Critic Agent
Make Claude self-correct its own writing by spinning up an internal critic that reviews drafts against your voice and style rules — up to 3 rounds, no manual editing needed.
Download this file and place it in your project folder to get started.
# Critic Agent Protocol
## Your Role
You write content that matches the user's voice exactly, and you self-correct using an internal critic process before delivering final output.
Before writing anything, read `voice-dna.md` (and any other context files in this folder) to understand the user's voice, banned phrases, and style.
## The Critic Process
When asked to "run the critic" (or "critic this", "review this against my voice"), follow this review process on the current output.
### Steps
1. Spin up a sub-agent to act as a Critic.
2. The Critic reviews the full output against ALL context files (`voice-dna.md`, and any other context files in the folder).
3. The Critic assigns a rating: **Needs Work**, **Good**, or **Excellent**.
4. If below Excellent, the Critic provides specific, actionable feedback referencing the exact rules or samples it's checking against.
5. Revise the output based on feedback.
6. Run the Critic again on the revised version.
7. Repeat until Excellent, or until 3 review rounds complete (whichever comes first).
## What the Critic Checks
### Voice Match
- Does this read like the writing samples? Same rhythm, sentence length, formality level?
- Any banned phrases present? (Check every single one.)
- Does it use words and phrases the user actually uses?
- Would someone who knows the user recognize their voice?
### Substance
- Does the output actually answer what was asked, or an adjacent version of it?
- Are claims specific and grounded, or vague and generic?
- Is anything padded or restated in different words to seem thorough?
### Final Bar
- Would the user send, publish, or present this without editing?
- If not, what specifically needs to change?
## Critic Output Format
For each round, the Critic produces:
```
## Critic Review — Round [N]
### Rating: [Needs Work / Good / Excellent]
### Issues Found
| Issue | Location | Feedback |
|-------|----------|----------|
| [Type] | [Where] | [Specific, actionable feedback referencing context files] |
### Summary
[1-2 sentences on what to fix or why it passes]
```
## Rules for the Critic
- **Be specific.** "The tone is off" is useless. "Paragraph 3 uses 'Furthermore' which is banned, and the structure is more formal than any writing sample" is useful.
- **Reference actual context files**, not general writing standards.
- **Don't over-polish.** Natural voice includes imperfections, casual language, and hedging. That's a feature, not a bug.
- **3 rounds max.** If it's not Excellent after 3 rounds, deliver the best version with a note on what could still improve.
- **Check every banned phrase** against the output. Every single one.
## Instructions
1. When the user asks you to write something, read `voice-dna.md` first
2. Write the initial draft matching their voice as closely as possible
3. When they say "run the critic," execute the full critic loop
4. Present the final version with a summary of what changed across rounds
5. If the user hasn't created a `voice-dna.md` yet, ask them to create one before running the critic
## Commands
```
"run the critic on this"
"run the critic"
"critic this"
"review this against my voice"
```
What This Does
Eliminates the manual editing step when Claude writes for you. Instead of spending 15-30 minutes tweaking tone, cutting AI slop, and making output sound like you — Claude does it itself.
When you say "run the critic on this", Claude spins up a sub-agent whose only job is to find problems with what the first version wrote. The critic reads the draft against your voice file and returns specific, actionable feedback like:
- "The intro reads like a LinkedIn post, not like you"
- "Paragraph 4 is just restating paragraph 2 in different words"
- "This sentence uses 'furthermore' which is on your banned list"
Claude rewrites based on that feedback, then the critic reviews again — up to 3 rounds — until the output is something you'd actually send without editing.
Prerequisites
- A voice DNA file that captures how you write (see setup below)
- Any writing project where Claude generates content for you
Quick Start
Step 1: Create Your Voice DNA File
Create a file called voice-dna.md in your project folder. This teaches Claude how you actually sound. Include:
- 3-5 writing samples that represent your real voice (emails, posts, docs you've written)
- Banned phrases — words and phrases you never use (e.g., "furthermore", "leverage", "in conclusion", "it's worth noting")
- Voice traits — your rhythm, sentence length, formality level, quirks
- Formatting preferences — how you structure things, use of headers, lists, etc.
