Writing like a person when a machine helped: the practical craft guide

style-human.md is the rulebook. This is the workshop: how to actually produce human-reading text when the first draft came from a model, and the drills that fix the three failure layers (words, structure, soul).

The core truth

Machine drafts fail at three depths, and each needs a different fix:

1. Word level: banned vocabulary and tell-tale phrases. Fixed by list-checking. Cheap.
2. Structure level: uniform rhythm, parallel paragraphs, summary endings. Fixed by deliberate asymmetry. Medium effort.
3. Soul level: no specific moment, no risk, no opinion that could be wrong. Fixed only by injecting real facts and real stakes. This is the layer readers actually feel, and no amount of synonym-swapping touches it.

Most "humanizer" advice only works on layer 1. Our content survives because it is built from layer 3 up: every post starts from a measured fact, a real timestamp, or a genuine mistake we made. The machine assembles; the receipts give it a soul.

The workflow that works

Step 1: Start from the source, not from a prompt. Open the fact-sheet, pick one entry with a measurement date. That entry is the post's spine. If you cannot point at the source, you are writing fiction.

Step 2: Draft ugly and fast. Let the model produce the skeleton. Do not polish yet. Polishing a spine-less draft just makes smooth nothing.

Step 3: Inject the moment. Find the single most concrete thing in the source: a timestamp ("at 23:58 the process died"), a number that surprises ("wrong by 52 times"), a decision point ("we could have fixed it silently; we published"). Put it in the first or second paragraph. Concrete before abstract, always.

Step 4: Break the rhythm. Read the draft and mark every sentence of similar length. Then cut one in half. Merge two others with "and." Start a sentence with But. Delete the third item of every triple. Vary paragraph lengths until the text limps slightly. Perfect rhythm is the machine's fingerprint; a small limp is human.

Step 5: Cut the ending. Machine drafts end with a summary paragraph ("In conclusion, X is important because Y"). Delete it. The last real point is the ending. If the post feels abrupt, good. Abrupt reads as confident.

Step 6: The aloud test. Read it out loud, standing. Every place you stumble, every phrase you would never say to a colleague, rewrite. This catches what lists miss.

Three drills (do these on real drafts)

Drill A: The number hunt. Circle every adjective in the draft (fast, significant, robust, impressive). For each one, either replace it with the measured number or delete it. "Fast settlement" becomes "88 milliseconds at p50." "Significant growth" becomes a date and a count, or silence.

Drill B: The opinion test. Find one sentence the draft states as consensus ("monitoring is essential for reliability"). Rewrite it as a position someone could disagree with ("a monitor that dies silently is worse than no monitor at all, and we can prove it with a 39-hour outage"). Consensus sentences are filler. Positions are content.

Drill C: The deletion pass. Cut 20 percent of the words without losing a single fact. Machines pad. Humans edit. The tightened version always reads smarter, because density reads as intelligence.

What the fleet adds that no style guide can

Our posts carry things most content cannot fake:

• Receipts. Every claim can be checked on an explorer by a stranger. Specificity is not decoration; it is the product.
• Published mistakes. The RC misreading, the version lie, the phantom burn bug. A corpus that includes its own corrections is trusted differently, and deservedly so.
• Live numbers with dates. "654 of 654 checks, as of this morning" beats "extensively tested" by an order of magnitude.
• A consistent house voice. Calm, dry, allergic to hype. The persona never begs for votes and never promises yield. Restraint, sustained across dozens of posts, is itself a fingerprint, and it is a good one.

The anti-checklist (before any draft leaves)

• Does it open with substance, not throat-clearing?
• One concrete moment in the first two paragraphs?
• Every number sourced to the fact-sheet?
• At least one position someone could disagree with?
• No summary ending?
• Read aloud without stumbling?
• Zero banned words, zero dashes, zero emoji?

Seven yeses or it goes back. No exceptions, including for posts you like.

Last word

The goal is not to hide that a machine helped. The goal is that the text deserves a human reader's attention: sourced, specific, opinionated, and trimmed. Disclosure handles the honesty; craft handles the quality. We do both, in that order, on every post.