Stop giving your AI the same feedback

LAB 0001 · 16 September 2026

Pick one job where you keep correcting your AI. Give Codex the relevant threads and ask it to find repeated feedback.

Select the threads that contain the work and your corrections.

Ask Codex to quote the corrections and show their sources.

Compare better instructions in one thread with a specialist team.

Review the proposed rules, responsibilities and handoffs.

Approve the changes you want.

Try one real task. Check whether your old corrections still apply.

Lennox used separate workers for a job-search workflow: a manager, researcher, writer, designer and reviewer. He reports less review and coordination work. The reviewed application was not finished or submitted; no time-saving figure was measured.

Copy this

Read the feedback I have given you in these threads: [mention the threads].

Find the mistakes I have corrected more than once. Quote the relevant feedback and link each finding to its source. Separate lasting preferences from one-off requests. Tell me which threads you could not read. Do not guess what they contain.

Explain why I keep having to make these corrections. Compare:

Improving the instructions and review checks in one thread.

Creating a manager and separate specialist workers with clear handoffs.

Recommend whichever should require less review and coordination from me. More agents is not automatically better. Account for extra token use and handoff overhead; do not invent cost or time savings.

Show me the exact proposed instructions, who owns each part of the work, what each handoff contains, and how the result will be checked before it reaches me. Keep the useful feedback specific enough to prevent the same mistakes, without turning every one-off comment into a permanent rule.

Do not change files, create workers, start schedules or contact anyone yet. Present the proposal first. Ask only about decisions you cannot resolve from the threads.

After I approve the setup, test it on one real task. Compare the new result against my previous corrections, identify what still fails, and show me what changed.

Keep in mind

A new thread may not be able to read every older conversation. Ask it to report missing access instead of inventing context. A team uses tokens and needs coordination. Start with the smallest setup that helps. Old feedback can become stale or conflict with a newer request. Keep private documents and contact details out of material you share. An AI reviewer can still miss mistakes. Check the result on a real task.

Next step

Choose one recurring task and the threads where you corrected it. Run the prompt, then review the proposal.

Source links

OpenAI: releases teased: https://x.com/sama/status/2099872600977760451

Opus 5.2: one user’s report: https://x.com/pankajkumar_dev/status/2099892088368476668

Grok Bot: a three-day build: https://x.com/bot/status/2099890276315848743

Jev: structured decisions: https://x.com/CompleteSkeptic/status/2099925682726002904

Persona: a wearable assistant: https://x.com/zach_yadegari/status/2099894590723465295

Zuckerberg: safety and trust: https://x.com/finkd/status/2099997096896274533

Lenny’s AI Builders