Skip to content

Notes12 notesUpdated 7 Nov 2026

Notes on AI work that holds up.

Guides, analysis and an open build log from making MyPromptVerse.

New notes by email
All12Guides6Build log4Analysis2

Latest · Analysis

Amnesiac intelligence: AI is learning to sleep

A year ago I thought the hard part of AI memory would be training experience into the model. Then Anthropic got Claude through Mt. Moon by fixing a note.

A year ago I thought AI memory meant training. Then Anthropic got Claude through Mt. Moon by fixing a note.

Read the note 11 min read
New here? Start with three guidesShowHide
  1. Prompt, Playbook, Operator, Knowledge: what goes whereGuide · 7 min
  2. What the model actually sees: context engineering for reusable AI workGuide · 8 min
  3. Prompt testing: how to know if your prompt is goodGuide · 8 min

All notes, newest first

October 2026

  1. Guide7 minPrompt, Playbook, Operator, Knowledge: what goes whereThe four kinds of reusable AI setup, and when a prompt you keep reusing should become a Playbook.7 min
  2. Build log4 minThe Quick button did nothingQuick, Structured and Deep only changed the prompt's wording. How we found out and what replaced them.4 min
  3. Build log4 minWe had 10,000 lines of multi-agent code. None of it called a model.An audit found our planner, scheduler and worker queue all ran as a local simulation.4 min
  4. Build log4 minEvery run said 'Passed'. Nothing had been checked.A hard-coded line stamped real runs 'Passed'. How we found it and what the app checks now.4 min
  5. Build log4 minFive words, 5,325 tokens: how our app turned 'say hello' into a product specA five-word greeting took about 30 seconds and 5,325 tokens. Most of the cause was our own defaults.4 min

Newsletter

New notes, by email.

Each guide, analysis piece and build log entry as it is published, plus new Library records. Occasional, practical, easy to leave.

Occasional, practical, written by Naeem. Easy to leave. See privacy.

Ready for your email.