Example structure:
# Voice DNA
## Writing Samples
[Paste 3-5 real examples of your writing here]
## Banned Phrases
- "Furthermore"
- "It's worth noting"
- "In today's fast-paced world"
- "Leverage"
- "Synergy"
- [Add your own]
## Voice Traits
- Short, punchy sentences. Mix in longer ones occasionally.
- Casual but not sloppy.
- Uses "you" a lot. Talks directly to the reader.
- Okay with sentence fragments. Starts sentences with "And" or "But."
- Avoids jargon unless it's industry-specific and necessary.
## Formatting
- Short paragraphs (2-3 sentences max)
- Minimal headers
- Bullet points for lists, not numbered
Step 2: Download the Critic Agent Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ~/your-project-folder/
Make sure both CLAUDE.md and voice-dna.md are in the same folder.
Step 3: Run Claude Code
cd ~/your-project-folder
claude
Ask Claude to write something, then say: "run the critic on this"
How It Works
The Loop
- Claude writes your first draft
- You say "run the critic on this"
- Claude spins up a critic sub-agent
- The critic reads the draft against
voice-dna.mdand all other context files - The critic assigns a rating: Needs Work, Good, or Excellent
- If below Excellent, the critic gives specific, line-level feedback
- Claude rewrites based on that feedback
- The critic reviews the revised version
- Repeat until Excellent or 3 rounds (whichever comes first)
What the Critic Checks
Voice Match
- Does this read like the writing samples? Same rhythm, sentence length, formality?
- Any banned phrases present? (Checks every single one.)
- Does it use words and phrases you actually use?
- Would someone who knows you recognize your voice?
Substance
- Does the output actually answer what was asked, or an adjacent version of it?
- Are claims specific and grounded, or vague and generic?
- Is anything padded or restated in different words to seem thorough?
Final Bar
- Would you send, publish, or present this without editing?
- If not, what specifically needs to change?
Example
You: "Write me a cold outreach email to a potential podcast guest"
Claude writes a draft, then you say:
"run the critic on this"
Critic (Round 1 — Needs Work):
| Issue | Location | Feedback |
|---|---|---|
| Banned phrase | Line 2 | "I hope this email finds you well" — on banned list |
| Voice mismatch | Paragraph 1 | Too formal. Your samples never open with pleasantries |
| Padding | Paragraph 3 | Restates paragraph 1 in different words |
| Generic | Throughout | No specific reference to the guest's work |
Claude rewrites. Critic reviews again.
Critic (Round 2 — Good):
| Issue | Location | Feedback |
|---|---|---|
| Sentence length | Paragraph 2 | Three long sentences in a row. Your samples mix short and long |
Claude revises. Critic reviews once more.
Critic (Round 3 — Excellent): Ship it.
Trigger Phrases
All of these activate the critic:
"run the critic on this"
"run the critic"
"critic this"
"review this against my voice"
Tips
- Invest in your voice DNA file — the critic is only as good as the reference material you give it
- Include real writing samples — not aspirational ones. Paste actual emails, posts, and docs you've sent
- Update your banned list — when you catch AI slop you hate, add it to the list
- Works on everything — emails, posts, docs, proposals, scripts, bios
- Don't over-polish — the critic is trained to preserve natural imperfections. Casual language and hedging are features, not bugs
Troubleshooting
Critic is too harsh / never reaches Excellent Simplify your voice DNA. Too many constraints make it impossible to pass. Start with 3 samples and 10 banned phrases.
Output loses personality after revisions Add a note to your voice DNA: "Natural imperfections are intentional. Don't smooth everything out."
Critic feedback is vague Make sure your voice DNA has concrete examples, not just descriptions. "Short sentences" is vague. Actual writing samples are specific.
Critic ignores banned phrases
List them explicitly in your voice-dna.md under a clear "Banned Phrases" header. The more structured, the better.