AI Prompting with Co-Audit
How to write instructions for the Co-Audit agent so it builds the right workflow on the first pass. State the goal, name the documents, list the columns, then iterate, with copy-and-paste recipes.
Co-Audit is the natural-language agent in the left sidebar of every request. You describe the workpaper you want and it builds the structure: the client-facing request description, the document groups, the AI Prompt Columns, the Formulas, and the references between them. The better your instruction, the closer the first draft is to done.
This page applies the universal prompting principles to Co-Audit specifically. If you haven't read the overview yet, start there.
What you'll learn
- The three-part shape of a strong Co-Audit instruction: goal, documents, columns.
- How to give Co-Audit a role and the right context.
- How to attach uploaded request documents to the session so Co-Audit can read them.
- How to iterate: refine in place instead of starting over.
- Copy-and-paste instruction recipes for common workpapers.
The shape of a good instruction
Co-Audit reads your instruction the way a new staff auditor would. The strongest instructions answer three questions, in plain English:
1. The goal
What is this workpaper for? What audit area or assertion does it cover?
2. The documents
What will the client upload, and for what period or scope?
3. The columns
What values should be extracted, calculated, or judged on each row?
When all three are present, Co-Audit can build the whole workflow in one shot. When one is missing, it guesses, and you spend the time you saved fixing the guess.
Vague — more rework
Build a credit card testing workpaper.Co-Audit has to guess what documents arrive, what period, and which columns you want.
Complete — less rework
Create a request asking the client to upload all credit card statements
for the period under audit, and build a workpaper that lists, for each
statement: the bank name, the statement period, the starting balance,
the ending balance, and a column flagging any significant or unusual
transactions.Goal, documents, period, and columns are all stated, so Co-Audit builds the full workflow.
Give Co-Audit a role and context
A single sentence of role and context sharpens everything downstream. Tell Co-Audit what kind of audit work this is and what standard or policy applies. It already knows the engagement variables, the documents in the request, and your firm's templates, so point it at the right ones.
Act as a staff auditor testing fixed asset additions for a manufacturing
client. The client will upload purchase invoices and approval forms for
the period under audit. Build a workpaper that traces each addition from
the invoice to its approval.Context is leverage. "For a manufacturing client," "under the plan's eligibility rules," "for a December 31 year-end" each tell Co-Audit how to shape the columns and the client-facing copy without you spelling out every field.
Add documents to the Co-Audit session
The richest context you can give Co-Audit is a real document. When files have already been uploaded to the request, you can attach one or more of them to the Co-Audit session so the agent reads the actual evidence while it builds the workflow. Instead of describing what a lease or a statement looks like, you let Co-Audit see it and shape the columns to match.
Click the + button in the Co-Audit panel
In the Co-Audit panel, click the + button next to the instruction box to open the document picker.
Select the document(s) to add
The picker lists the files already uploaded to the request. Select the one or more documents you want Co-Audit to read, then confirm. They attach to the current session as context.
Reference them in your instruction
Write your instruction as usual. Co-Audit now reads the attached documents directly, so you can point at what's in them: "build the schedule based on the terms in the attached lease," or "use the columns that match the fields on the attached statement."
When this helps most
Attach a document whenever the structure of the workpaper depends on the document's real layout: building an amortization schedule from an actual lease or loan agreement, matching columns to the exact fields on a sample statement, or letting Co-Audit confirm what a document contains before it commits to a set of columns.
Only files already uploaded to the request appear in the picker. If the document you want isn't listed, upload it to the request's Requested Files or Files area first, then add it to the session.
Recipes you can copy and paste
Drop one of these into the Co-Audit panel and adapt the specifics. Most follow the goal → documents → columns shape, and together they show the range of work you can hand to Co-Audit: extraction, tracing, calculation, classification, anomaly detection, summarization, and reconciliation.
Extraction workpaper — bank statements
Showcases the basics: pull a fixed set of values off each document into one row per file.
Create a request for the client to upload all of their bank statements
for the period under audit, and build a workpaper that lists, for each
statement: the bank name, the account number (last four digits only),
the statement period, the beginning balance, the ending balance, and a
column that flags any transactions over materiality as significant.Tracing and tie-out — invoice testing
Showcases pairing extraction with a Testing Attribute that renders a pass/exception judgment.
