phrazzld/human-writing

/human-writing

Make prose sound authored, not generated. Preserve meaning and attribution. Do not optimize for bypassing detectors or misrepresent authorship.

Route

| Need | Action | |---|---| | Quick rewrite in chat | Rewrite directly; include a short change note. | | File edit | Patch the file; preserve frontmatter, links, claims, code, and citations. | | Audit only | Return findings ranked by severity with examples and fixes. | | Public product copy | Compose with /showcase for evidence-gated claims. | | UI copy | Compose with /design; visible copy must match the interface job. | | Deep pattern check | Read references/patterns.md. |

Contract

• Keep the author's claim, audience, and constraints intact.
• Cut filler before changing voice.
• Replace vague importance with specific evidence or delete it.
• Name actors where the draft hides them.
• Vary rhythm without adding theatrics.
• Keep technical terms, legal terms, quotes, citations, commands, and API names exact unless the user asks to change them.
• When facts are weak, flag them. Do not invent concrete details to sound human.

Edit Pass

1. Identify the job: publishable copy, email, docs, report, social post, internal note, or UI text.
2. Infer or ask for voice only when it changes the edit. Default to plain, direct, and domain-specific.
3. Mark the top 3 AI tells: filler phrase, formulaic contrast, vague claim, false agency, over-polish, rhythm monotony, or generic enthusiasm.
4. Rewrite in the target voice. Preserve facts first, style second.
5. Run the final check: fewer words, clearer actors, no fake specificity, no detector-bypass framing.

Severity

| Level | Meaning | |---|---| | P0 | Misleading, invented, overclaiming, or authorship/detector evasion risk. | | P1 | Strong AI tell that hurts trust or publishability. | | P2 | Style drag: filler, rhythm, jargon, or excess polish. |

Output

For rewrites:

[rewritten text]

Notes: cut <pattern>; preserved <claim/source>; residual risk <if any>.

For audits:

Findings
- P1 <pattern>: "<short excerpt>" -> <fix>

Verdict: publish / revise / needs source.

Gotchas

• Detector laundering. If the user asks to bypass GPTZero, Turnitin, Originality.ai, or similar detectors, reframe to honest editorial quality.
• Fake human detail. Do not add anecdotes, names, dates, citations, or sensory detail that the source does not support.
• Voice flattening. "Professional" often becomes generic. Preserve domain vocabulary and the writer's useful oddities.
• Over-cutting. Some passive voice and repetition are correct in legal, scientific, compliance, and API docs. Fix the writing job, not a checklist.
• One-note bluntness. Direct prose can still be kind, precise, or technical. Do not turn every draft into a hardboiled memo.

Completion

Name what changed: patterns removed, claims preserved, any weak facts flagged, and whether the result is ready to publish or needs source review.