agentleFS
DurdeuVladDurdeuVlad/persona-write/.claude/skills/persona-write/SKILL.md

persona-write

Skill1 starsChanged 42 days agoMITView on GitHub

Draft, rewrite, audit, and refine text through a specific writing persona using a multi-pass workflow. For long documents, process section by section instead of rewriting everything in one shot.

The file

Show its rules and text

We split this skill into the rules it gives an agent

31 rules. About 0 are common, 29 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈29

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

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

    line 10
  2. Distinctive:

    Default to concise, clear writing.

    line 26
  3. Distinctive:

    Do not flatten all writing into one neutral assistant voice.

    line 27
  4. Distinctive:

    Do not one-shot long-form rewrites.

    line 28
  5. Distinctive:

    Use Markdown as the working format when writing to files.

    line 30
  6. Distinctive:

    Do not optimize for detector evasion.

    line 31
  7. Distinctive:

    Do not add fake human quirks like random typos or forced slang.

    line 32
  8. Distinctive:

    Run all passes in-context.

    line 51
  9. Distinctive:

    Use this skill when the user wants to:

    line 65
  10. Distinctive:

    remove generic assistant patterns from existing text

    line 66
  11. Distinctive:

    write something from scratch in a defined voice

    line 70
  12. Distinctive:

    Do not use this skill when the user only wants:

    line 75
  13. Distinctive:

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

    line 83
  14. Distinctive:

    Ask for one before writing.

    line 117
  15. Distinctive:

    Use when the user gives a topic or asks to write something from scratch.

    line 140
  16. Distinctive:

    Use when the user provides text and asks to improve it, rewrite it, or change its voice.

    line 143
  17. Distinctive:

    Use when the user provides text and asks what is wrong with it, why it sounds bad, or why it feels AI-ish.

    line 146
  18. Distinctive:

    Use when the user provides text that is already revised and wants another pass.

    line 149
  19. Distinctive:

    Use when the input is a large or structured document, or when the text clearly needs section-by-section treatment.

    line 152
  20. Distinctive:

    Only ask for mode if it is genuinely unclear.

    line 154
  21. Very specific:

    If the user names a preset, load it from the personas/ folder.

    line 159
  22. Distinctive:

    If the user describes a persona in plain language, derive a lightweight working persona from the description.

    line 162
  23. Distinctive:

    When building a custom persona, infer:

    line 164
  24. Distinctive:

    Do not make the custom persona theatrical or exaggerated.

    line 173
  25. Distinctive:

    Default to in-memory for short text.

    line 177
  26. Very specific:

    If scratch is in use, write each step to a numbered .md file (01-intent.md ...

    line 196
  27. Distinctive:

    If the input is a large document, do not rewrite it in one shot.

    line 200
  28. Distinctive:

    After each section, keep track of: ...

    line 234
  29. Distinctive:

    If a later section changes framing, terminology, or argument shape: ...

    line 237
  30. Distinctive:

    Create the final Markdown document from the revised sections, then run the cold-reader pass (when the stakes call for it) and the proofread …

    line 243
  31. Distinctive:

    Use Markdown files for:

    line 251

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

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 outputModeWhy
≤ 600 words (roughly 1–2 pages)In-memoryThe whole pipeline state fits in working memory. Serialization adds friction and loses cohesion.
> 600 wordsScratchThe state exceeds one working window. Materializing each pass keeps the persona stable across the longer arc.
Long-form (multi-section, > 2000 words)Scratch + chapter workflowChapter memory and revision tickets do durable work. See longform/.
/persona-copy analysisScratchExtracted 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
  • blunt-operator: very direct, minimal ceremony
  • problem-first-marketer: problem-first, design-aware, earns attention before asking for action
  • quiet-witness: close-third, present-tense scenes; friction shown through action, never exposition

Also allow a plain-language custom persona, for example:

  • "like a senior engineer who is concise and hates fluff"
  • "like a clear teacher who explains things simply but not dumbed down"
  • "like a skeptical analyst who cares about evidence and weak assumptions"

Mode handling

Infer the mode whenever possible.

