Codex tallied 180 hours on a product from git and chats, and helped clear up tax loose ends
The author of the "LLM under the hood" channel (LLM pod kapotom) asked the Codex agent to estimate his labor from git history, work chats and the log of his conversations with Codex. The result: 180 hours on his latest product and 20 hours on a training module. The same approach with a "company brain" — a shared base of chats, email and bank exports — helped him find and pay outstanding taxes. If an agent has traces of your work, you can hand it the estimates and the routine.
AI-processed from TG: @llm_under_hood (LLM под капотом); edited by Hamidun News
How many hours did you actually spend on your last product? Most people answer from memory, and get it wrong. The author of the Telegram channel "LLM under the hood" (LLM pod kapotom) stopped guessing and gave the job to the Codex agent.
It went through the git commit history, work chats and the log of conversations with Codex itself, and came up with about 180 hours of development on the latest product and about 20 hours on preparing a training module. This is an example of an "AI-native" workflow, where AI doesn't just answer questions but works with the real traces of your work. The author uses "company brain" to mean a single base that pulls together chats, email and bank exports.
The post is personal and has no benchmarks, but its techniques are easy to repeat.
Estimating time from traces of work
The first technique is simple. When he needs to know how much time a project took, the author doesn't try to remember — he asks Codex to calculate it. For that, the agent needs access to three sources: the git history of all projects, work chats, and the history of work with Codex itself. The author exports the latter into a separate conversation index, meaning an index of dialogues with the agent.
The author makes one important caveat himself: the 180 and 20 hours are only his personal time. The agent can't yet calculate the team's total hours, because it doesn't have the same access to the traces left by other people.
The practical value is clear. After-the-fact estimates are needed for planning, pricing and conversations with investors, and memory is an unreliable source for them. Commits and correspondence at least provide a reproducible basis. Still, this is a model's estimate, not a timesheet: time spent thinking away from the keyboard doesn't show up in git.
The "company brain" and accounting loose ends
The second technique is heavier. To sort out accounting records and open issues, the author connects all work chats (or simply all of them), the email archive and bank exports to the company brain. He then asks the agent to untangle the problem the way an experienced Silicon Valley process manager would.
For local data, the author built a separate bridge service that runs on a Mac. It does three things:
- gives read-only access to the local Telegram and email client databases, so the agent can't change or delete anything;
- keeps a separate cache in a search-friendly form;
- serves data through an interface with progressive disclosure: short summaries first, details on request.
The last point matters for token economics. Dumping all correspondence and exports into the context at once is expensive and lowers answer quality. When the agent first sees a table of contents and opens only the pieces it needs, it spends fewer tokens and is less likely to get lost in irrelevant material.
According to the author, this let him pinpoint the outstanding taxes, pay them quickly, reconcile the results and neatly sort documents into folders. Along the way, the agent also helped file a bug report with the bank about a failure in its statement export system.
Overnight ideas and limitations
The third technique is more about mindset. When you look at the processes of more mature companies through the lens of a company brain, you see how much you could improve in your own. Ideas immediately come up about which data sources to connect and where to embed them.
The author names a favorite prompt: at night he asks a model he calls Fable 5.2 to act as an experienced YC partner hired for a lot of money to accelerate the company. The task: review the changes in the company brain over the past few days, do research, and propose one small thing that would improve scaling, redistribute responsibility, or make life easier. The "one small thing" constraint is useful in itself: you get a step that can actually be carried out, rather than a long report nobody reads.
Things to consider if you want to repeat this:
- The post contains no quality measurements: it's unknown how many mistakes the agent made with the taxes or how they were checked. Financial conclusions are best verified by hand.
- Access to chats, email and the bank is a serious privacy risk. Read-only access and a local bridge reduce it but don't eliminate it. Employees and counterparties whose messages end up in the base should be told.
- Counting time from git and chats works well for one person. For a team, you need shared sources and the participants' consent.
- The model name and service details come from the author's own account; there is no independent verification.
The practical takeaway: an agent's value grows with the quality of the data it can access. If you collect your work history in one place, give read-only access safely, and feed the data in portions, the same model can serve as an analyst, an accountant's assistant and a night-time process consultant. The author ended with a question to readers about which techniques they use. It's a good prompt to check which data sources in your company aren't yet connected to an agent.
- Nov 21, 2026Amazon SageMaker HyperPod Now Saves Live LLM Requests to S3 at Three Levels
- Nov 21, 2026OpenAI Disbands Robotics Team Again: GPT-6 Astra Completed 74.8% of Tasks in Simulation Versus 51%
- Nov 21, 2026AI Warehouse Clerk Broke a Hidden Rule: DeepSeek V4 Pro Failed in 18% of Cases, Gemini 3.8 Flash in 4.8%
- Nov 21, 2026AWS gives agents a shared LLM cache skill: up to 40% cheaper, first token 80% faster
Key Questions Answered
What are the key details on Estimating time from traces of work?
The first technique is simple. When he needs to know how much time a project took, the author doesn't try to remember — he asks Codex to calculate it. For that, the agent needs access to three sources: the git history of all projects, work chats, and the history of work with Codex itself. The author exports the latter into a separate conversation index, meaning an index of dialogues with the agent. The author makes one important caveat himself: the 180 and 20 hours are only his personal time.
What are the key details on The "company brain" and accounting loose ends?
The second technique is heavier. To sort out accounting records and open issues, the author connects all work chats (or simply all of them), the email archive and bank exports to the company brain. He then asks the agent to untangle the problem the way an experienced Silicon Valley process manager would. For local data, the author built a separate bridge service that runs on a Mac.
What are the key details on Overnight ideas and limitations?
The third technique is more about mindset. When you look at the processes of more mature companies through the lens of a company brain, you see how much you could improve in your own. Ideas immediately come up about which data sources to connect and where to embed them. The author names a favorite prompt: at night he asks a model he calls Fable 5.2 to act as an experienced YC partner hired for a lot of money to accelerate the company.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.