> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cogniagent.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Thorough review

> Make your AI coworker check its own work properly before handing it back — for the jobs where being wrong is expensive.

Before your coworker calls a job done, have it check its own work properly.

## What it does

Turn on Thorough review with `/review` and your coworker adds a real checking pass before it finishes: re-reading what it produced, testing its conclusions against the source, and looking for the thing it got wrong rather than confirming what it already believes.

It's off by default because it costs more and takes longer. On work that matters, it's the cheapest insurance available.

## When it earns its keep

* Anything going to a customer or an executive
* Figures someone will make a decision on
* A document that will be published or filed
* Work you won't have time to check yourself
* A long job where an early mistake would carry through everything after it
* The first time a coworker does an unfamiliar kind of task

## When to skip it

Exploratory work, drafts you'll read closely anyway, and quick lookups. Reviewing a first draft you're about to rewrite is paying for care you don't need.

## Good to know

<Tip>
  **Pair it with `/plan`.** Plan mode checks the approach before, review checks the result after. On genuinely important work, both.
</Tip>

<Note>
  Review checks the work against what you asked for. It can't tell you that you asked for the wrong thing — that's what [Plan mode](/cowork/chat/plan-and-build) is for.
</Note>

<Tip>
  For a repeated job where quality always matters, put the checks in a [skill](/cowork/skills/creating) instead. "Before finishing, verify every figure against the source" becomes part of the procedure rather than something you remember to switch on.
</Tip>

## Common questions

<AccordionGroup>
  <Accordion title="Does it actually catch things?" icon="magnifying-glass">
    Yes — most often arithmetic that drifted, a claim that isn't quite what the source said, and requirements from your original message that got dropped along the way.
  </Accordion>

  <Accordion title="How much slower is it?" icon="clock">
    Noticeably, because it's doing real additional work. On a task worth reviewing, that's the trade you want.
  </Accordion>

  <Accordion title="Can I leave it on all the time?" icon="toggle-on">
    You can, and you'll pay for care you often don't need. Better to switch it on for the tasks that warrant it.
  </Accordion>

  <Accordion title="What does it do if it finds a problem?" icon="wrench">
    Fixes it and tells you what it fixed. You see the correction rather than a silently different answer.
  </Accordion>

  <Accordion title="Is this the same as a reviewer helper?" icon="user-group">
    Similar in spirit. A [helper](/cowork/skills/helpers) is a named specialist you write with your own standards; this is a general checking pass with nothing to set up.
  </Accordion>
</AccordionGroup>

## Next steps

<CardGroup cols={2}>
  <Card title="Plan mode and Build mode" icon="compass-drafting" href="/cowork/chat/plan-and-build">
    Checking before, not after.
  </Card>

  <Card title="Custom helpers" icon="user-group" href="/cowork/skills/helpers">
    A reviewer with your standards.
  </Card>

  <Card title="Effort and model" icon="gauge-high" href="/cowork/chat/effort-and-model">
    The other quality dial.
  </Card>

  <Card title="Seeing what changed" icon="code-compare" href="/cowork/files/changes">
    Check it yourself.
  </Card>
</CardGroup>
