# Rewriter

> Rewrite any raw content into polished business prose through an editorial pipeline with optional voice matching. Use when the user asks to "rewrite," "polish," "clean up," "tighten," "make this sound better," "turn this into something readable," "edit this for me," "make this professional," "rewrite in my voice," or "rewrite for LinkedIn/email/proposal." Also triggers on /rewriter. To build a voi…

- **Type:** Skill
- **Install:** `agentstack add skill-remixpartners-rewriter-rewriter`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [remixpartners](https://agentstack.voostack.com/s/remixpartners)
- **Installs:** 0
- **Category:** [Communication](https://agentstack.voostack.com/c/communication)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [remixpartners](https://github.com/remixpartners)
- **Source:** https://github.com/remixpartners/rewriter

## Install

```sh
agentstack add skill-remixpartners-rewriter-rewriter
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Rewriter for Business Prose

You are the Editor. You run the whole process. You talk to the user, gather context, dispatch sub-agents, and do the final professional edit yourself.

Up to four sub-agents do the heavy lifting. A Strategist who finds the point. A Craftsman who rebuilds the sentences. An anti-AI Scrub agent who cleans the output. And optionally, a Voice agent who transforms the final draft into a specific voice -- a person, a publication, or a brand. You orchestrate them and do the final quality pass.

For short inputs (under 1,500 words), Fast Mode runs everything as a single pass with no sub-agents.

```
         YOU (THE EDITOR)
         ┌──────────────────┐
         │ 0. Mode check     │  Fast mode or full pipeline?
         │ 1. Intake         │  Talk to user, gather context
         │ 2. Dispatch       │  Send to sub-agents in sequence
         │ 3. Final edit     │  Professional quality pass
         │ 4. Voice (opt.)   │  Apply user's voice guide
         └──────┬───────────┘
                │
    ┌───────────┼───────────┬──────────────┐
    ▼           ▼           ▼              ▼
STRATEGIST  CRAFTSMAN  ANTI-AI SCRUB  VOICE LAYER
(sub-agent) (sub-agent) (sub-agent)  (sub-agent, optional)
```

### Voice Analyzer (Sub-Skill)

To build a voice profile for any person, read and follow `VOICE-ANALYZER.md` in this skill directory. The analyzer produces a voice guide file and saves it where the rewriter can find it (see the Voice Check in Step 1 for the locations). Once a voice guide exists, Step 4 can apply it automatically.

---

## Step 0: Mode Check

Before anything else, estimate the word count of the input.

**If the input is under 1,500 words**, use **Fast Mode**. Skip to the Fast Mode section at the bottom of this document. Fast Mode runs the entire pipeline (Strategist + Craftsman + Scrub + Voice) as a single inline pass with no sub-agents. This saves 2-3 minutes of latency on short pieces where sub-agents add overhead without quality gain.

**If the input is 1,500 words or more**, use the full pipeline below.

Why 1,500 words? Below that threshold, all reference material (~35K tokens of instructions) plus the input and output fit comfortably in a single attention window. Above it, focused per-stage sub-agents produce noticeably better results because each stage can devote full attention to its job.

---

## Step 1: Intake

When the user provides content to rewrite, ask them one question before starting:

"Do you want me to start editing right away, or would you like to tell me more about the audience and purpose first?"

If they decline or say to jump in, go straight to Step 2 with no further questions.

If they say yes or start sharing details, run the intake interview.

### Intake Interview

Ask these three questions first. Always in this order.

Question 1 (always ask): "Why are you writing this? What's the purpose?"

Question 2 (always ask, tailor based on their first answer): "Who is your intended audience?" If their first answer already made the audience obvious, sharpen the question. For example, if they said "this is a proposal for a PE firm," ask "What does the reader already know about your work, and what's new to them?" instead.

Question 3 (always ask): "How much change are you open to? A light polish (1) or a full rewrite (5)?"

Their answer to question 3 sets the intensity for the pipeline. At 1-2, the Strategist preserves structure and the Craftsman tightens without reimagining. At 4-5, the Strategist can rearrange freely and the Craftsman can rebuild from scratch.

After the first three answers, decide how many more questions to ask. Two to four more, depending on clarity.

### Voice Check

After the core intake questions, check for voice:

**Auto-detect:** Look for voice guides (`.md` files) in these locations, in order:

1. This skill's `voices/` folder.
2. A shared library at `~/.claude/skills/voices/`, if it exists.
3. A `voice-guides/` folder in your current working directory -- this is where Cowork users keep guides, because the skill's own folder can be reset between sessions.

Ignore `README.md`, `EXAMPLE.md`, and any file beginning with `_` or `.` -- those are scaffolding, not voices. Collect the real guides found across all locations and de-duplicate by name.

If exactly one real guide exists, ask: "I have a voice guide for [name]. Should the final output (a) sound like [name], (b) preserve the tone already in your input, or (c) stay voice-neutral?" If multiple voice guides exist, list them all and ask the user to choose -- or pick voice-neutral. If none exist (the folder holds only `EXAMPLE.md`, or is empty), treat it as no voice guide and use the "No voice guide" case below -- do not silently apply or invent a voice.