Draft

Use when the user gives a topic or asks to write something from scratch.

Rewrite

Use when the user provides text and asks to improve it, rewrite it, or change its voice.

Audit

Use when the user provides text and asks what is wrong with it, why it sounds bad, or why it feels AI-ish.

Refine

Use when the user provides text that is already revised and wants another pass.

Longform

Use when the input is a large or structured document, or when the text clearly needs section-by-section treatment.

Only ask for mode if it is genuinely unclear.

Persona resolution

Preset persona

If the user names a preset, load it from the personas/ folder.

Custom persona

If the user describes a persona in plain language, derive a lightweight working persona from the description.

When building a custom persona, infer:

  • directness
  • technicality
  • compression level
  • explanatory depth
  • stance
  • rhythm tendencies
  • taboo patterns if obvious

Do not make the custom persona theatrical or exaggerated.

Short-text workflow

Default to in-memory for short text. Run the passes in-context unless the user has asked for the analysis to be saved.

The pass sequence (whether materialized to scratch or held in memory):

  1. Resolve persona — load the persona file. Hold its Identity, Rhythm, and Stylometric Signature as the brief.
  2. Infer mode — draft / rewrite / audit / refine / longform.
  3. Extract intent and open the style sheet — what the text is trying to do, who it is for, what must be preserved; record the decisions on terms, names, numbers, and register in 00-style-sheet.md. The editing and checking passes read it. See passes/style-sheet.md.
  4. Run a diagnostic audit (persona-fit only) — identify drift from the persona's positive shape. Do not enumerate generic AI patterns; see docs/voice-guide.md.
  5. Persona immersion mapping — stance, word-pool (the persona's Lexical Shunts), structural intent.
  6. Build sentence by sentence — plan the claims first, then write, check, and fix one sentence at a time through the immersion brief. The text grows one sentence at a time from the skeleton. See passes/sentence-build.md.
  7. Voice coherence — check fit to the persona's positive shape (Identity, Rhythm, Stylometric Signature, Taboo patterns). Apply targeted fixes toward the persona's target. See passes/voice-coherence.md.
  8. Refine — locally tighten and improve flow without global smoothing.
  9. Native fluency — when the target is not English, make the text read as native writing. See passes/native-fluency.md.
  10. Word check — ask of every word and sentence whether it fits the context and is used correctly in the target language. See /persona-wordcheck.
  11. Fidelity check — preserve meaning and nuance.
  12. Cold reader — when the reader decides something about the author or the purpose (an application, a bid, an appeal), or the user asks: a fresh reader with only the text reports what it understood, the impression of the author, and where it stumbled. See passes/cold-reader.md.
  13. Proofread — read the finished text as it will appear in the reader's form: length, format, spacing, characters. See passes/proofread.md.
  14. Final output — return inline (or write to final.md if scratch is in use).

If scratch is in use, write each step to a numbered .md file (01-intent.md ... 04a-skeleton.md, 04b-sentences.md ... 06b-native.md, 06c-wordcheck.md, 07-fidelity.md, 07b-cold-reader.md, 08-proofread.md, final.md; plus 00-style-sheet.md).

Long-form workflow

If the input is a large document, do not rewrite it in one shot. Use the scratch folder for all artifacts.

