A Single Claude Code Agent Writes an Article in 7.8–8.8 Minutes; a Chain of Roles Takes 10.5–18.9
An author on Habr built an "article factory" from a chain of AI roles, but the documentation grew larger than the code, so he kept a single Claude Code agent. Across three technical briefs, the agent wrote text in 7.8–8.8 minutes, while the chain took 10.5–18.9. These are machine run times only: the author did not measure setup and repairs, so the savings are unknown. Count the costs of accepted results, not the speed of a single run.
AI-processed from Habr AI; edited by Hamidun News
Automation with AI almost always looks like a win at the moment of the first run: a result appears within minutes. The question is how many hours went into it before and after. The author of a Habr analysis built an "article factory": several AI roles in sequence turned a technical brief (TZ, a description of what text is needed) into a finished article.
Then he replaced that chain with a single Claude Code agent. Across three briefs, the agent finished in 7.8–8.
8 minutes, while the chain of stages took 10.5–18.9 minutes.
But that is only the machine's working time. The author did not count how much human time went into setup, checks, and repairs, so it is impossible to say whether he came out ahead or behind on time. Still, his three cases clearly show where the hidden work hides.
Article factory: when documentation outgrows the code
The first version ran on Python and SQLite with several sequential roles. The instructions and reference materials came in packs of 176–222 thousand characters, and most training runs failed the checks. Instead of an article, the author got output he had to untangle. When the documentation grew larger than the code, he froze the system.
He rebuilt the task around a single Claude Code agent: a short prompt and five skills, meaning sets of instructions for specific tasks. Code checked the result, and a hook fired when the agent finished its response, preventing it from stopping early. Anthropic describes hooks as commands that run automatically on a chosen event in Claude Code.
The ranges 7.8–8.8 and 10.5–18.9 minutes do not overlap, but human time is not included in them. There is also a token nuance: the agent read 5–8 million tokens from cache. A cache is a temporary store on the service side that holds already processed text so it does not have to be processed again. This is the volume read, not human time. A short prompt does not mean the entire context became small.
The overnight run that failed before it started
After a Claude Code update, the path to the program changed, and the scheduled overnight run failed on every line. This was a failure of the launch itself, not of text quality: there was no content to check. The author did not count the recovery time.
For schedules, he suggests a checking order (this is a recommendation, not a fix already applied):
- the command in the schedule actually launches Claude Code;
- the same command runs one trial task under conditions close to the overnight ones;
- a launch error is recorded separately from a result-check error.
The separation is useful: "did not launch" and "launched, but the result failed the check" are fixed in different ways.
A listings digest with last year's events
A family events digest delivered events from 2025. When a post had no date, the model filled in the current week, and the result was a recommendation for an event that had already passed. After that, the author began checking dates in code. For a digest, two dates matter: when the post was published and when the event takes place. If the event date cannot be established, it is better to hand the entry to a person than to ask the model to guess.
The limitations are honest: the author did not count how many issues have gone out without errors since the fix, so the digest's reliability cannot be assessed. The code verifies dates but does not confirm that the text describes the right event.
How to count savings and when to stop
The author proposes comparing manual work and agent-assisted work over the same period with the same requirements for the result. Four items go into the accounting: preparation, execution, review, and corrections. What needs counting is the cost of accepted results, including failed attempts. Rejected text and a digest with a past event do not count as finished results.
Two more rules apply. Waiting for the agent and human occupancy are different measures: if Claude Code works for eight minutes while you are busy with something else, those are not eight minutes of your work. If you watch the run and intervene, that time is recorded separately. And errors discovered after acceptance also belong to the result. The author cites the essay Counting the Bandages by Solomon Shalom Lijo: without counting the errors that slipped through, weak checking looks like acceleration.
The decision is simple. Continue if automation delivers accepted results and, over the period, takes less human time, including repairs; a rare failure is no reason to switch it off. Simplify if maintenance gets in the way of results: this is how the author removed the chain of roles, and the digest can be limited to collecting events. Switch off if, after simplifying, there is still a lot of manual work or errors cannot be reliably filtered out.
The practical takeaway: a fast launch does not prove savings. Automate what you can verify with code, and keep records of costs for at least several weeks.
- Nov 22, 2026DeepSeek Veteran Open-Sources Linger: A Local AI Agent That Responds Up to 5x Faster
- Nov 22, 2026Anthropic Shares Fell 8.5% in a Day, Nearly Back to the $83 IPO Price
- Nov 22, 2026CCRent Resells Claude to Chinese Vibe Coders: $100 Plan Costs $200
- Nov 22, 2026Linger Runs an AI Agent on a 16 GB Mac With No Token Fees and Promises a 5x Speedup
Key Questions Answered
What are the key details on Article factory: when documentation outgrows the code?
The first version ran on Python and SQLite with several sequential roles. The instructions and reference materials came in packs of 176–222 thousand characters, and most training runs failed the checks. Instead of an article, the author got output he had to untangle. When the documentation grew larger than the code, he froze the system. He rebuilt the task around a single Claude Code agent: a short prompt and five skills, meaning sets of instructions for specific tasks.
What are the key details on The overnight run that failed before it started?
After a Claude Code update, the path to the program changed, and the scheduled overnight run failed on every line. This was a failure of the launch itself, not of text quality: there was no content to check. The author did not count the recovery time.
What are the key details on A listings digest with last year's events?
A family events digest delivered events from 2025. When a post had no date, the model filled in the current week, and the result was a recommendation for an event that had already passed. After that, the author began checking dates in code. For a digest, two dates matter: when the post was published and when the event takes place. If the event date cannot be established, it is better to hand the entry to a person than to ask the model to guess.
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.