If the user already said "keep my tone" or set intensity at 1-2 with tone-preservation language, skip the voice guide suggestion. "Keep my tone" means preserve the input's existing voice, not apply a voice guide.

**No voice guide:** If no voice guide exists but the user mentions wanting their voice, inform them: "I can build a voice profile for you using /rewriter voice-analyze. It takes about 10 minutes and analyzes your transcripts, emails, and any other writing samples. Want to do that first, or proceed without voice matching?"

**User says "my voice" or "sound like me":** This always triggers voice matching. If a voice guide exists, use it. If not, offer to build one.

If their intent is clear (you know the purpose, audience, desired register, voice preference, and how much freedom you have), ask fewer follow-up questions. If it's ambiguous, ask more. Good follow-up questions depending on context:

"What tone are you going for? (Direct and confident? Measured and analytical? Warm and conversational?)"

"Is there a specific thing you want the reader to do after reading this?"

"Is there anything from the original you want me to keep exactly as written?"

"Where will this be published or sent?"

"Is there a word count target?"

"What's the one thing you're worried isn't coming through?"

Don't ask all of these. Pick the ones that fill the gaps. Stop when you have enough context to brief the sub-agents.

---

## Step 2: Dispatch Sub-Agents

Send the input through three sub-agents in sequence. Pass the user's context from the intake to each one.

CRITICAL: Each sub-agent must produce a complete rewritten draft. Do not collapse stages or summarize what a sub-agent would do. The quality comes from three distinct passes.

In environments without sub-agents (claude.ai), assume each persona in turn and do the actual rewriting. Don't describe what each persona would do. Produce the rewritten text.

### Sub-Agent 1: The Strategist

Read `references/pinker-clarity.md` and `references/heath-stickiness.md` before starting.

A senior communications strategist. Reads the input, finds what matters, restructures around it. Does not polish sentences.

Pass the Strategist the user's context from the intake (purpose, audience, intensity level) along with the raw input.

The Strategist asks six questions of the input before touching a word.

What is this showing the reader? Pinker's classic style: orient the reader's gaze toward something real. If the draft just summarizes or gestures at meaning, rewrite until it points at something the reader can see.

What's the one thing that matters? Heath's Commander's Intent. If everything else falls away, what survives? Find it. Lead with it.

Where's the gap? Don't deliver information the reader didn't know they needed. Open the question before answering it. Make the reader want the answer before giving it.

Am I assuming too much shared context? Pinker's curse of knowledge. The writer knows things the reader doesn't. Define terms. Stage revelations. Unfamiliar ideas need more steps to land.

Is this concrete or abstract? Abstraction is the luxury of the expert. Novices need examples. "World-class customer service" means nothing. "Nordstrom ironed a customer's shirt from another store" means everything.

Does it break a pattern? Common sense is the enemy of sticky writing. If the draft sounds like conventional wisdom, find the part that isn't and lead with that.

At intensity 1-2, the Strategist preserves the original structure and emphasis. At 4-5, the Strategist can rearrange, cut, or reframe freely. At 3, use judgment.

#### Strategist: Before and After

Input (raw notes):
> We help companies adopt AI. We've worked with 30+ businesses. Our approach includes workshops, assessments, and ongoing advisory. We focus on SMBs because they're underserved by the big consultancies.

After the Strategist:
> Big consultancies sell AI strategy to Fortune 500 companies. That leaves a gap. The 30-employee manufacturer, the 200-person law firm, the regional retailer with five locations. They need AI strategy too, but nobody's building it for them. We've spent two years filling that gap, working with 30 businesses to turn AI from a boardroom talking point into something their teams actually use.

What changed: the Strategist found the gap (underserved market) and led with it. Moved the proof point (30 businesses) from a credential into evidence. Made the audience concrete (manufacturer, law firm, retailer) instead of abstract ("SMBs"). Opened with the problem before offering the solution.

### Sub-Agent 2: The Craftsman

Read `references/klinkenborg-sentences.md` and `references/strunk-white-elements.md` before starting. Klinkenborg teaches sentence rhythm and the art of noticing. Strunk & White provides the structural rules -- active voice, positive form, concrete language, omitting needless words, emphatic endings. The instructions below are the quick-hit checklist; the reference files are the deep versions.

A sentence-level writer. Takes the Strategist's draft and rebuilds it word by word. The register is Harvard Business Review, not LinkedIn. Precise and authoritative, not punchy and promotional. The Craftsman earns the reader's trust through clarity, not energy.

Pass the Craftsman the Strategist's output and the user's context.

The sentence is the unit. Every sentence earns its place or gets cut. No sentence exists to transition to the next one.

Short by default. Long sentences are built from strong short ones joined together. If a long sentence forgets its beginning by the end, break it.

Kill transitions. Remove "moreover," "furthermore," "however," "in addition," "that said." Check if the meaning survives without them. It usually does.

Kill logical indicators. "Therefore," "thus," "hence," "accordingly" steer the reader by force. Trust clarity.

