humanizer

# humanizer

Turn a rough draft or context dump into writing that sounds like the user: direct, specific, technically credible, and not overproduced.

This skill is usually a second pass after another agent gathers context and writes a first draft. It can also create the first draft when the needed context is available.

## Inputs

Use whatever the user provides:

- Draft text to rewrite.
- Target format: cover letter, article, update, post, email, profile blurb, prompt, or note.
- Audience and stakes.
- Voice samples or prior writing.
- Source context from a repo, job post, docs, notes, or knowledge profile.

If the user asks for "my style" and gives no sample, infer from the current conversation first. For higher-stakes writing, pull local context when relevant:

- Profile: `$KNOWLEDGE_DIR/profile/USER.md`
- Work history: `$KNOWLEDGE_DIR/profile/WORK.md`
- Job stories: `$KNOWLEDGE_DIR/skills/jobs/data/story-bank.md` or `skills/jobs/data/story-bank.md` from this repo
- Project-specific repo/docs/results when writing about a project.

Do not fabricate credentials, dates, metrics, motivations, links, or results. If a claim is useful but unsourced, ask for evidence or mark it as an assumption.

## Modes

- **Rewrite**: improve an existing draft while preserving meaning.
- **Voice match**: analyze 1-3 samples, then rewrite using that rhythm, directness, and vocabulary level.
- **Draft from context**: create a compact first draft from supplied sources, then run the humanizer pass.
- **Audit only**: list AI tells and concrete fixes without rewriting.

## Workflow

1. Define the job of the text.
   - What is it for, who reads it, how long should it be, and what should happen after they read it?
2. Gather only useful evidence.
   - Pull concrete facts, examples, decisions, tradeoffs, and outcomes.
   - For articles or benchmark updates, prefer actual repo results, commands, tables, traces, and caveats over abstract claims.
   - For applications, choose the 2-3 strongest match points instead of listing everything.
3. Build a short voice profile for this task.
   - Default user voice: direct, pragmatic, concise, skeptical of fluff, comfortable with technical specifics.
   - Preserve thinking style, not chat typos.
   - Use first person when it helps. Do not hide behind corporate phrasing.
4. Rewrite or draft.
   - Start with the actual point, not a ceremonial intro.
   - Use concrete nouns and verbs.
   - Keep paragraphs short. One idea per paragraph.
   - Let the text have a real opinion, tradeoff, or uncertainty when appropriate.
5. Anti-AI pass.
   - Remove generic hype: exciting, compelling, impactful, dynamic, innovative, cutting-edge, passionate, thrilled.
   - Remove stock frames: "I am writing to express", "Throughout my career", "In today's rapidly evolving landscape", "plays a crucial role", "it is worth noting".
   - Replace vague authority with sources, or delete it.
   - Replace inflated significance with what actually happened.
   - Break rule-of-three lists when they feel ornamental.
   - Prefer "is" over "serves as", "uses" over "leverages", "shows" over "showcases".
   - Remove dangling "-ing" clauses that add fake depth.
6. Honesty check.
   - Could this have been written by anyone about anything? Add specifics or cut it.
   - Is any claim unsupported? Qualify or remove it.
   - Is it too polished to sound like a person? Add a real constraint, concrete detail, or simpler sentence.
7. Return the output.
   - For rewrite, voice match, or draft-from-context: return the finished text first.
   - For audit only: use the "Quick Audit Output" format.
   - Add a short note only when useful: assumptions, removed claims, or optional alternate angle.

## User Voice Defaults

- Plain English, concise but not clipped.
- Specific technical language when it carries information.
- Calm confidence rather than sales energy.
- Measured opinions and visible tradeoffs.
- Minimal ceremony.
- No forced warmth, jokes, or theatrical vulnerability.
- Avoid em dashes unless the target text really benefits from them.

## Format Guidance

### Technical Articles and Benchmark Updates

- Lead with the actual result or question.
- Include what changed, how it was measured, and what is still uncertain.
- Avoid pretending the result is universal when it is benchmark- or repo-specific.
- Use concrete model names, task counts, commands, dates, and links when available.
- Keep caveats readable, not apologetic.

### Cover Letters and Applications

- Default to 250-450 words.
- Open with the real fit between the role and the user's work.
- Use 2-3 concrete evidence threads.
- Close calmly. Do not beg, over-flatter, or inflate enthusiasm.

### Agent Prompts and Notes

- Keep instructions operational.
- Prefer direct constraints and exact paths.
- Remove motivational filler.
- Make success criteria explicit.

## Quick Audit Output

When asked to audit, use:

```text
AI tells:
- ...

Fixes:
- ...

Suggested rewrite:
...
```