persona-write

Files (47)

• SKILL.md
• ai-patterns.md · dictionaries/
• banned-phrases.md · dictionaries/
• manager-speak.md · dictionaries/
• anti-ai-guidelines.md · docs/
• contributor-guide.md · docs/
• longform-mode.md · docs/
• output-handling.md · docs/
• persona-chain-mode.md · docs/
• persona-theory.md · docs/
• philosophy.md · docs/
• pipeline.md · docs/
• simple-usage.md · docs/
• voice-guide.md · docs/
• writing-quality-rubric.md · docs/
• audit.md · examples/
• longform.md · examples/
• quickstart.md · examples/
• rewrite.md · examples/
• style-sheet.md · examples/
• chapter-memory.md · longform/
• chapter-workflow.md · longform/
• consistency-pass.md · longform/
• document-map.md · longform/
• global-brief.md · longform/
• revision-tickets.md · longform/
• cold-reader.md · passes/
• diagnostic-audit.md · passes/
• fidelity-check.md · passes/
• intent-extraction.md · passes/
• native-fluency.md · passes/
• persona-mapping.md · passes/
• proofread.md · passes/
• refine.md · passes/
• rewrite.md · passes/
• sentence-build.md · passes/
• style-sheet.md · passes/
• voice-coherence.md · passes/
• blunt-operator.md · personas/
• clear-teacher.md · personas/
• pragmatic-builder.md · personas/
• problem-first-marketer.md · personas/
• quiet-witness.md · personas/
• sharp-technical.md · personas/
• skeptical-analyst.md · personas/
• proofread.py · scripts/
• test_proofread.py · scripts/

SKILL.md (first lines)

# Persona Write

## Purpose

Use this skill to write through a **specific kind of persona**, not generic assistant voice.

This skill is for:
- drafting text
- rewriting existing text
- auditing weak or AI-ish text
- refining already revised text
- processing long documents section by section

The goal is writing shaped by a specific attention model, compression level, and stance.

## Core rules

- **Positive shape over negative shape.** Voice work means executing the persona's positive signature, not avoiding a global ban-list. See `docs/philosophy.md`.
- Persona comes first.
- Preserve meaning before style.
- Default to concise, clear writing.
- Do not flatten all writing into one neutral assistant voice.
- Do not one-shot long-form rewrites.
- For long text, process one major section or chapter at a time.
- Use Markdown as the working format when writing to files.
- Do not optimize for detector evasion.
- Do not add fake human quirks like random typos or forced slang.

## Scratch folder & Pipeline traceability

Scratch files are **durable memory**, not virtuous traceability. The mode is decided by length and task type, not by user request.

### Two modes

| Target output | Mode | Why |
|---|---|---|
| **≤ 600 words** (roughly 1–2 pages) | **In-memory** | The whole pipeline state fits in working memory. Serialization adds friction and loses cohesion. |
| **> 600 words** | **Scratch** | The state exceeds one working window. Materializing each pass keeps the persona stable across the longer arc. |
| **Long-form** (multi-section, > 2000 words) | **Scratch + chapter workflow** | Chapter memory and revision tickets do durable work. See `longform/`. |
| **`/persona-copy` analysis** | **Scratch** | Extracted persona files and verification reports are saved artifacts by definition. |

Estimate target length from the user's brief. If the user explicitly asks for the analysis to be saved on a short task, switch to scratch on request.

### In-memory mode

Run all passes in-context. Return the result inline with a one or two sentence note on the main adjustments. No folder is created.

`sentence-build` still applies in this mode: keep the skeleton, the sentence log, and the style sheet in context, build one sentence at a time, and show all three as compact lists under the result so the build can be checked.

### Scratch mode

1. **Create a task folder:** `scratch/YYYY-MM-DD-[task-slug]/`
2. **Record each pass to its own numbered `.md` file:** `01-intent.md`, `02-audit.md`, `03-mapping.md`, `04a-skeleton.md`, `04b-sentences.md`, `04-draft.md`, `05-coherence.md`, `06-refine.md`, `06b-native.md`, `06c-wordcheck.md`, `07-fidelity.md`, `07b-cold-reader.md`, `08-proofread.md`, `final.md` (and `00-style-sheet.md`, plus `00-research.md` when `/persona-research` ran).
3. The scratch folder is gitignored.

The reasoning, with cross-model experimental evidence, is in `docs/voice-guide.md`. The underlying principle is in `docs/philosophy.md`.

## When to use this skill

Use this skill when the user wants to:
- remove generic assistant patterns from existing text
- rewrite text in a stronger voice
- sound more direct, technical, practical, or clear
- improve text that is over-hedged, padded, or voiceless
- write something from scratch in a defined voice
- revise a large document while keeping voice and structure consistent

## When not to use this skill

Do not use this skill when the user only wants:
- grammar correction with minimal rewriting
- exact preservation of wording
- detector evasion
- a full large-document rewrite in one pass

## Multi-persona chain mode

When the user provides more than one persona, interpret them as an ordered chain.

Accepted formats:
- `sharp-technical -> problem-first-marketer`
- `sharp-technical, problem-first-marketer`
- natural language: "write as a sharp technical person, then review as a problem-first marketer"

### Chain semantics

- **First persona owns the text.** It writes or rewrites the draft.
- **Every following persona is a reviewer.** It does not become a co-author.
- **Reviewers run a two-step process** — no user permission required:
  1. Identify problems from their perspective (what is off, unclear, misframed, wrong for the audience)
  2. Apply surgical fixes — sentence-level, paragraph-level, local framing, targeted cuts or additions
- **Reviewers must not** rewrite the whole piece in their own voice, replace the owner persona, or flatten the text into generic prose.
- **After all reviewers finish, the owner persona runs a mandatory final reconciliation pass.** This pass smooths seams, restores the owner's voice where it was displaced, preserves useful reviewer improvements, and ensures the final text reads as one coherent piece.
- **Final output should still primarily sound like the first persona.**

For multiple reviewers, apply them in order. Each reviewer reads what the previous one produced.

Example — three-persona chain:
1. `sharp-technical` writes
2. `clear-teacher` reviews and applies surgical fixes
3. `problem-first-marketer` reviews and applies surgical fixes
4. `sharp-technical` runs the final reconciliation pass

→ Full details: `docs/persona-chain-mode.md`

## Main user flow

### If the user already gave a persona
Proceed directly.

### If the user did not give a persona
Ask for one before writing.

List the preset options in plain English:

- **sharp-technical**: direct, concise, technical
- **pragmatic-builder**: practical, clear, useful
- **clear-teacher**: easy to follow, explanatory without fluff
- **skeptical-analyst**: careful, critical, evidence-aware

From DurdeuVlad/persona-write, MIT.

More from DurdeuVlad/persona-write