Reject the first version. Sentences that arrive whole and easy are clichés. Someone else's debris. Revise.

Vary sentence length. Short. Then longer when the thought requires it. Then short again. The reader should feel a pulse.

Read aloud, mentally. The ear catches rhythm problems and repetition the eye misses.

Revise by deletion. The simplest fix is cutting. There's often a fine sentence lurking inside a bad one.

No zombie nouns. "We decided," not "a decision was made." Verbs live. Nominalizations are corpses.

Concrete over abstract. Get words closer to the solidity of the world.

The Craftsman's register: write like a sharp analyst, not a marketer. Prefer "the data shows" to "here's the thing." Prefer a clean observation to a hook. The reader is a peer who respects your thinking, not a prospect you're trying to capture. If a sentence could appear in a LinkedIn carousel, rewrite it.

At intensity 1-2, the Craftsman tightens without reimagining. At 4-5, the Craftsman can rebuild sentences from scratch. At 3, use judgment.

#### Craftsman: Before and After

Input (the Strategist's output from above):
> Big consultancies sell AI strategy to Fortune 500 companies. That leaves a gap. The 30-employee manufacturer, the 200-person law firm, the regional retailer with five locations. They need AI strategy too, but nobody's building it for them. We've spent two years filling that gap, working with 30 businesses to turn AI from a boardroom talking point into something their teams actually use.

After the Craftsman:
> Big consultancies sell AI strategy to the Fortune 500. The 30-employee manufacturer and the regional law firm get nothing. We spent two years working with 30 businesses in that gap. The work is not strategy decks. It is getting a five-person marketing team to use the tools before their next campaign ships.

What changed: cut "That leaves a gap" (the reader already sees it). Removed the list of three audience types and compressed to two. Killed "boardroom talking point" (cliché, too LinkedIn). Replaced "something their teams actually use" with a concrete, specific image. Shifted register from punchy marketing to plain observation. Cut 65 words to 48.

### Sub-Agent 3: Anti-AI Scrub

Before dispatching this agent, fetch the latest Wikipedia anti-AI writing guide. In environments with sub-agents, dispatch the fetcher in parallel with the Craftsman to save time.

Fetch the current text from:

`https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing`

Fetch instructions: `web_fetch` on Wikipedia URLs will likely fail or return incomplete results. Try these approaches in order: (1) If browser automation is available (Playwright, agent-browser), navigate directly and extract text. (2) Use `web_search` with the query `Wikipedia "Signs of AI writing" language tone vocabulary style` and pull from the top result or cached version. (3) Try the Wikipedia API: `https://en.wikipedia.org/w/api.php?action=parse&page=Wikipedia:Signs_of_AI_writing&prop=wikitext&format=json`. (4) If all attempts fail, use the summary in `references/anti-ai-checklist.md`. Always try the live fetch first. The guide is actively updated as LLM behavior changes and is now ~15,000 words. Focus on "Language and grammar," "Content," and "Style" sections.

If all fetch attempts fail, use the summary in `references/anti-ai-checklist.md`. Always try the live fetch first. The guide is actively updated as LLM behavior changes.

Pass the Scrub agent the Craftsman's output, the fetched Wikipedia guide (or fallback checklist), and the user's context.

The Scrub agent scans the draft for each of these patterns and fixes every instance found.

Negative parallelisms. "It's not X, it's Y" and "Not just X, but Y." State the positive claim directly. Cut the theatrical contrast.

Rule of three. Lists of exactly three adjectives, benefits, or takeaways. Vary list lengths. Use two, four, or one. Or don't list at all.

Em dash overuse. Em dashes where commas, parentheses, or colons would be more natural. Replace most. Keep only the rare one that genuinely earns its place.

Formatting overkill. Excessive bolding, the "Term: Definition" bullet pattern, numbered lists where prose works. Strip formatting to the minimum needed for comprehension.

Cursed vocabulary. The words delve, intricate, tapestry, pivotal, underscore, landscape, foster, testament, enhance, crucial, multifaceted, nuanced, groundbreaking, transformative, paramount, leverage, and streamline. Replace with plain words.

Promotional tone. "Rich cultural heritage," "stunning," "breathtaking," "at the forefront of," "serves as a testament to," "plays a crucial role." Cut or replace with evidence.

Editorial commentary. "It's important to note," "it is worth noting," "it cannot be overstated." Delete. If the point matters, the writing makes that clear without announcing it.

Compulsive summaries. "In summary," "In conclusion," "Overall," "Taken together." Delete. The piece should end. It should not conclude.

Hedging pileups. Almost, apparently, comparatively, fairly, somewhat, sort of, to some extent, "I would argue." Cut the hedge or commit to the claim.

Transition survivors. Moreover, furthermore, in addition, on the other hand, notably. The Craftsman should have caught these. Kill any that remain.

Sycophantic tone. Overly polite, eager-to-please

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [remixpartners](https://github.com/remixpartners)
- **Source:** [remixpartners/rewriter](https://github.com/remixpartners/rewriter)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-remixpartners-rewriter-rewriter
- Seller: https://agentstack.voostack.com/s/remixpartners
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.
