Claude Subagents: The Exact Prompts To Catch Its Own Mistakes

When you ask the chat that wrote something to go back and check it, it reads its own work with the same blind spots that made the mistake. It’s like proofreading your own email at midnight. You miss the typo, and it swears it looks perfect.

The fix is a subagent, a second Claude that starts with a blank context and reviews your work like a coworker seeing it for the first time. Fresh eyes catch what the first one talked itself into.

Same Claude, blank slate

A subagent runs in its own separate context window, with its own instructions, and hands back only a summary. It can’t see the reasoning that produced the work, so it can’t inherit those blind spots. It can run on the same model as your main chat, so the reviewer is just as smart as the writer.

Where you use it

Subagents live in Claude Code, Anthropic’s coding agent (terminal, desktop app or IDE). Paste any prompt below and Claude spins one up. To save a reviewer you’ll reuse, just ask Claude to create a subagent for it. It saves a small file in .claude/agents/ for one project or ~/.claude/agents/ for all of them. (Anthropic’s subagent docs)

No Claude Code? The same idea works anywhere. Open a brand-new chat, paste only the finished work, and use the prompt with "Spawn a subagent to" swapped for "You are reviewing this cold." The magic is the clean context.

Fresh eyes on anything

The one you’ll reach for most. Works on a doc, an email, a plan, a caption or code.

Cold Review

Spawn a subagent with completely fresh eyes to review the work above. Give it only the final work, not our conversation, and tell it what the work is for: [what this is and who it's for]. The subagent should not assume anything I said earlier is correct, or that any choice was made for a good reason. Have it report back in this format: 1. WRONG: anything incorrect, broken or contradictory 2. MISSING: what a reader or user would need that isn't there 3. WEAK: what's vague, padded or unclear 4. FIX LIST: one prioritized list of changes, ranked by how much each improves the result Rules: be blunt and specific, quote the exact line for every issue, and if something is strong, say so in one line and move on. Don't rewrite the whole thing. Show me the fix list before you change anything.

Attack it on purpose

Run this before any decision, pitch or plan. It stress-tests your thinking now, so reality doesn’t do it later.

Devil’s Advocate

Spawn a subagent whose only job is to argue against the plan above. Give it the plan and the goal: [what I'm trying to achieve]. Have it take the opposite side and build the strongest possible case that this is wrong. It should: 1. List the 3 weakest assumptions, and what happens if each one is false 2. Point out the holes in the logic 3. Name the risks I'm glossing over, including cost, time and who could say no 4. Describe the most likely way this fails 6 months from now Push it to be the smartest critic in the room, not a cheerleader. No softening and no "but overall it's great." End with the single objection I most need a good answer to before I move forward.

Catch confident wrong answers

AI sounds the most sure of itself right when it’s wrong. Send a subagent in to verify, not to agree.

Fact-Checker

Spawn a subagent to fact-check the work above with a skeptical eye. Give it only the work, not our conversation. Its job is to hunt for claims that are wrong, exaggerated, outdated, unverified or made up. That includes numbers, names, dates, prices, product features, quotes and any "everyone knows" statement. For each claim, return a table row: CLAIM (quoted) | VERDICT (correct / wrong / outdated / can't verify) | WHY | ACCURATE VERSION or WHAT I'D NEED TO CHECK Rules: if it can search the web, check primary sources (official sites, the original study) and link them. If it can't, say "can't verify" instead of guessing. Assume at least one thing in here is wrong, and go find it.

See it through their eyes

For anything that makes sense to you but lands flat on everyone else. Fill in who it’s for.

Reader Reaction

Spawn a subagent that role-plays my exact audience and reacts to the work above the way that real person would. Who it's for: [describe your reader, customer or boss in one or two lines, including what they care about and how busy they are] What I want them to do after reading: [the action] Give the subagent only the work, not our conversation. Have it read cold and tell me, in its own voice: 1. The exact line where it would stop reading or start skimming 2. What confuses it 3. What it quietly doubts or disagrees with 4. Whether it would take the action, and the one thing that would change its mind Make it react like a real, slightly impatient human, not a helpful assistant trying to be nice. Then, as yourself, give me the 3 edits that would fix the biggest problems.

Three experts, one verdict

My favorite. Three independent opinions instead of one, and Claude picks the right experts for whatever you’re working on. None of them can see each other’s notes, so there’s no echo chamber.

Expert Panel

Spawn three subagents to review the work above, each in its own separate context so they can't see each other's notes. First, you choose the three most useful expert lenses for this specific kind of work and assign one to each (for example a sharp editor, a real subject-matter expert, a skeptical buyer, a lawyer or a numbers person). Tell me who you picked and why, in one line each. Give each expert only the work and its goal: [what this is for]. Have each expert return its top three changes, ranked by impact, with the exact line each change applies to. Then step back and give me: - AGREE: where 2 or more experts flagged the same thing - DISAGREE: where they conflict, and which side you'd take - FIX FIRST: the 3 changes I should make first Don't make any changes until I pick which fixes to apply.

One Rule Under All Of It

Never let your AI become an echo chamber. This works because the reviewer can’t see the reasoning that made the mistake. Heads up: every subagent counts toward the same usage limits as your main chat, so save the panel for work that matters.

Make ClaudeCatch Its Mistakes

Make Claude
Catch Its Mistakes