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.
Punchcard runs on AI in three places: the Co-Audit agent that builds your workflows, the AI Prompt Columns that extract values from documents, and the Testing Attributes that evaluate audit judgments. All three are driven by the prompts you write. The quality of what you get back is mostly a function of the quality of what you put in.
This page is the foundation. It covers the prompting principles that hold everywhere in Punchcard, distilled from Anthropic's prompt engineering best practices and adapted to audit work. The two pages that follow apply these principles to the specific surfaces.
What you'll learn
- The mindset that makes Punchcard's AI accurate: treat it like a brilliant new staff auditor.
- Seven prompting principles that apply to Co-Audit, AI Prompt Columns, and Testing Attributes.
- Copy-and-paste recipes you can drop into a prompt and adapt.
- Where to go next for Co-Audit and Workpaper prompting.
The core mindset
Think of Punchcard's AI as a brilliant but brand-new staff auditor. It is fast, capable, and reads every page of every document, but it has no context on your firm's norms, this engagement's specifics, or what "good" looks like unless you tell it. The more precisely you describe the task, the better the result.
The golden rule
Show your prompt to a colleague who has minimal context on the task and ask them to follow it. If they would be confused or have to guess, the AI will be too. A prompt that a new hire could execute without asking questions is a prompt Punchcard can execute well.
Seven principles that apply everywhere
Be clear and direct
State exactly what you want. Vague instructions force the AI to guess, and guesses are where errors come from. Specify the output, the format, and any constraints up front.
Less effective
Summarize this invoice.More effective
Return the total amount due shown on the invoice as a number with two
decimal places. If multiple totals appear, return the final amount due
after taxes and adjustments.Add context and motivation
Explaining why you want something helps the AI generalize to edge cases you didn't spell out. A reason is worth more than another rule.
Less effective
Ignore handwritten notes.More effective
Extract only the typed, printed total from the statement. Ignore
handwritten annotations, since those are the client's working notes and
not part of the official record.Specify the format
Tell the AI the exact shape of the answer you want: a number, a date format, a Yes/No, a single word. Unspecified formats come back inconsistent and break downstream formulas.
Return the statement period as a date range in MM/DD/YYYY – MM/DD/YYYY
format. If only a single statement date is shown, return that date in
MM/DD/YYYY format.Tell it what to do, not what not to do
Positive instructions outperform prohibitions. Instead of listing everything to avoid, describe the one correct behavior.
Less effective
Don't include the vendor's address or phone number.More effective
Return only the vendor's legal entity name as it appears in the header.Ground answers in the document
Punchcard's AI works against the actual source documents and returns a citation pill for every value. Lean into that. Ask it to base its answer on what the document says and to flag when the document doesn't support an answer, rather than filling a gap with a plausible guess.
Based only on the lease agreement, return the monthly base rent. If the
agreement does not state a monthly base rent, return "Not stated" rather
than inferring it from other figures.Iterate and self-check
Your first prompt is a draft. Run it, click the citation pill to see where the value came from, and refine the wording on the column header. For judgment columns especially, add a verification step.
After determining your answer, re-read the relevant section of the
document and confirm the figure matches before responding.Validated cells are locked. You can tune a prompt and rerun a column as many times as you need without overwriting work an auditor has already signed off on. Iterate freely.
Use examples when format matters
A few well-chosen examples (sometimes called few-shot prompting) are one of the most reliable ways to lock in a format or handle a tricky edge case. When a value comes in many shapes across documents, show the AI the shape you want.
Return the check number from the document. Examples of the format:
- "Check #1042" → 1042
- "CHK 000587" → 587
- "No. 12,330" → 12330
Return the number only, with no prefix, leading zeros, or punctuation.Keep examples relevant (mirror your real documents) and diverse (cover the edge cases that trip up a single rule). Two or three good examples usually beat a long paragraph of instructions.
Anti-patterns to avoid
Using an AI Prompt for math
For any calculation (recomputing a payment, footing a column, a variance tie-out) use a Formula column, not an AI Prompt. AI sometimes copies a value forward instead of recomputing it. Formulas always recompute. See The Workflow tab.
Typing @ references as plain text
@ references are not plain text. You have to type @, narrow the list, and select the item so it becomes a blue pill. Text that merely looks like @Invoice Number will not resolve at run time.
Listing everything to avoid
Long lists of "do not" rules are harder to follow than a single clear description of the correct output. Lead with what you want.
Where to go next
AI Prompting with Co-Audit →
Prompt the natural-language agent that builds your workflows. State the goal, the documents, and the columns, then iterate.
AI Prompting for Workpapers →
Write AI Prompt Columns and Testing Attributes that extract clean, citable values and sound audit judgments.
How is this guide?