# Three nights of work (/claude-and-chatgpt-2/case-2-nishas-first-client/three-nights-of-work) --- type: Document title: "Three nights of work" description: "What the AI Worker did while Nisha slept, what she checked each morning, a run that stopped, a second opinion from ChatGPT, and a missing lease contract." status: stable order: 0.63 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-11-10T15:02:33Z chapter: "F00D" part: Front expert_status: required concepts: [ "While she sleeps" ] last_verified: 2026-11-09 sources: - id: memory title: "Anthropic, Let Claude coordinate ongoing work with Projects" resource: https://code.claude.com/docs/en/claude-projects - id: new-threads title: "Anthropic, Let Claude coordinate ongoing work with Projects" resource: https://code.claude.com/docs/en/claude-projects - id: limit title: "Anthropic, Let Claude coordinate ongoing work with Projects" resource: https://code.claude.com/docs/en/claude-projects - id: run-sheet title: "Nisha's First Client: practical run sheet (the practical has not yet been run)" resource: "appendix-practical-run-sheet.md" generated: at: 2026-11-10T15:02:33Z by: process:panaversity trust_tier: unverified build_id: sha256:46fd53c5d7ab1c14f6347908d45c9fdaecda4d914439923ee878bdcecd57803e dirty: true ksor_version: 0.0.60 --- > The run log, the catches and the times in this part of the story are **illustrative** until the practical is recorded. See [About this case](overview.md). ![One night in the AI Worker's life. The subtitle says the desk forgets, the Library remembers, and a stopped run waits, then continues. Two rows of boxes. Night 1, Tuesday to Wednesday: at 9 p.m. Nisha closes her laptop. At 1:00 a.m. the scheduled task starts a run from its written instruction. Threads work at their desks from the Library. Bridge version 1 is saved to the Library. The desk is cleared but the Library still holds version 1. At 6:00 a.m. Nisha checks on her phone. Night 2 and Thursday: at 1:00 a.m. the run starts as planned. The usage limit is hit, and the thread stops and waits (a dashed box, for stopped). In the morning there is no new version and version 1 is still there. The limit resets and the thread continues on its own. At lunch Nisha finds bridge version 2 in the Library. A band says every run reads the project instructions, project memory and the Library files, and web lookups reach only allowlisted sites. Gold boxes are Nisha's steps and white boxes are AI work. A callout says a run knows only its instruction and what was saved outside the desk. A note says the times are illustrative until the practical is recorded.](img/one-night-in-the-ai-workers-life.png) ## Wednesday morning Nisha woke at six and opened the project on her phone, on the **Library** tab. The first run had finished in the night ([Exhibit 3](exhibits.md#exhibit-3-the-nightly-run-log)). A new file was in the Library: the bridge, version 1, with the date. She read it with her coffee. The bridge started with Sutlej's IFRS profit before tax, Rs 120.0 million, and ended with a US GAAP figure of Rs 103.2 million ([Exhibit 4](exhibits.md#exhibit-4-the-us-gaap-bridge)). A line under it said the figure would be Rs 102.0 million if the warehouse lease turned out to be a different kind of lease. The inventory thread had found the reversed write-down. The development thread had found the costs recorded as an asset this year, and the Rs 3.0 million charged this year for costs from earlier years. The lease thread had worked out the difference in lease cost. The reviewer thread had left a note. It said two things in the files did not agree. The engagement letter gave the warehouse rent as Rs 12.0 million a year, but Sutlej's lease note said Rs 10.0 million. One of them was wrong. And the total in Sutlej's inventory note was Rs 2.0 million higher than the inventory in the trial balance. The bridge used the trial balance figure. The worker had found both problems, but it could not say which figures were right. The answers were with Sutlej, and only Nisha could ask. She wrote both points in her notes, to raise with Sutlej. It was 6:25. She had to leave for work at seven. ## Wednesday evening That evening she did three things. First, she wrote to Sutlej's accountant. She asked which rent figure was right, why the two inventory figures differed, and for a copy of the warehouse lease contract. The lease note gave the interest rate, 8%, but not what happened at the end of the lease. Under US GAAP that decides whether the lease is an operating lease or a finance lease. The lease thread had assumed an operating lease and listed this as an assumption, as the brief asked. The answer was in the contract. Second, she asked for a second opinion. As in Case 1, she gave the bridge to ChatGPT and asked for a strict review, without saying who had prepared it. ChatGPT agreed with the arithmetic ([Exhibit 5](exhibits.md#exhibit-5-who-caught-what)). But its list of problems was long. It had seen only the bridge, not Sutlej's files, so it also listed what it could not check. One of its main points was about the lease. Nobody had read the contract, so the kind of lease was not settled. Until it was, Rs 103.2 million was one possible answer, not a final figure. Third, she filtered the points before passing them on, as she had in Case 1. Many of the points needed Sutlej's documents, so they had to wait. She kept two and added them to the project instructions, because instructions reach every new thread exactly as she writes them. One said to show the two lease answers side by side and to call neither one the result. The other said to explain the accounting words in the lease line in plain English for the board. Then she wrote a short progress note to Mark herself. It listed the three areas, the lines she had checked so far and the questions open with Sutlej. She read it twice and sent it from her own email. She used **project memory** for something smaller. She told the coordinator: *"Remember: check rent against the lease note, not the engagement letter."* Claude saved the note to project memory, and every cloud thread reads it when it starts. But memory is notes Claude writes in its own words, and Anthropic calls such notes "instructions Claude keeps to, not enforced settings".[^memory] ## Thursday morning On Thursday morning, the Library had no new version. The run had started at one, as planned. It had stopped partway through with a limit error. Her plan's usage limit had been reached, and a thread that stops at the plan's limit waits. By default, it continues on its own when the limit resets.[^limit] Nothing had been lost. Version 1 was still in the Library. But nothing new had been made either. At lunch, after the limit had reset, Nisha opened the **Library** tab again. Version 2 was there. The thread had continued on its own and finished. Version 2 showed the two lease answers side by side and explained the lease line's accounting words in plain English, the two points she had added to the instructions on Wednesday. That mattered. Changes to instructions reach new threads, not threads already running, and a run in a project can continue an earlier thread.[^new-threads] Wednesday night's run had started a new thread, so it had read the new instructions. No amount had changed. The bridge showed Rs 103.2 million as an operating lease and Rs 102.0 million as a finance lease ([Exhibit 4](exhibits.md#exhibit-4-the-us-gaap-bridge)). She checked the inventory and development lines against Sutlej's notes and the rules, one by one. They held. The lease line she could not check yet. Only the contract could say which kind of lease it was, and Sutlej had not sent it. Next, [Thursday night](thursday-night.md). [^memory]: As verified on November 9, 2026, Anthropic says project memory is "Notes Claude keeps about the project ... stored as files", that "Every cloud thread reads the index file MEMORY.md when it starts", and that preferences Claude saves this way are "instructions Claude keeps to, not enforced settings". Let Claude coordinate ongoing work with Projects, Anthropic. [^new-threads]: As verified on November 9, 2026, Anthropic says "Changes to instructions, repositories, plugins, and environment in Project settings reach new threads, not threads already running." Let Claude coordinate ongoing work with Projects, Anthropic. [^limit]: As verified on November 10, 2026, Anthropic says "A thread that stops at your plan's limit waits and by default continues on its own when the limit resets", that to keep it from continuing you "turn off Auto-continue when usage limits reset in Project settings > General", and that, with the setting off, you "send the thread or conversation a message after the reset" to continue it. Let Claude coordinate ongoing work with Projects, Anthropic.