing. Content production became its own long experiment. Over roughly a year, I tried skills, supplied design references, examined weak results, and refined instructions for images and video. I then turned those accumulated rules and workflows into reusable skills for the studio’s content plugin. A new conversation could start with those instructions instead of repeating all the earlier trial and error. The underlying model had not permanently learned my preferences; I had made the working knowledge explicit. Keyword research was another example. While preparing this article, I initially considered “AI coding workflow” as its main phrase. The Keyword Planner research for this article reported 390 average monthly searches for that term and 8,100 for “claude code workflow,” worldwide, in English, over October 2025 through September 2026. The latter’s September figure was 4,400, so the annual average also needed context. “Codex vs Claude Code” was larger, at 49,500, but people searching it expect a comparison. My story would not give them one. That is a practical use of the research workflow behind my Search Planner plugin. It helped choose wording that matched both the subject and measured demand. It did not promise a ranking. Sales and support added more reasons to keep the work connected. Purchases from connected Stripe or Paddle accounts, prioritized incoming email, and customer conversations now come into the same environment. An AI chat assistant can use a knowledge base prepared from available product information, with a path to a human when needed. Those needs explain why I kept extending the studio beyond coding. Many of these problems appear with a single AI subscription. Several projects and long conversations are enough to create them. Switching providers is useful in some situations, while finding the right session and keeping a desktop available matter throughout the day. I still have to decide what to build, review changes, and respond when an agent needs judgment. I want to spend less time locating the conversation that needs that judgment, or waiting to use a computer occupied by a test. The spare-machine setup is a particularly concrete improvement for me: the application can be tested on another computer while my own mouse and keyboard remain available for the next project. I’m Anoosh Ohadi, a software developer and the founder of AigoPilot. I write about the practical work of building and running software with AI. My Claude Code and Codex Workflow: The Problems I Had to Solve Myself https://pub.towardsai.net/my-claude-code-and-codex-workflow-the-problems-i-had-to-solve-myself-8cab221d8a51 was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.