AI Guides › Workbench
By Nigel Guy · 6 min read
You finish a draft, plan or piece of code, and in the same chat you type "can you check this properly?" The answer is a polite tick with two small suggestions. That feels like a review, but the reviewer has just watched itself write the thing, knows every reason behind every choice, and has the whole conversation nudging it towards "yes, that holds up". It is marking its own homework with the answer key open.
The rule: never review in the context that produced the work. Hand the work to fresh reviewers who have not seen your reasoning, give each one a different job, and run them side by side.
"Fresh context" is the mechanism. You can get it several ways.
| Option | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| New chat in Claude | Blank slate; you paste in only the work | Free plan exists; Pro lists at $20 a month on monthly billing (about £15 at time of writing; check the £ price at checkout) | Writing, plans, emails, anything without files | You are the messenger: paste three times, compare by hand |
| Subagents in Claude Code | Each runs in its own context window with its own system prompt and tool access, and returns only a summary to your main session | Claude Code is listed as included with Pro and Max; Max is from $100 a month (about £75 at time of writing; check at checkout) | Code, documents in a folder, anything Claude can read from disk | Subagent requests count towards the same usage limits as your main conversation |
| Saved reviewer file | A subagent defined once in .claude/agents/ (this project) or ~/.claude/agents/ (all projects) |
Same as above | Reviews you repeat weekly | Needs a clear description, or Claude will not know when to delegate |
| Another model or tool | A different vendor's assistant as reviewer | Varies; check the vendor | Catching blind spots shared by one model family | Different formatting, no shared files |
Claude Code docs say each subagent "runs in its own context window", has independent permissions, and returns only a summary to the main conversation. That is exactly the separation you want. Anthropic changes plans and limits often, so check the pricing page and the Claude Code documentation before you rely on any figure here.
Save the thing under review to a file or a single pasted block. Do not summarise it. A reviewer who sees your summary is reviewing your opinion of the work.
Give reviewers the work, the stated goal and the audience. Leave out your doubts, your preferred outcome and phrases like "I think this is pretty solid". Anything you say about how good it is will be echoed back.
Three copies of "review this" produce three similar reviews. Split by job: one reads as the intended recipient, one tries to break it, one checks claims against evidence. The three prompts below do that.
In Claude Code, ask for it plainly: "Run these three prompts as three separate subagents in parallel, give each only the file [FILE_PATH], and show me each answer unedited." Docs list parallel research as a standard pattern, and by default up to 20 subagents can run at once. Keep it to three here. Each one spends your usage, and running several draws more than a single reviewer does.
In an ordinary chat window, open three new chats and paste one prompt into each.
Read the three answers side by side. Issues raised by two or more reviewers go to the top. A lone objection is not automatically right; check it against the work.
Fill in the audience, the goal and the work.
You are a member of this audience: [AUDIENCE_DESCRIPTION]. You are meeting
the material below for the first time and have no background beyond what
it states.
The author's goal is: [GOAL_OF_THE_WORK].
Read the material once, as a real reader would, then report:
1. In two sentences, what you believe it is asking of you or telling you.
2. The first point where you got lost, bored or doubtful, quoting the line.
3. Anything you would need to know that it never says.
4. Your honest verdict: would you act on it? Yes, no or only with changes,
and why.
Rules: do not praise anything unless you can say what specific effect it
had on you. If the material is too thin to judge, say what is missing
rather than guessing. Before answering, check that every point you make
quotes or points to a specific place.
MATERIAL:
[PASTE_WORK_OR_FILE_PATH]
Fill in the context in which the work will be used.
You are a sceptical specialist whose job is to find how the work below
fails once it is used for real. Context of use: [WHERE_AND_HOW_IT_WILL_BE_USED].
Goal: [GOAL_OF_THE_WORK].
Produce a table with columns: Failure, Trigger (what has to happen),
Severity (high, medium, low), Cheapest fix. List at most eight rows, worst
first. Then name the one assumption the whole thing depends on.
Rules: only report failures you can tie to specific text in the work. Do
not pad with generic risks. If you find fewer than eight, stop at what you
found. If you need information about the context that I have not given,
ask me before you start. Before answering, remove any row that would apply
equally to any piece of work.
WORK:
[PASTE_WORK_OR_FILE_PATH]
Fill in the work and any sources you hold.
You are a fact-checker. List every factual claim, number, name, date and
cause-and-effect statement in the work below. For each, mark it:
SUPPORTED (the supplied sources back it), UNSUPPORTED (no source given),
or CONTRADICTED (a supplied source disagrees, quote it).
Sources I am providing: [SOURCES_OR_NONE].
Rules: do not use memory to mark anything SUPPORTED; if I gave no sources,
mark claims UNSUPPORTED and say which would be easiest to verify and where.
Do not rewrite the work. End with the three claims that would do the most
damage if wrong. Before answering, confirm that you quoted each claim
rather than paraphrasing it.
WORK:
[PASTE_WORK_OR_FILE_PATH]
If you run the breaker often, create it with /agents or ask Claude to write a file in ~/.claude/agents/. The format is YAML frontmatter, then the prompt body:
---
name: breaker
description: Finds how finished work fails in real use. Use after drafting, before sending.
tools: Read, Glob, Grep
---
[Paste Prompt 2 here, leaving the placeholders for the task message.]
Limiting tools to read-only means the reviewer cannot quietly "fix" your file and then approve its own change.