Lessons Learned (Living Log)

We keep a running log of what worked, what didn’t, and what changed as a result.

Entries

Short, plain sections produce better reader‑side AI summaries
Keep paragraphs tight
Use explicit headings
State reading ease level by school style boundary 'high-school senior'
Context is limited, summarising a single short document gives better results than a a massive PDF.
- Limit below 200 lines of text (perhaps 4000 words max) for best chance of usable output

Reduce waffle in main flow.
Keep main text focused
Link to context/citations
- OpenAI markdown parser added
- Ensure research reports are exported in json, then converted to fully annotated markdown
Citations from 'deep research' tools often keep the point, but lose the connection to the underlying source.
- Labour report has mostly reliable claims, but is missing some citations, to be resolved

Exploration in drafts/chats, clarity in print
ChatGPT and Claude respond well to direct instruction, as it encourages more careful output, and in the case of GPT-5, to engage thinking mode, where it produces much better quality outputs.
This is most likely because GPT-5 marshalls requests to older models, which have vastly different capabilities.
Research from ChatGPT can only be considered 'mostly' reliable if it comes from 'thinking' or 'deep research' modes.
Most models are great with markdown as input and output
Server load is a MASSIVE influence on LLM response quality. Do NOT use a service under heavy load, or at heavy load times. The results at peak demand times are unusable and cannot be trusted. This applies to all platforms.
AI responses have a tendency to rabbit-hole just to have something to say, so you can chase your tail forever if you want. Try to focus on the objective, small issues matter, but eyes-on can tell you what really matters more than ChatGPT or Grok can.