← sources

How do you stop your agent's emails reading like an agent wrote them?

source: reddit (reddit:AI_Agents)published 2026-11-07 10:50ingested 2026-11-08 02:35

retained text

Genuine question first, then what I did to it.

Token prompting only brought me so far. Drafts still had the usual tells and the humanisers I tried on top switched a number or name, a bad trade for an email to a client.

What I made is reright, a rewrite step the agent calls with its draft. It returns the text sounding like you, after checking numbers, names, links and code match the draft, the language the same and length close. If a check fails the agent gets its own words back. For extra safety you can switch on "Check before it's used" and approve every rewrite yourself. Those approvals and edits are also how it learns your style.

It plugs into Claude Code (MCP plus a hook) and into Claude and ChatGPT as a connector.

On my own run over public AI Humanizer Benchmark texts, which the benchmark team didn't review, it switched a number, name or quote in 1% of texts. The other 9 tools did that 30% to 46% of the time. Detectors aren't its thing, it came third of 10.

There's a free plan with 3,000 words a month. I'd really like to hear other approaches though, tool or no tool. Link in the comments.

document fetched 2026-11-08 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xq3lzj/how_do_you_stop_your_agents_emails_reading_like