Generative AI - Short & Sweet
FIELD NOTES / WORKING WITH GPT-6
GPT-6 Astra has launched, shortly after Claude Fable 5.1. Both are excellent so far. On the right tasks, they feel superhuman. But even that power needs limits.
I went to bed while GPT-6 kept working. By morning, I had paid €750.
The task still wasn’t done.
My laptop fan was still roaring. I had left Astra running on Codex’s Ultra setting, with automatic recharge on. I thought it was nearly finished.
It was an expensive way to learn that “keep going” needs a limit.
I had asked for a document assistant: find relevant passages, send them to the model, answer with sources. That pattern is called RAG.
I gave Astra plenty of requirements, left some choices open, and expected it to handle the rest. Instead, the project grew extra code and moving parts. I wanted a useful answer. I got a bigger project.
Since then, I’ve run more than 150 tests in my own work. These are the three habits I’d give you first.
01 / Define “done” before you start
State the result, what stays out, and the checks it must pass. Put known requirements in one brief. Let the model ask about gaps that change the result.
For my document assistant, “done” could mean correct answers with sources to five agreed questions, including one the documents cannot answer. That last check matters: a confident guess is a failed answer.
In partnership with
AI made PMs faster. Multiplayer mode is still broken.
A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.
Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.
And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.
AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.
02 / Put the limit outside the prompt
“Spend no more than €10” is a request, not a hard spending cap. Check automatic recharge and use enforceable runtime or usage limits where available. If you cannot enforce a limit, supervise the run.
Start with a short stage. Inspect it before extending it.
03 / Make extra effort earn its cost
Try Medium for routine work. Increase effort when a specific quality check fails. More reasoning can help hard tasks; it also uses time and tokens. The highest setting does not belong on every task.
A brief you can copy
Build a local assistant that answers from these PDFs and cites its sources. No login, dashboard or extra agents. First show the smallest design and five test questions. After my approval, build it, run those checks and stop. Report the result in five lines.
Add your actual limits in the tool before you run it.
Get the full field guide. Free.
Reply ASTRA and I’ll send you all 15 rules, the full brief and my cost-control checklist. If you like, add one task where AI went off track.
Reply here, or email [email protected] with “ASTRA”.
Until next time,
Martin
