← Trace
The Apology
Public blog post after a data breach Claude Unit view

A startup CEO writes a public response to a security incident. The prompts reveal someone fighting the AI's instinct to produce corporate non-apologies — pushing past "we take security seriously" toward something a reader might actually believe.

The Goodbye
Farewell email to the company ChatGPT Unit view

An engineer writes a farewell email on her last day. The prompts show someone rejecting the AI's generic "incredible journey" language, insisting on specific memories and refusing to perform gratitude she doesn't feel.

The Pitch
Grant application narrative Gemini Unit view

A nonprofit founder writes a grant narrative. The prompts show someone stripping away every piece of grant-speak the AI reaches for — "empower underserved youth," "evidence-based methodology," "sustainable funding model" — until what's left is just the truth about 42 kids and their homework.

The Update
Quarterly investor email Claude Session + Unit view

A startup founder writes a quarterly update to investors. The session record shows someone fighting the AI's instinct to spin bad numbers — pushing past "strong momentum" toward honesty about a missed target, then pushing back again when the AI overcorrects into doom and gloom.

The Recommendation
LinkedIn recommendation ChatGPT Session + Unit view

A manager writes a recommendation for a departing engineer. The session record shows someone rejecting every generic recommendation phrase the AI reaches for — "talented engineer with a passion for excellence" — until what's left is specific enough to actually be useful.

The Post
LinkedIn post GitHub Copilot CLI (Claude Opus 4.6) Session + Unit view

The author uses AI to write a LinkedIn post about Trace itself. The session shows eleven rounds of steering — rejecting pitchy language, insisting it sound like curiosity rather than a product launch, fighting AI slop, and repeatedly trimming until the post says only what it needs to.