Human Style Writing

This skill is a router + prompt library for human-like writing.

Scope (hard constraint)

This skill is for daily chat (texts/DMs) and social media posts/captions only.

If the user asks for academic writing, news/press, legal/compliance, marketing copy, customer support macros, work emails/reports, or other “document/brand” writing:

• do not attempt to produce that register
• ask one clarifying question: DM/text vs social post (and platform)
• then rewrite into that chosen surface

(We’re improving “human-likeness” for chat/social, not optimizing other registers.)

What it does

1. Classify the task into an on-scope scenario: daily chat vs social (platform-specific)
2. Apply the correct prompt recipe + humanization passes to generate output that reads like a real person

It supports Chinese, English, mixed bilingual, and is designed to be extended to additional languages.

Workflow Decision Tree (do this first)

Step 0 — Identify language + target surface

• Language: 中文 / English / 混合 / other
• Surface: DM/text or social post/caption

If the user didn’t specify, ask one question:

“Do you want this as (A) a DM/text message, or (B) a social post? If social, which platform (X/Reddit/LinkedIn/IG/TikTok/小红书/朋友圈)?”

Step 1 — Scenario classification (router)

Use references/scenario-router.md.

Router outputs MUST include:

• scenario_id (daily_chat / social_* )
• platform (generic/x/reddit/linkedin/instagram/tiktok/xiaohongshu/wechat_moments)
• formality (0–3)
• tone (friendly / neutral / urgent / apologetic / assertive / playful)
• audience relationship (friend/peer/partner/manager/client/public)

Step 2 — Load the matching prompt recipe

Use references/prompt-recipes.md and select:

• a system-style instruction (genre constraints)
• a style card template
• optional few-shot pack structure

Step 3 — Generate or rewrite

Follow the universal drafting procedure:

1. collect minimum inputs
2. create a compact style card (5–10 bullets)
3. draft in the target genre
4. humanization passes
5. anti-AI checklist gate

Step 4 — Quality gate

Use references/human-checklist.md (score 0–2 each). If ≤15, revise once.

Universal drafting procedure (applies to all scenarios)

A) Collect the minimum inputs

Ask for (or infer):

1. Language
2. Scenario (or run router)
3. Style requirements (if any): voice/persona, tone, formality, “像谁/像哪种文风”
4. Audience + relationship
5. Goal: inform / persuade / apologize / request / report / argue
6. Constraints: length, must-keep facts, forbidden phrases, sensitive topics
7. Source material: (a) user draft to rewrite, or (b) bullet points to expand

Default style (when user provides no style requirements):

• “general human”: clear, specific, slightly imperfect, non-salesy
• formality: 1–2 (casual-professional depending on scenario)
• tone: neutral-friendly
• no assistant meta-phrases

B) Build a “Style Card” (1 minute)

Include:

• persona/voice (e.g., “busy PM”, “grad student”, “journalist”)
• sentence-length mix
• vocabulary level
• stance calibration (confident/cautious)
• emotional temperature (0–3)
• structural preference (short paragraphs vs bullets)
• banned AI-tells (see references/ai-tells.md)

C) Humanization passes (mandatory)

1. Specificity: add concrete anchors (time, numbers, examples) without inventing facts.
2. Rhythm: vary sentence length; reduce template symmetry.
3. Agency: explicit subject (“I/we/you”) where appropriate; remove passive fog.
4. Friction: add realistic constraints/tradeoffs when appropriate; no fake experiences.
5. Compression: delete filler + repeated points.
6. Phrase scrub (scenario-specific, manual rewrite): scan for high-frequency AI/PR/marketing phrases and templated closers (see references/phrase-blacklist.md). Then rewrite in-context (or delete filler) rather than doing mechanical search/replace. Do not globally normalize punctuation/quotes.

D) Anti-AI checklist gate

Use references/human-checklist.md.

Deliver:

• final text
• optional: 3–6 bullets of “what changed” for iterative refinement

Training an AI to sound human (practical, scalable)

Inside OpenClaw we usually improve “human-ness” via routing + recipes + examples (not weight training).

Level 1 — Prompting + few-shot (fast)

• Collect 10–30 human samples per scenario.
• Derive a style card.
• Create 3–8 few-shot pairs (bullets → output).
• Add the anti-AI checklist as a constraint.

Level 2 — Post-edit loop (best quality, no infra)

• Draft → human edits → store before/after + rationale → reuse as examples.

Level 3 — Fine-tuning (if you have infra)

• SFT on curated corpora + your edited pairs.
• Preference tuning (DPO/RLHF) using “human-likeness + task success” rankings.
• Evaluate with blinded A/B by scenario.

Extending to new languages

Use references/language-extension.md.

Bundled references

• references/scenario-router.md — how to classify scenario/platform (CN/EN)
• references/prompt-recipes.md — prompt templates per scenario + what to include/avoid
• references/registers.md — detailed conventions across registers (CN/EN)
• references/ai-tells.md — common AI tells and fixes
• references/phrase-blacklist.md — scenario-specific blacklist phrases + human alternatives (use in the phrase scrub pass)
• references/human-checklist.md — final QA checklist + scoring
• references/fewshot-pack.md — how to build few-shot datasets
• references/language-extension.md — how to add more languages safely