Reviewing What Claude Code Did

Good instructions are half the job

You've learned how to give clear, specific instructions. That's the first half of working with Claude Code. The second half is checking what comes back.

Claude Code is good at producing results that look right. Clean formatting. Confident language. Well-organized output. But looking right and being right are not the same thing.

Get into the habit of reviewing what Claude Code gives you before you act on it. Not because the tool is unreliable, but because even your best colleague deserves a second pair of eyes.

How Claude Code shows its work

When Claude Code finishes a task, it doesn't just say "done." It shows you what happened. What you see depends on what you asked it to do.

When it answers a question, you get a plain text response. Ask "how many rows are in this spreadsheet?" and you get a number with a brief explanation. Ask "what are the top five customers by revenue?" and you get a table or list.

When it changes a file, you see a summary of what changed. Claude Code shows which files it touched, what it added, and what it removed. For text files, it highlights the specific lines that are different — green for new content, red for removed content. These are called diffs, and they're how you can see exactly what changed without reading the entire file.

You don't need to understand every line of a diff. Focus on the parts you care about: did the right file change? Do the new values look correct? Was anything removed that shouldn't have been?

When it runs a command, you see the command itself and its output. If Claude Code converted a batch of files, you'll see which files were processed and whether any errors came up. If it analyzed data, you'll see the results printed out.

When it creates something new, like a report or a web page, it tells you the file name and where it saved it. You can then open that file to check the result directly.

Tip: You don't need to understand the technical details. Focus on the result: is it what you asked for? Does it make sense?

The trust-then-verify gap

AI output looks polished. Well-formatted, confidently stated, reads like it was written by someone who knows what they're doing. Language models are built to produce fluent, authoritative-sounding text.

Fluent doesn't always mean accurate.

AI tools can produce results that "appear plausible but contain fabricated or inaccurate information," to use the academic phrasing. People — even experienced ones — tend to trust results that sound confident, without checking whether they're actually correct. There's something about clean formatting and a declarative tone that switches off the critical-thinking reflex.

In practice, this shows up as:

• Numbers that are close but not exact. Claude Code might report "427 rows" when the actual count is 431. The number looks reasonable, so you don't question it.
• Summaries that quietly drop something. A summary of a report might read well but leave out the one paragraph that changes the conclusion.
• Misread data. Claude Code might treat a "date modified" column as a "date created" column. The results look right, but they're based on the wrong thing.
• Confident guesses. Claude Code doesn't always flag uncertainty. If a spreadsheet has ambiguous column headers, it picks an interpretation and moves forward.

None of these are disasters. But if you forward a report to your manager with slightly wrong numbers, or make a decision based on a misread dataset, the consequences land on you — not on the tool.

The gap between "this looks right" and "this is right" is yours to close.

What to check, and how

Checking doesn't mean auditing every line. It means building a few quick habits that catch the most common issues.

Does the result match what you asked for?

Start with the obvious. Read what Claude Code produced and compare it to what you requested.

If you asked for a summary of Q4 data and the output includes Q3 numbers, something went wrong. If you asked for a table sorted by revenue and it's sorted by date, that's a miss.

This is the check people skip most often. The result looks professional, the brain says "close enough," and you're already on to the next thing.

Do the numbers pass the smell test?

You know your business. If the monthly revenue total looks wildly different from what you'd expect, pause.

You don't need to recalculate everything. Just ask yourself: does this number make sense? Is it in the right ballpark? If your company typically does $200K per month and the report says $2M, that's worth questioning even if you don't know exactly where the error is.

Spot-check a few specific data points. Pick three to five numbers from the output and verify them against the original source. Open the spreadsheet, find those rows, and confirm they match.

Did it change anything unexpected?

When Claude Code edits files, scan the summary for surprises. If you asked it to update phone numbers in a contact list, check that it didn't also rename columns, reorder rows, or delete any records in the process.

Claude Code usually does exactly what you ask. But occasionally it "helps" by tidying up things you didn't mention. If you see changes you didn't request, ask about them before moving on.

Open the actual file

This is the most reliable check: look at the output file directly.

If Claude Code created a report, open it. If it cleaned up a spreadsheet, open the spreadsheet. Don't rely only on the summary Claude Code shows you in the terminal — look at the actual file.

For a data file, scroll through a few sections. For a document, read the introduction and conclusion. For a web page, open it in your browser.

This takes an extra 30 seconds, and it catches things that summaries miss.

When Claude Code gets numbers wrong

Data is where careful checking matters most.

Claude Code works with your actual files, so it's usually more accurate than chat-based AI tools that truncate or sample large datasets. But "more accurate" is not "always right."

A few common data issues worth watching for:

• Column misreading. If your spreadsheet has a column called "Amount," Claude Code might not know whether that's revenue, cost, or something else. Check that it read your columns the way you intended.
• Missing rows. If a file has formatting quirks or hidden rows, Claude Code might skip some records. Compare the row count in the result to what you know is in the original.
• Rounding and blank cells. When Claude Code calculates averages or totals, it makes choices about rounding and how to handle empty cells. Those choices might not match yours.
• Dates. Date columns are a common source of quiet errors. "January" in your data could shift to "December" if time zones change during conversion.

These don't happen all the time. But they happen often enough that a 30-second spot check is worth the effort.

Tip: When you need to verify data results, ask Claude Code: "Show me the first 10 rows you used for this calculation." Comparing those rows to the original file is the fastest way to catch interpretation errors.

Building the review habit

This doesn't need to be a formal process. More like a reflex — a quick scan before you move on.

Here's a pattern that takes less than a minute:

1. Read the summary. Does it describe what you asked for?
2. Check the scope. Did it touch only the files or data you intended?
3. Spot-check three values. Pick a few numbers and compare to the original.
4. Open the file. Look at the actual result, not the terminal summary.
5. The manager test. Would you send this to your manager right now? If something feels off, dig in before moving on.

Over time, this becomes automatic. You'll develop a sense for when something needs a closer look and when it's clearly fine.

The goal isn't to distrust Claude Code. It's to trust yourself to catch the occasional issue before it matters.

Asking Claude Code to double-check itself

You can ask Claude Code to verify its own work. After it produces a result, try something like:

• "Can you double-check that total? Walk me through the calculation."
• "Are there any rows in the original data that didn't make it into this summary?"
• "Show me where in the file you got that number."

Claude Code will go back through its steps and often catch its own mistakes. This isn't foolproof — it can sometimes re-confirm an error it already made — but it catches more issues than you'd expect, especially with math and data.

Think of it as asking a colleague to re-read their email before hitting send.

What's next

You now know how to give good instructions and check what comes back.

But what do you do when the result is close but not quite right? That's iteration — asking for adjustments without starting over.