A checklist is only useful if you actually run it, which means it has to be short enough to use under time pressure and clear enough that each item earns its place. This one is built for exactly that: a working tool you can run before trusting any AI interpretation of a table or chart, with a brief justification for every check so nothing feels like ceremony.
It is organized into three phases—before you prompt, while you prompt, and after you get an answer. The before phase prevents the most errors, the during phase shapes a checkable answer, and the after phase catches whatever slipped through. You will not need every item every time, but the high-stakes items are flagged so you know what is non-negotiable.
Keep it nearby the first dozen times. Most of these checks compress into habit quickly, and once they do you will run the whole thing in a couple of minutes. The point is not bureaucracy; it is catching the confident, well-formatted errors that otherwise reach your reports unchallenged.
Phase 1: Before You Prompt
Most interpretation errors are decided here, before the model sees anything. These checks are the highest-leverage on the list.
The pre-prompt checks
- Is the data clean? Aligned columns, no merged cells, decorative rows removed. Ambiguous structure invites wrong-column errors.
- Does every column have a clear header with units? Implied units cause magnitude errors that look tidy but are off by orders of magnitude.
- Do you have the underlying data, or only an image? Text gives exact reading; images give estimates. Know which you are working with.
- For images, what is the axis scale? A truncated or logarithmic axis can invert the entire interpretation.
These mirror the input-quality emphasis of the field guide, which treats clean data as the real lever.
Phase 2: While You Prompt
These checks shape a prompt that produces a verifiable answer rather than a vague impression.
The during-prompt checks
- Did you state what the data represents? One sentence of context eliminates a class of guesses.
- Is your question specific and checkable? A named cell or computed value can be verified; a summary cannot.
- Did you define any ambiguous terms? "Sales" could mean units or revenue—say which.
- Did you ask the model to show its work? Cited cells and formulas turn a black box into an audit trail.
This sequence is the same discipline laid out in the step-by-step process, condensed into checks.
Phase 3: After You Get the Answer
The verification phase. Skipping it is the single most common mistake, so treat these as non-negotiable for anything that matters.
The post-answer checks
- Spot-check two or three cited values against the source. Confirms the model read the right cells.
- Recompute the headline metric independently. Catches arithmetic and wrong-cell errors on the number that matters most.
- Confirm units and scale. Guards against the costliest, easiest-to-miss error.
- Did the model flag what it estimated? Estimates from images deserve extra scrutiny.
- Does the answer's size make sense? A sanity check on magnitude catches gross errors instantly.
These are the checks that saved the team in the data prompting case study and that the best practices guide insists you never skip.
Calibrating the Checklist to the Stakes
Not every question needs the full list. The right amount of rigor depends on the consequence of being wrong.
How to scale it
- Quick, low-stakes glance: clean the data, ask a checkable question, sanity-check the size. Done.
- Client-facing or decision-driving: run the full list, no exceptions.
- High-volume automated use: encode the checks into your workflow so they run every time.
The mistake is treating a high-stakes question with low-stakes rigor. Match the effort to what a wrong answer would cost, and keep the post-answer sanity check even at the lightest level.
Turning the Checklist Into a Habit
A checklist you have to consciously remember is a checklist you will eventually skip. The goal is to internalize it so it runs almost automatically.
How the items become reflexes
The fastest way to absorb the list is to run it deliberately on a dozen real tables, justifications and all. After that, the high-leverage items—clean headers, stated units, shown work, recomputed headline—fire without conscious effort, and the list becomes a backstop you consult only when something feels off. This is the same progression beginners follow in the for-beginners guide: deliberate at first, automatic later.
Keep the list visible until then
Until the reflexes form, keep the checklist where you actually work—pinned in your notes or your team's shared space. Visibility is what makes a checklist get used under deadline pressure, which is exactly when skipping it does the most damage.
Embedding the Checklist in a Team Workflow
Individual use is valuable; making the checklist a team standard is where the real reliability comes from, because quality stops depending on who did the work.
Practical ways to embed it
- In review: require that any client-facing data answer show its cited cells, so a reviewer can run the post-answer checks against something concrete.
- In templates: bake the context sentence, explicit units, and the request for shown work into reusable prompts.
- In automation: for high-volume work, turn the post-answer checks into code—validate structure, require cited cells, and add an automated recomputation of key figures.
A team that shares this checklist catches errors the way the team in the data prompting case study did—routinely, cheaply, and before anything reached a client. The framework behind these checks, including which stage each one belongs to, is laid out in the data interpretation framework.
When to Update the Checklist
A checklist is a living tool. The items that matter most depend on the data you actually work with, so revisit it periodically.
Tuning it to your data
If your tables always carry implied units, promote that check to the top. If you rarely use images, you can deprioritize the scale check. After every error that slips through, ask which check would have caught it and whether that check needs to be more prominent. Over time the list converges on the failures your specific work is prone to, the same source-fixing instinct the common mistakes piece recommends.
Frequently Asked Questions
Which phase prevents the most errors?
The before-you-prompt phase. Clean data with clear headers and known scale prevents wrong-column reads, magnitude errors, and axis misinterpretations—the most common and costly failures. Fixing input quality up front does more than any amount of clever prompting or after-the-fact checking.
Do I really need to run this every time?
No. Calibrate to the stakes: a quick glance at a small, clear table needs only a few checks, while a client-facing or decision-driving answer warrants the full list. The one check to keep even at the lightest level is a sanity check on whether the answer's size makes sense.
What is the single most important post-answer check?
Recomputing the headline metric independently. It catches arithmetic mistakes and wrong-cell reads on exactly the number that will drive a decision or appear in a report. Combined with a quick units-and-scale confirmation, it covers the errors that cause the most damage.
How do I use this for automated, high-volume work?
Encode the checks into the workflow itself: validate input structure, require the model to return cited cells, and add an automated recomputation or range check on key figures. When volume is high, manual checking does not scale, so the checklist becomes part of the system rather than a human ritual.
Why include a justification for each item?
Because a check without a reason gets skipped under pressure. Knowing that implied units cause order-of-magnitude errors, or that truncated axes invert interpretations, makes you actually run the check. The justifications turn the list from ceremony into something you trust.
Key Takeaways
- The checklist has three phases—before, during, and after prompting—and the before phase prevents the most errors.
- Clean data, clear headers with units, and known axis scale are the highest-leverage pre-prompt checks.
- A checkable question plus shown work turns the answer into something you can verify in seconds.
- The non-negotiable post-answer checks are recomputing the headline metric and confirming units and scale.
- Calibrate rigor to the stakes, but always keep a sanity check on whether the answer's size makes sense.