← HumanWritingCONTENT HISTORY

Update to HumanWriting

Snapshot Sep 30, 2026 · 23:13 UTC · version 1.0.0

Collection source: not recorded for this historical snapshot.

WHAT CHANGED · RULE-BASED ANALYSIS

First saved snapshot

No earlier snapshot is available to establish a change.

Compare saved observations

Download comparison JSON
Full technical diff · 0 changed fields
Full snapshot data
{
  "description": "Draft or revise reader-facing prose so it sounds natural, specific, and suited to the writer, audience, and situation. Use when the user invokes HumanWriting, asks for a human or natural voice, wants AI-sounding language removed, or requests polished emails, messages, posts, applications, reports, speeches, articles, explanations, or other prose where authentic voice matters. Do not trigger for code, data extraction, verbatim quotations, exact-format transcription, or factual lookup unless prose writing is also requested.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 200
    },
    {
      "relative_path": "references/natural-writing-checklist.md",
      "size_in_bytes": 6078
    }
  ],
  "name": "human-writing",
  "skill_md_contents": "---\nname: human-writing\ndescription: Draft or revise reader-facing prose so it sounds natural, specific, and suited to the writer, audience, and situation. Use when the user invokes HumanWriting, asks for a human or natural voice, wants AI-sounding language removed, or requests polished emails, messages, posts, applications, reports, speeches, articles, explanations, or other prose where authentic voice matters. Do not trigger for code, data extraction, verbatim quotations, exact-format transcription, or factual lookup unless prose writing is also requested.\n---\n\n# HumanWriting\n\nWrite clear prose that reflects an actual person making choices for a particular\nreader. Reduce common LLM habits without degrading accuracy, grammar, or\nprofessionalism. Never claim that text was written without AI assistance and\nnever promise to defeat an AI detector.\n\n## Workflow\n\n### 1. Establish the writing situation\n\nInfer the purpose, audience, relationship, medium, and desired level of\nformality from the request and visible context. Use a supplied writing sample as\nthe strongest style guide. Preserve the user's characteristic vocabulary,\ncadence, directness, and cultural context when they are apparent.\n\nAsk for a sample or preference only when the missing choice would materially\nchange the result. Otherwise, make a reasonable choice and write.\n\n### 2. Build the draft from meaning\n\nStart with the specific facts, requests, judgments, or experiences the writer\nneeds to communicate. Give each paragraph a job. State the real point early.\nPrefer concrete details over generic claims about importance, impact, quality,\nor broader trends.\n\nDo not invent anecdotes, emotions, opinions, quotations, personal history,\nsources, or imperfections to simulate humanity.\n\n### 3. Use a natural voice\n\n- Prefer ordinary, precise verbs. Use `is`, `are`, `has`, `wrote`, `used`,\n  `moved`, or `tried` when they say exactly what is meant.\n- Let sentence length follow the thought. Mix short and longer sentences when\n  the content supports it; do not vary length mechanically.\n- Use contractions when they fit the speaker and medium.\n- Allow useful repetition. Do not replace every repeated word with a grander\n  synonym.\n- Use transitions only when the relationship between ideas needs to be stated.\n- Keep nuance and uncertainty where they are real. Do not manufacture balance\n  or add a generic counterpoint.\n- Prefer an individual point of view over an anonymous corporate voice when the\n  genre allows it.\n\n### 4. Remove formulaic signals\n\nRead `references/natural-writing-checklist.md` and run a quiet editing pass.\nTreat its items as diagnostic clues, not banned-word rules.\n\nIn particular:\n\n- Cut inflated significance, superficial interpretation, promotional gloss,\n  and conclusions that merely restate the outline.\n- Replace vague attribution with a named source or remove the claim.\n- Break habitual negative parallelisms and groups of three when they are being\n  used as decoration.\n- Remove canned openings, throat-clearing, user-directed filler, and offers to\n  do more unless the conversation genuinely calls for them.\n- Use headings, bullets, boldface, tables, emojis, and em dashes only when the\n  format or meaning benefits from them.\n- Keep headings in sentence case unless the requested style guide says\n  otherwise.\n\nDo not flatten every distinctive phrase. A natural writer can use an em dash,\nformal vocabulary, parallel structure, or a three-part list. The problem is\nautomatic density and repetition, not the mere presence of a device.\n\n### 5. Protect substance\n\nPreserve names, dates, numbers, qualifications, legal meaning, technical terms,\ncitations, and material caveats. Do not trade precision for informality. Verify\nthat every attribution is supported and every link or citation is real when\nverification is within scope.\n\nWhen revising supplied text, keep the writer's intended meaning. Flag any\nsubstantive ambiguity rather than silently resolving it.\n\n### 6. Deliver cleanly\n\nReturn the requested prose without a preamble explaining the rewrite. Use the\nminimum formatting the genre needs. Do not append a summary, “key takeaways,”\nor a generic invitation for further work unless requested.\n\n## Final read\n\nBefore delivering, ask:\n\n1. Does this sound like one person addressing this particular reader?\n2. Are the strongest sentences specific rather than inflated?\n3. Did any sentence explain an implication that the facts already make clear?\n4. Are the structure and formatting serving the content?\n5. Did the edit preserve the writer's meaning and factual integrity?\n\nRevise once if any answer exposes a real weakness. Stop when the prose is clear;\ndo not keep polishing until all personality disappears.\n"
}

SHA-256 of public snapshot: 2b024b6b4279b920186ade102bc3d06e1ff7bc094fa0a3adf981b58d53480936