Installation kit · v1.0 · reviewed 2026-10-01
For: "I review vendor contracts every week and always check for the same issues."
Overview
What you're building
A review workspace loaded with your criteria. You upload a document, and it returns a structured findings table using the same checklist each time, with each finding tied to a quoted passage so you can verify it quickly.
Why this setup
Consistency is the point of a review. A standing checklist in the workspace means item 7 gets checked on the fortieth document just like the first. Requiring quotes for every finding makes the AI's output checkable rather than something you have to trust.
The workflow architecture
AI workspace
Holds your instructions and reference files so you never re-explain.
Reusable instructions
Review method and output format.
Reference document
Your criteria list or style guide.
Your review
You verify each flagged item against the quote.
You haven't picked a tool yet, so this kit is platform-neutral. Both Claude and ChatGPT support this setup.
| Claude | ChatGPT |
|---|---|
| Claude Projects: instructions + project knowledge files | ChatGPT Projects: instructions + uploaded files |
| Optional Claude Skills for reusing a format across chats (plan-dependent) | Personalization settings can hold short preferences |
| Paid plan needed for Projects | Plan determines Project and file limits |
Optional: a stable checklist used across many chats can be packaged as a Claude Skill if your plan supports Skills. The Project is sufficient for most people.
What to prepare
- □Your review criteria written as a numbered list (or your existing style guide/policy).
- □Two past documents: one that passed and one with known problems, for testing.
- □What your output should look like (comments, a table, a pass/fail summary).
Build it once
Copy-ready materials
Edit anything in [BRACKETS]. Changes are kept when you save the kit.
Project Instructions
Defines the review method.
You review [DOCUMENT TYPE, e.g. vendor contracts] against the criteria in the reference file "REVIEW CRITERIA". FOR EACH DOCUMENT 1. Check every criterion, in order. Do not skip any. 2. Output a table: # | Criterion | Status (Pass / Issue / Not found) | Exact quote | Suggested fix 3. Every "Issue" must include a direct quote from the document. If you can't quote it, mark "Not found". 4. End with: overall summary (3 sentences max) and the top 3 items for me to look at first. RULES - Do not rewrite the document unless I ask. - Do not give legal or regulatory conclusions; flag items for my judgment. - If the document is a different type than expected, say so before reviewing.
Review Criteria
Edit into your own list, then upload.
REVIEW CRITERIA – [DOCUMENT TYPE] v[1] [DATE] 1. [e.g. Term length is stated and is 12 months or less] 2. [e.g. Auto-renewal clause present? Flag notice period] 3. [e.g. Payment terms are net 30 or longer] 4. [e.g. Liability is capped] 5. [e.g. Our standard confidentiality language is included] 6. [e.g. Names, dates, and amounts are consistent throughout] 7. [Style: e.g. Uses sentence-case headings] 8. [Anything that always needs a human decision]
How Claude can help execute this
Why use this instead of a regular chat?
A chat reviews one document you paste. A task can review a batch of files against your criteria and produce one findings table per document.
- Task execution needs a paid Claude plan. Add your plan in My AI Setup for a more specific recommendation.
- This involves sensitive information. Keep Claude asking before it acts (the default), and approve each step.
Availability depends on your Claude plan and device. Add them in My AI Setup for a more specific recommendation.
1. Set up once
- · Create a Project with the review instructions and criteria file.
2. Claude capabilities used
- · Multi-step task execution (Cowork-style)
- · Files saved to your Claude account
3. Access, files & permissions
- · The documents to review
- · A paid Claude plan for task execution
4. What Claude can do
- · Check each document against every criterion
- · Produce findings with quotes
5. What you review or approve
- · Every flagged item against its quote
- · Any final decision — this is a first-pass screen
6. What changes each run
- · The batch of documents
7. Test it safely
- · Run a batch containing one known-problem document and one known-good document
8. If it doesn't work
- Findings are less thorough in big batches.
- Use smaller batches (3–5 documents) per task.
Run the task
If your Claude has one unified message box (no Chat/Cowork choice)
- Open the Project, then describe the task in any conversation. Claude decides whether it's a quick answer or a task.
If you see separate “Chat” and “Cowork” options
- In the message box, select “Cowork” in the bottom-left corner, then describe the task.
- Projects work the same way in Cowork and in chat.
Task request
Review each file in [FOLDER/ATTACHMENTS] against REVIEW CRITERIA. Produce one findings table per document, with quotes, and a summary of the top issues across all of them.
Fallback: Review one document at a time in the Project, as described in Build It Once.
By default, Claude asks before taking an action. Keep that on for anything involving money, personal records, or files you can't replace.
How to test it
Success looks like
- ✓ All known problems are flagged with correct quotes.
- ✓ No more than one or two false alarms on the good document.
How to use it next time
- Open the workspace and upload the new document.
- Say: 'Review this. Context: [anything unusual].'
- Check each Issue against its quote before acting on it.
- When you find a missed issue, add it to the criteria file and re-upload (bump the version).
Set up once — reuse
Criteria list · Review instructions · Output table format
Changes each time
The document under review · Any special context for this document
Troubleshooting
- Findings without quotes, or quotes that don't exist.
- Search the document for the quote. If missing, reinforce: 'A finding without a verbatim quote is not allowed.'
- It skips criteria on long documents.
- Ask it to review in sections ('criteria 1–4 first'), or split the file.
- Too many minor flags.
- Add a severity column and ask for 'Issues of severity Medium or higher only' in the summary.
Limitations, privacy & safety
- · This is a first-pass screen, not professional legal, compliance, or editorial sign-off.
- · Scanned PDFs may not be read accurately; use text-based files when possible.
- · Very long documents may exceed what the tool reads at once.
Privacy & safety
- · Contracts and drafts are often confidential. Confirm your organization permits uploading them.
- · Redact personal data (names, addresses, ID numbers) when it isn't needed for the review.