Overview

What you're building

Two stages. Build once: a reference spreadsheet of past records, each labeled by type with its scope, amount, date, and source, plus written adjustment rules. Use each time: describe the new request, and your AI workspace finds the most comparable past records, applies your rules, and drafts a preliminary estimate showing which records it used, what it assumed, and how uncertain it is. You review it before anyone sees it.

Why this setup

AI chats don't keep a reliable database of their own, so the spreadsheet is your source of truth and does the math. The AI's job is the judgment-heavy part: reading the new request, finding comparable records, and explaining the comparison. Invoices alone often leave out scope, hours, exclusions, and dates, so the kit has you capture those once, which is what makes later comparisons meaningful.

The workflow architecture

• AI workspace

  Holds your instructions and reference files so you never re-explain.

• Spreadsheet or tracker

  Reference spreadsheet: one row per past record, labeled by type, with formulas for any calculation.

• Reference document

  Adjustment rules you approve (e.g. inflation, size, rush), written in plain sentences.

• Reusable instructions

  Estimating instructions: how to pick comparables, apply rules, and flag uncertainty.

• Document template

  Estimate layout with sources, assumptions, and a range.

• Your review

  A human check before the estimate is shared or relied on.

You haven't picked a tool yet, so this kit is platform-neutral. Both Claude and ChatGPT support this setup.

If you estimate often, the Estimating Instructions can later become a Claude Skill so the steps load automatically. Check Claude's current documentation for Skills availability on your plan before relying on it.

What to prepare

• □Your past records (invoices, bills, timesheets, budgets), as many as you can gather, ideally 10 or more.
• □The labels you want to group them by (project type, size, location, client type).
• □Any details the records leave out: what was in scope, hours, exclusions, and anything unusual about the job.
• □How you adjust today: inflation, size, rush, complexity. Write down only rules you'd actually stand behind.

Build it once

1. 1Create a dedicated workspace named "Historical Reference & Estimation" in the AI tool you choose. Both Claude and ChatGPT call this a "Project". Claude offers Projects on all plans; ChatGPT availability depends on your plan.
2. 2Paste the main instructions block from this kit into the workspace's instructions.
3. 3Upload the reference files listed below, and start every future conversation from inside this workspace.
4. 4Build once: create a spreadsheet from the Reference Table template. One row per past record; fill in what you know and mark gaps as UNKNOWN.
5. 5Build once: optionally paste a few invoices into a chat with the Record Extraction prompt to speed up data entry, then check each row against the original.
6. 6Build once: agree on your categories (e.g. project types) and use the same labels every time.
7. 7Build once: write your adjustment rules using the Adjustment Rules template. Put the math in spreadsheet formulas, not in the chat.
8. 8Build once: create a Project or saved workspace, add an export of the spreadsheet and the rules as files, and paste the Estimating Instructions.
9. 9Keep it current: when a job finishes, add its row and re-upload the export so the workspace isn't working from old data.

Copy-ready materials

Edit anything in [BRACKETS]. Changes are kept when you save the kit.

REFERENCE TABLE  (last updated: [DATE])
Record ID | Type / Category | Short scope description | Size or quantity | Amount (cost, hours, or units) | Date completed | Source (invoice #, file name) | Included | Excluded | Unusual factors | Data quality (Complete / Partial)

Rules:
- One row per past record. Use the same Type labels every time.
- Write UNKNOWN rather than guessing missing details.
- Keep amounts as numbers so formulas work.

I'll paste past [INVOICES / BILLS / RECORDS]. For each one, return one row using these columns:
Record ID | Type / Category | Short scope description | Size or quantity | Amount | Date | Source | Included | Excluded | Unusual factors | Data quality

- Use only what is written in the record. Write UNKNOWN for anything not stated.
- Suggest a Type from this list only: [YOUR CATEGORIES]. If none fits, write NEEDS LABEL.
- Do not total, adjust, or estimate anything.

