De AI Prose

The rule

Readers can tell when a machine wrote the text, and it costs trust the same way sloppy code does. After you write any prose a person will read, read it back once and cut the tells. Plain words, varied sentences, say the thing instead of announcing you're about to say it.

Fires when

About to show a human a README, a doc page, a commit or PR body, a changelog entry, a UI string, store copy — anything a person will read as writing rather than run as code.

Tells to cut

• Signposting sentences: "Here's the thing.", "That's the point.", "This is the important part.", "By the end you'll know what, why, and how."
• Closers like "That's it." or "And that's all there is to it."
• Em-dash overuse as the default connector. The constant em-dash rhythm is a fingerprint.
• Parallel triads everywhere ("a texture, a config, a script"), over-balanced sentences, the list-of-three cadence.
• Arrow constructions in prose ("write one function -> the whole thing appears").
• A table for everything; bold sprinkled on every other phrase as decoration.
• Bullet lists where every item opens with a bold label ("Fast: ...", "Simple: ..."), parallel from top to bottom.
• Cutesy section names and peppy filler ("the payoff", "the ones that bite").
• Paragraphs that are all the same length and shape. Too tidy.

How to apply

Plain, varied sentences. Some short. Let one run long and a little uneven, like a person typing.

Say it directly instead of signposting it. Use em-dashes rarely. Prefer periods, commas, "and". Vary connectors. Cutting an em-dash means rebuilding the sentence (a comma, a colon, parentheses, or two sentences), not swapping a different mark into the same joint.

Keep bold for genuinely key terms, not decoration. Drop a table where a sentence reads more naturally.

Read it back and ask: "would a tired engineer writing this actually phrase it this way?" If it sounds like a polished AI explainer, flatten it.

Before / after

Before (AI voice):

Here's the thing about our caching layer — it's fast, it's simple, and it's reliable. The payoff? Sub-millisecond reads.

After (human):

The cache keeps reads under a millisecond. It's a plain in-memory map with a TTL, nothing clever.

What changed: dropped the signpost, dropped the triad, dropped the rhetorical question, said the concrete thing.

Red flags

This skill must obey its own rule. The read-back pass applies here too. If this file sounds like AI, it has failed on its face.