Build a workpaper to test the expense population. The client will upload
the supporting invoice for each selection. For each invoice, extract the
vendor name, the invoice number, the invoice date, and the invoice total.
Then add a testing attribute that confirms the invoice total supports the
amount recorded in the selection, citing the policy that an invoice must
match the recorded amount within a reasonable rounding tolerance.Calculation schedule — lease amortization
Showcases reading one attached document and generating a row-per-period schedule of computed values.
The client will upload a single lease agreement. Read it and build an
amortization schedule with one row per month over the lease term. For
each month show the period, the beginning liability, the interest
expense, the principal reduction, and the ending liability. Use the
period start and discount rate from the engagement variables.Classification — categorize ledger activity
Showcases having the AI assign each row to a category and explain its reasoning, not just extract a value.
The client will upload the general ledger detail for the repairs and
maintenance account. Build a workpaper with one row per transaction
showing the date, vendor, description, and amount, then add a column
that classifies each transaction as "routine maintenance," "capitalizable
improvement," or "needs review," with a one-sentence rationale for the
classification.Anomaly detection — scan for red flags
Showcases pointing the AI at a population to surface risk, rather than testing a predefined attribute.
The client will upload the full disbursements register for the year.
Build a workpaper listing each disbursement and add columns that flag
potential anomalies: round-dollar amounts, payments dated on a weekend
or holiday, duplicate invoice numbers, and any vendor not appearing
elsewhere in the population. Summarize why each flagged item stands out.Summarization — contract abstract
Showcases turning a long, unstructured document into a structured plain-English summary.
The client will upload signed customer contracts. For each contract,
read it and produce a summary workpaper with the counterparty, the
contract term, the total contract value, the payment schedule, any
renewal or termination clauses, and a plain-English note on how revenue
should be recognized under the stated terms.Reconciliation — three-way match
Showcases tying values across multiple document types in the same row.
For each selection the client will upload the purchase order, the
receiving report, and the vendor invoice. Build a workpaper that extracts
the quantity and amount from each of the three documents, then add a
testing attribute that concludes whether the purchase order, receiving
report, and invoice agree, raising an exception that names any document
that disagrees and by how much.Confirmation testing — compare replies to the books
Showcases comparing an externally sourced figure to a client-provided value and judging the difference.
The client will upload the returned accounts receivable confirmations.
For each confirmation, extract the customer name, the balance confirmed
by the customer, and the confirmation date. Add a column with the balance
per the client's records for that customer, a formula for the difference,
and a testing attribute that concludes whether the confirmation supports
the recorded balance or requires follow-up.Iterate — add to an existing workpaper
Showcases extending a workflow you've already built instead of starting over.
Add a column that recalculates the total as unit price times quantity,
and a second column showing the variance between that recalculation and
the invoice total the client provided.Use Formulas for math
When you ask Co-Audit for a calculation (a recalculation, an interest figure, a variance, a footed total), it should build a Formula column, not an AI Prompt. If it builds an AI Prompt instead, tell it: "make that a formula so it recomputes." Formulas always recompute; AI prompts sometimes copy a value forward.
Iterate in place, don't start over
Co-Audit edits the workflow you already have. When something needs to change, describe the change instead of rebuilding from scratch. The panel stays open as you work, so you can ask for a change and watch the Workflow tab update next to it.
Describe the change in plain English
The client provided an annual report instead of monthly statements.
Update the procedures to read the annual report, and change the schedule
to one row per quarter instead of one per month.Submit and watch the Workflow tab
Co-Audit edits the existing document groups, columns, and references in place rather than starting a new workflow.
Spot-check the result
Co-Audit is fast and accurate at structure but still makes mistakes. Watch for red text in a custom column (usually a broken @ reference from a column-name mismatch) and double-check any column that should recalculate a number.
Common fixes
Continue your journey
← AI Prompting overview
The universal principles behind every prompt in Punchcard.
AI Prompting for Workpapers →
Write the AI Prompt Columns and Testing Attributes that Co-Audit generates, by hand or to refine them.
How is this guide?
AI Prompting
How to write prompts that get better results from Punchcard's AI. Universal best practices for Co-Audit, AI Prompt Columns, and Testing Attributes, with copy-and-paste recipes you can adapt.
AI Prompting for Workpapers
Write AI Prompt Columns and Testing Attributes that return clean, citable values and sound audit judgments. Format specs, @ references, grounding, and copy-and-paste prompt recipes.