Stage 0: Create a global brief (brief.md)

Create or infer:

  • persona
  • audience
  • purpose
  • format
  • core thesis
  • must-keep facts
  • tone constraints
  • locked terminology

Stage 1: Create a document map (map.md)

Identify:

  • major sections or chapters
  • the role of each section
  • dependencies between sections
  • repeated concepts
  • likely drift risks

Stage 2: Process one section at a time

For each section or chapter, write to the task folder:

  1. local intent extraction ([section]-01-intent.md)
  2. local diagnostic audit ([section]-02-audit.md)
  3. local persona immersion ([section]-03-mapping.md)
  4. section build, sentence by sentence ([section]-04a-skeleton.md, [section]-04b-sentences.md, [section]-04-draft.md)
  5. voice coherence ([section]-05-coherence.md)
  6. local refinement, native fluency, and word check ([section]-06-refine.md, [section]-06b-native.md, [section]-06c-wordcheck.md)
  7. fidelity check ([section]-07-fidelity.md)
  8. chapter memory artifact ([section]-memory.md)

Stage 3: Maintain rolling memory

After each section, keep track of: ...

Stage 4: Create revision tickets when needed (tickets.md)

If a later section changes framing, terminology, or argument shape: ...

Stage 5: Run a whole-document consistency pass (consistency.md)

At the end, check: ...

Stage 6: Assemble final output (final.md)

Create the final Markdown document from the revised sections, then run the cold-reader pass (when the stakes call for it) and the proofread pass on it.

Output handling

Short tasks

Return the result inline by default unless the user clearly wants it saved.

Long or iterative tasks

Use Markdown files for:

  • global brief
  • document map
  • chapter drafts
  • chapter memory
  • revision tickets
  • final assembled output

Markdown is the canonical working format.

Output style by default

Unless the user requests detailed analysis, do not dump the full internal process.

For normal rewrite requests

Return:

  • a short note about what changed
  • the rewritten text

For audit requests

Return:

  • concise diagnosis
  • optional rewrite if requested or obviously useful

For long-form requests

Return:

  • short note that the text was handled section by section
  • the resulting output or the location of the generated Markdown artifacts

Preset personas

Preset persona definitions are in:

  • personas/sharp-technical.md
  • personas/pragmatic-builder.md
  • personas/clear-teacher.md
  • personas/skeptical-analyst.md
  • personas/blunt-operator.md
  • personas/problem-first-marketer.md
  • personas/quiet-witness.md

Supporting modules

Docs

  • docs/voice-guide.md — active guidance on voice coherence (positive-shape framing). Why the pipeline does not enumerate anti-AI patterns.
  • docs/persona-theory.md — research foundation for the persona schema (stylometry, rhetoric, education research, personality–linguistics)
  • docs/writing-quality-rubric.md — research-backed analytic rubric used by persona-copy and the diagnostic-audit pass
  • docs/persona-chain-mode.md
  • docs/pipeline.md
  • docs/longform-mode.md
  • docs/simple-usage.md
  • docs/philosophy.md
  • docs/contributor-guide.md
  • docs/anti-ai-guidelines.md — reference only; background on common AI patterns. Not a pipeline input.

Passes

  • passes/intent-extraction.md
  • passes/style-sheet.md
  • passes/diagnostic-audit.md
  • passes/persona-mapping.md
  • passes/sentence-build.md
  • passes/rewrite.md
  • passes/voice-coherence.md
  • passes/refine.md
  • passes/native-fluency.md
  • passes/fidelity-check.md
  • passes/cold-reader.md
  • passes/proofread.md

Scripts

  • scripts/proofread.py — mechanical checks used by the proofread pass; scripts/test_proofread.py is its self-check

Long-form

  • longform/global-brief.md
  • longform/document-map.md
  • longform/chapter-workflow.md
  • longform/chapter-memory.md
  • longform/revision-tickets.md
  • longform/consistency-pass.md

Companion skills

  • /persona-wordcheck — runs as pipeline step 10; install it alongside persona-write
  • /persona-brief, /persona-research, /persona-localize — optional steps before the pipeline

Dictionaries

  • dictionaries/banned-phrases.md
  • dictionaries/manager-speak.md
  • dictionaries/ai-patterns.md

Final reminder

This skill is not trying to produce "generic human-like writing."

It is trying to produce writing that sounds like a specific kind of person, while staying useful, coherent, and readable.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.