Claude and ChatGPT add unrequested promises when editing AI-like emails
A user tested how Claude (using Opus 5.5) and ChatGPT (using GPT-6) handle AI-generated emails with specific phrasing and numbers. Both models altered hedging language like ‘is expected’ to definitive statements and inserted unrequested promises (e.g., ‘I will contact you again’ or ‘I will confirm separately’), even when instructed to avoid changes. Claude’s skill feature—custom rulesets for…
Key points
- Claude’s skill feature preserved hedging terms (*‘is expected’* → unchanged) but moved promises to suggestions; ChatGPT’s skill still altered phrasing despite rules
- Both models added unrequested promises (e.g., *‘I will confirm’* or *‘reduces errors’* instead of *‘is expected to’* without skill constraints
- Prompt-based edits (for Gemini, Claude, ChatGPT) kept original phrasing intact, while skills showed mixed compliance with explicit rules
The test email included fictional company details, conditional deadlines, and hedged claims (e.g., ‘unconfirmed’ payroll software compatibility). Numbers (¥198,000 initial cost, 30-person cap) remained unchanged across all edits. The author notes that ‘skills’ (ZIP-formatted rule files) require manual upload via platforms like ChatGPT’s skills screen, with no confirmation on pricing plan compatibility. A paid guide (¥980) offers the core skill, prompt templates, and implementation steps.
Model pages: Opus 5.5 → · GPT-6 →
The story so far
3 episodes →- Claude and ChatGPT add unrequested promises when editing AI-like emailsthis story
When You Ask AI to Fix AI-like Emails, It Adds Unrequested Promises | Finishing Skills That Don't Change the Meaning (Comparing Claude and ChatGPT)
note.com · 12 October 2026
Loading the full article…
This text was published by note.com and written by 会社のAI係｜コピペで使う業務スキル. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Marketing & Small Business
All →- Character.AI files for Chapter 11 bankruptcy and agrees to sell assets to Disney for $20 million · 20 src
- Open-source tools replace expensive data science software for Free · 1 src
- Non-coding executive shares Claude Code memory and handoff system for executives · 1 src
- PCMag shares 6 ways to improve Google AI overviews · 1 src
- Claude AI offers four work courses for $19.99 during Deal Days · 1 src
Comments
via GitHub Discussions