Repetitive Document Review

Some documents get checked against the same list every single time — government filings, contracts, settlement paperwork. A person misses different items depending on the day. An agent applies the same standard every time.

The goal isn’t to remove the human. It’s to clear the front end so the human only does the final check.

What you get

• Attach a document in the chat and the agent runs the whole checklist against it.
• For anything missing or inconsistent, it tells you which item and why.
• The findings are written up as a review report on a canvas.
• You read the report and make the call.

What to turn on

Attachments are readable without any extra tool. Turn on the skill above when you need more precision on structurally complex documents, like PDF tables.

⚠️ Korean HWP files (.hwp) cannot be attached in chat. Supported formats include PDF, Word (.doc/.docx), Excel (.xls/.xlsx), CSV, PPT, and images. Export HWP to PDF first. The full list is in Attaching files.

Step 1: Pin down the checklist first

This recipe lives or dies on checklist quality. A vague standard produces vague results.

Ask the agent:

I want a checklist for reviewing the grant applications we submit.
I'll tell you the rejection reasons we've had — turn them into items
and save them to checklist/application.md.

Then give it the real rejection reasons. A good item is one where two people can’t disagree about the verdict.

Step 2: Write the system prompt

Role: submission document reviewer.

Every review:
1. Read checklist/application.md for the list of items to check.
2. Go through the attached document item by item, in order.
3. Mark each item as pass / missing / inconsistent.
   - Note where in the document the verdict came from.
   - If an item cannot be verified from the document, mark it
     "cannot verify" — never mark it pass.
4. Write the review report on a canvas, structured as:
   - One-line verdict (ready to submit / needs fixing)
   - Problem items (name · verdict · evidence · how to fix)
   - Summary of items that passed
5. If anything needs fixing, put the count on the first line of the reply.

Principle: if it isn't certain, it doesn't pass.
Ambiguous items are marked "cannot verify" and handed to a person.

The last line matters most. If ambiguity resolves to “pass,” the review is worthless. Leaving “cannot verify” behind is far better.

Step 3: Run a first review

Test it on a document whose outcome you already know.

Review this application. (attach file)

The most convincing test is feeding it a document that was actually rejected and seeing whether it catches the reason. If it doesn’t, the checklist is missing an item — just say so:

This one was rejected for a missing attachment but came back as pass.
Fix the checklist so it catches that.

Step 4: Promote the checklist into a skill

Once the checklist stabilizes, turn it into a skill. Other agents can then inherit the same standard, and changes only need to happen in one place.

Turn the review procedure we're using into a skill.

The agent drafts one and drops a review card in the chat. See Creating skills.

Tuning

• One checklist per document type — checklist/application.md, checklist/settlement.md, then just say “review the settlement doc.”
• Keep a log — “after each review, append the date, document name, and problem count to review/log.md” shows you which items keep failing.
• Fix the report format — if a department expects a specific layout, put that layout straight in the prompt.

Next steps

• Attaching files — formats, size, and count limits
• Creating skills — turning a review procedure into something reusable
• Morning Briefing & Change Watch — work that runs without anyone asking

Advanced

Why “cannot verify” has to exist

The worst failure mode for a review agent is not reporting what it missed. With only pass/fail, ambiguous items drift toward pass — the report looks clean while containing items nobody actually verified.

A third verdict keeps those items on the page so a person can look at exactly those. The value of a review isn’t knowing what was checked; it’s knowing precisely what wasn’t.

How long attachments live

• Images stay in the chat for 90 days, but documents, spreadsheets, and slides can’t be reopened or downloaded from the chat later. They’re delivered to the agent, read, and used on the spot.
• Keep your own copy of anything you’ll need again.
• Files per message: up to 4 on Free, Starter, and Pro; up to 10 on Enterprise. Split large batches across messages.

Scanned documents and image PDFs

A scanned PDF with no selectable text can’t be extracted. The agent reads it as an image instead, which is less accurate. Don’t run precision checks — number reconciliation especially — against scans; use the original file.

Cost notes

• Longer documents (tens of pages) cost more per review. If the checklist only needs the first few pages, saying “check pages 1–3 only” helps a lot.
• Review quality drives the outcome. Start on the default Standard tier; if verdicts on ambiguous items keep wobbling, move the agent to Ultra (expect roughly 10× the SU).
• Canvas reports cost in proportion to length. “Detail the problem items; one line each for passes” keeps it bounded.