Best Practices · Optional
Verify Everything
Why fluent output still needs checking, and a practical routine for verifying Claude's work before it leaves your desk.
Claude writes with confidence whether or not it is right. That is the single most important thing to internalize about working with it. The output reads like a competent colleague produced it, the figures are formatted neatly, the tone is assured, and none of that tells you whether the numbers are correct. Verification is not a sign that you distrust the tool. It is the professional discipline that lets you use the tool at all.
If you take one habit from this whole guide, take this one: never let an unverified figure, quote, or claim reach a deliverable. Everything else is detail.
Why fluent output can still be wrong
A model produces the most plausible next words given what it has seen. Most of the time that plausibility lines up with the truth, which is exactly why Claude is useful. But plausibility and accuracy are not the same thing. When a document is vague, when a figure is buried, or when you ask about something the source does not actually say, Claude can fill the gap with something that sounds right and is not. This is usually called a hallucination, and the dangerous version is not the obvious mistake. It is the small, confident, specific-looking number that slots neatly into your memo.
Build the check into the prompt
The first line of defense costs you nothing: ask Claude to mark its own weak spots. The scrutiny instruction from Anatomy of a Good Prompt is not optional polish, it is the part that surfaces what you would otherwise have to hunt for yourself. A model that has been told to flag gaps will point you straight at the sentences worth checking.
For every figure, claim, or conclusion in your answer, tell me where in the source it comes from (the section, page, or table). If something is your inference rather than stated in the document, label it clearly as an inference. If the document does not contain the information, say so plainly rather than estimating.
This turns a wall of confident prose into something auditable. Now each claim carries its own pointer back to the source, and your job shifts from re-deriving everything to spot-checking the pointers. The ready-made versions of prompts like this live in the Prompt Library.
A verification routine you can repeat
Verification is faster when it is a routine rather than a vague intention to be careful. Run these five passes, in order, on anything analytical before it goes near a client deliverable or an investment committee paper.
-
Check every number against the source
Take each figure Claude reports (revenue, EBITDA, growth, multiples) and find it in the original document. Not a similar number nearby: the exact figure, in context. Most errors hide in numbers that were transcribed, combined, or restated.
-
Trace each quote and claim
For any direct quote or specific factual claim, confirm the source actually says it. Watch for claims that are true in spirit but overstated in wording, and for conclusions presented as fact that are really Claude’s reading.
-
Separate what was stated from what was inferred
Decide which parts came from the document and which Claude reasoned into existence. Inferences can be valuable, but they are yours to accept or reject, not facts to repeat. Mark them as judgment, not source.
-
Look for what is missing
A fluent summary can quietly omit an inconvenient risk or a caveat in a footnote. Skim the source for material points the answer left out, especially anything that would change the conclusion.
-
Apply your own judgment last
Finally, read the output as the professional you are. Does the conclusion make commercial sense? Does it match what you already know about the sector and the target? Claude supplies the draft; the judgment stays with you.
How much to verify
Not every task carries the same weight, and pretending otherwise just makes verification feel like a tax you start skipping. Match the rigor to the stakes.
Match the check to the risk
| Task | Risk if wrong | How to verify |
|---|---|---|
| Reformatting text you wrote | Low | A quick read for sense |
| Summarizing a teaser for an initial screen | Medium | Spot-check headline figures and the stated thesis |
| Extracting financials for a model | High | Check every figure against the source, line by line |
| A claim going into an IC paper or client deliverable | Very high | Full trace: every number, quote, and conclusion to source |
The higher the consequence of an error, the more of the source you check by hand.
The pattern is simple. Drafting and reshaping are low-risk because you can see the source yourself. Anything where Claude is the bridge between a document and a decision is where you slow down.
A worked example
Suppose you ask Claude to pull the headline financials from the Project Atlas information pack, the mid-market B2B field-service software business growing in the mid-20s percent. It returns revenue of around $60m, EBITDA margin in the low twenties, and net retention “above 110 percent”. The first two you can find on the summary page in seconds. The retention figure is the one to chase: it is precise, it is attractive, and it may have come from a single management slide, a blended cohort, or Claude’s own rounding. You do not assume it is wrong. You find exactly where it came from before it informs a single line of your thesis.
What to take from this
Claude makes you faster by giving you a strong first draft and doing the heavy lifting of synthesis. It does not, and should not, make the final judgment. Keep the scrutiny instruction on every analytical prompt, ask Claude to point at its own sources, and run the five-pass check before anything matters. Used that way, the speed is real and the risk stays controlled. For the wider set of habits that go with this, see What Not to Do and the one-page Cheat Sheet.