ADJUSTMENT RULES  (approved by: [NAME], [DATE])
- Older records: [e.g. add X% per year since completion]
- Size: [e.g. scale by quantity / square footage]
- Rush or short timeline: [ ]
- Complexity factors: [what counts and how much]
- Minimum comparable records needed for an estimate: [e.g. 3]
- Records to never use: [e.g. discounted favors, unusual one-offs]

When I describe a new request:
1. Restate the request: type, scope, size, timing. List anything I didn't say that matters as QUESTIONS.
2. Find the most comparable records in the REFERENCE TABLE file. Use the same Type first, then the closest scope and size. List each with its Record ID and why it's comparable, and note differences.
3. If fewer than [MINIMUM] comparable records exist, or they are Partial quality, say "NOT ENOUGH COMPARABLE RECORDS" and explain what's missing. Do not produce a confident number.
4. Apply ONLY the ADJUSTMENT RULES file. Name each rule you used. Do not invent new adjustments.
5. Give a range (low / likely / high), not a single figure, using the ESTIMATE LAYOUT.
6. This is a preliminary estimate for me to review. Never present it as a final price or commitment.

PRELIMINARY ESTIMATE: [request in one line]  —  FOR INTERNAL REVIEW

COMPARABLE RECORDS
[Record ID] | [Type] | [Amount] | [Date] | Why comparable | Key differences

ADJUSTMENTS APPLIED
[Rule] → [effect]

ESTIMATE RANGE
Low: [ ]  Likely: [ ]  High: [ ]
(Check these figures in your spreadsheet.)

ASSUMPTIONS
- [ ]

UNCERTAINTY: Low / Medium / High because [reason]

QUESTIONS BEFORE SHARING
- [ ]

□ Each comparable record exists in the spreadsheet and the amount matches.
□ The comparables really are similar in scope, not just the same label.
□ Only approved adjustment rules were used; totals checked in the spreadsheet.
□ Assumptions and exclusions are stated to the client.
□ It is labeled preliminary / an estimate, not a fixed price.

How to test it

1. 1Pick a finished job that is in your spreadsheet, hide its row, and ask for an estimate. Compare with what it actually cost or took.
2. 2Ask for an estimate for a type you have no records for and confirm it says NOT ENOUGH COMPARABLE RECORDS.
3. 3Check that every adjustment it lists is one of your written rules.

Success looks like

• ✓ Every figure traces to a Record ID in your spreadsheet.
• ✓ It gives a range with assumptions, not a single confident number.
• ✓ It declines when the history is too thin.

How to use it next time

1. Open the workspace and describe the new request (type, scope, size, timing).
2. Read the comparable records and check they're really similar.
3. Confirm the range in your spreadsheet and answer the open questions.
4. Share the estimate yourself, labeled as preliminary.
5. When the job is done, add its actual figures as a new row.

Troubleshooting

It picks records that share a label but aren't really similar.

Add more scope detail to rows, and add to the instructions: "Explain the key differences for each comparable."

The math is off.

Do calculations in spreadsheet formulas and treat the AI's figures as a pointer to check.

It doesn't know about recent jobs.

Re-upload a fresh export of the spreadsheet; the workspace only sees the files you've added.

It invents an adjustment (e.g. a market factor).

Repeat rule 4 and remove anything it adds that isn't in your rules file.

Limitations, privacy & safety

• · Invoices often leave out scope, hours, exclusions, and dates. Without those, comparisons are rough, so capture them in the spreadsheet.
• · The AI has no persistent database. It only sees the spreadsheet export and files you add to the workspace.
• · An estimate is only as good as the comparable records. With few or unusual records, it should tell you it can't estimate reliably.
• · Estimates are preliminary. A person reviews them before they're used for pricing, budgets, or commitments.

Privacy & safety

• · Past invoices may include client names and pricing. Remove details you don't need for comparison and share the workspace only with people who should see it.