Ad hoc prompting for data interpretation works until it does not. You get a good answer, then a confidently wrong one, and you cannot quite say what changed. The fix is a repeatable model—a named set of stages you apply in order—so that good results come from process rather than luck. This article introduces one such model, called LENS, designed for anyone who interprets tables and charts with AI often enough that consistency matters.
LENS stands for Label, Examine, Narrow, Substantiate. The four stages map onto the natural arc of a data question: establish what you are looking at, understand its shape, ask something specific, and verify the result. The value of naming them is that you can tell, at any moment, which stage you are in and which you skipped when something goes wrong.
A framework is not a script to recite. It is a mental model that makes good moves automatic and bad omissions visible. By the end of this article you will be able to apply LENS to almost any table or chart and to diagnose, after a bad answer, exactly which stage failed.
L — Label: Establish What the Data Is
The first stage is telling the model, and yourself, what the data represents before any analysis begins.
What happens here
You state the rows, the columns, the units, the time period, and any rules of the data (multiple-select, percentages of what, currency scale). "This table shows weekly active users by region for 2024; counts are whole numbers." This single sentence removes the guesswork that causes most misinterpretation.
When it matters most
Always, but especially when the data is unfamiliar or the headers are terse. Skipping Label is the root cause of magnitude and category errors, the exact failures catalogued in the common mistakes piece.
E — Examine: Understand the Data's Shape
Before asking a pointed question, have the model confirm it reads the structure correctly.
What happens here
Ask the model to restate the table: how many rows and columns, what each column measures, and the range of values. For a chart image, have it confirm the axis scale and what each series represents. This is a cheap diagnostic that catches structural misreads before they contaminate an answer.
When it matters most
When the data is messy, large, or provided as an image. Examine is where you catch a truncated axis or a misaligned column while it is still harmless, the same early catch that anchors the step-by-step process.
N — Narrow: Ask Something Specific and Checkable
With the data labeled and its shape confirmed, ask a question that has a verifiable answer.
What happens here
You pose a question whose answer is a number, a named cell, or a computed difference—not an open-ended summary. "Which region had the most active users in Q3, and what was the count?" Define any ambiguous terms in the question itself so the model has a precise target.
When it matters most
On every analytical question, but especially before you ask for any summary. Narrowing first confirms the model reads the table correctly, which is why the best practices guide insists on checkable questions before comprehensive ones.
S — Substantiate: Verify and Ground the Answer
The final stage is making the model show its work and then checking it.
What happens here
Require the cells used and the formula for any calculation, then verify: spot-check cited values, recompute the headline number, confirm units and scale, and ask what the model estimated. For trends, demand quantified change rather than narrative. Only after Substantiate do you trust the answer enough to act on it.
When it matters most
Always for anything that drives a decision or reaches a client. Substantiate is the stage people skip and later regret, and it is the heart of the data prompting checklist.
Applying LENS End to End
The stages are designed to run in sequence, but the real power is using them to diagnose failures after the fact.
Using the model to debug
- Wrong category or magnitude? You skipped or rushed Label.
- Misread structure or axis? Examine was thin.
- Vague, unverifiable answer? Narrow was too broad.
- Confident but unchecked? You stopped before Substantiate.
This diagnostic ability is the practical payoff of a named framework. Instead of vaguely sensing that something went wrong, you can point to the stage that failed and fix it. It also makes the practice teachable—new team members can learn the four stages and apply them consistently, which is how the team in the data prompting case study standardized their workflow.
A Worked Pass Through LENS
Seeing the stages applied to a concrete question makes the model click. Take a table of monthly active users by region for 2024.
The four stages in action
- Label: "This table shows monthly active users by region for 2024; values are whole-number counts, no scaling." The model now knows exactly what it is reading.
- Examine: "Restate the table—how many regions, how many months, and the approximate range of values." The model confirms it parsed twelve months across four regions, catching any structural misread early.
- Narrow: "Which region had the highest active users in any single month of 2024, and what was that count?" A single, checkable answer.
- Substantiate: "Show the cell you used," then verify it against the source and sanity-check the magnitude. Only now is the answer trustworthy.
Each stage took seconds, and the discipline turned a question that could have produced a confident wrong answer into one you can defend. The same sequence underlies the step-by-step process, expressed there as numbered steps rather than named stages.
Where LENS Fits Alongside Other Tools
A framework is a way of thinking; it works best paired with a concrete checklist and a set of practices. Knowing how they relate keeps you from treating any one as the whole answer.
The division of labor
- LENS tells you the order and purpose of each move, and lets you diagnose which stage failed.
- A checklist gives you the specific items to verify within each stage, especially during Substantiate.
- Best practices supply the reasoning behind individual moves, like why trends must be quantified.
Used together, the framework structures your thinking, the data prompting checklist operationalizes the verification, and the best practices guide explains the why. LENS is the scaffold the other two hang on, which is what makes it worth learning first.
Frequently Asked Questions
Do I have to run all four stages every time?
For low-stakes glances you can compress Label and Examine into a quick read and lean on Narrow and Substantiate. For anything client-facing or decision-driving, run all four. The stages are most valuable as a diagnostic: even when you compress them, knowing which one you skipped explains most failures.
How is LENS different from a checklist?
A checklist is a flat list of items to verify; LENS is a staged model that tells you the order and the purpose of each step. The framework is better for diagnosing why an answer failed, because each failure maps to a specific stage. The two complement each other—use LENS to think, a checklist to verify.
Which stage do people skip most?
Substantiate, the verification stage, because the answer already looks polished and checking feels redundant. That is exactly the trap: polish is unrelated to correctness. Examine is a close second, since people jump straight to their question without confirming the model read the structure.
Can LENS handle chart images as well as tables?
Yes. For images, Label includes stating the axis scale, Examine includes having the model confirm what each series and axis represents, and Substantiate includes asking for ranges rather than precise figures. The stages are the same; the specifics adapt to the medium.
How do I introduce LENS to a team?
Teach the four stages and the failure each prevents, then require it for client-facing analysis. Because every failure maps to a named stage, the framework is easy to learn and easy to enforce in review—you can point to which stage was skipped rather than debating the answer in the abstract.
Key Takeaways
- LENS—Label, Examine, Narrow, Substantiate—is a four-stage model for interpreting tables and charts with AI.
- Label establishes what the data is; skipping it causes magnitude and category errors.
- Examine confirms the model reads the structure correctly, catching axis and column misreads early.
- Narrow asks a specific, checkable question before any summary; Substantiate verifies and grounds the answer.
- The framework's real payoff is diagnostic: each failure maps to a stage, making errors easy to locate, fix, and teach.