When AI writes your prompt, make it mark its guesses.
What the Reel says
When AI writes your prompt, make it mark its guesses.
Tired of redoing your Friday client update?
So I had GPT-5.5 write the prompt.
9
email parts, its call
Reused weekly, one wrong pick hits every Friday.
Then I added: put guesses in [brackets].
It flagged 7. One: 300–600 words, too long.
Fix its picks once, then reuse every week.
Sources and assumptions
- Claude Platform Docs: Prompting best practices (section 'Be clear and direct'): 'Claude responds well to clear, explicit instructions. Being specific about the output you want can help enhance results.' and the 'Key test: Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too.' Backs the fix step: replace the guessed details with your real ones. (checked 2026-11-04)
- Claude blog: Get to production faster with the upgraded Anthropic Console (Mar 6, 2025): Background only: Anthropic's Console has a prompt generator ('Describe what you want to achieve, and Claude will use prompt engineering techniques ... to create an effective, precise, and reliable prompt'). We did NOT test the Console, and the old prompt-generator docs page now redirects to the best-practices page, so the Reel must not describe or name the Console tool. (checked 2026-11-04)
Assumptions:
- All tests ran on GPT-5.5 through the OpenAI API (label 'GPT-5.5 (OpenAI API), Nov 2026'), not the ChatGPT app. The planned Claude runs failed in this research session (the CLI wasn't allowed to run) and are recorded with errors. Never say Claude produced these outputs.
- Client 'Rosa' and the website-project notes are fictional.
- Word count of the fixed output was counted by hand: about 75 words, counting the subject line and headings.
- '7 guesses' is the number of bracketed items in the flagged draft's 'Guesses I made' list. '9 parts' is the number of bullet points under 'Please include' in the plain draft.
- The plain draft does say 'Do not invent details that are not in my notes', so the Reel must not claim it would make up facts. Its guesses are about format: tone, sections and length (it set no length).


