Everything on this path so far has been about what Claude can do. This page is the honest one.
Claude is sometimes confidently wrong. Treating its output as automatically correct will eventually cost you. The fix is a habit, not a setting.
Some of the specifics here will be less true in two years, as models improve. The habit will still be the right one.
The ways it’s wrong
Roughly in order of how often you’ll meet them:
- Made-up specifics. It cites a file that doesn’t exist, a number that isn’t in your spreadsheet, a quote nobody said. The shape of the answer is right; the details are invented. The word for this is hallucination.
- Stale knowledge. Its training has a cutoff. Recent events, new versions, the change you made five minutes ago — it can miss all of those unless it goes and looks.
- Confident wrongness. The dangerous one. It doesn’t reliably know when it’s guessing, so a guess and a fact arrive in the same tone of voice.
- Wrong scope. You asked for one change; it also changed a related-looking thing that wasn’t related.
The good news is that nearly all of it is catchable in seconds — if you look.
The habit
Pick one move and do it every time Claude tells you something that matters:
- It says a file contains something? Make it show you the exact line. Or open the file and look.
- It made a change? Read the diff, not the summary. You already know how.
- It says it did something and it worked? Check the thing it says worked.
Call it the same instinct as glancing in the mirror before changing lanes: cheap, automatic, occasionally the whole ballgame.
Reading is safer than writing
A frame worth keeping:
- Asking Claude to read and explain is low-risk. The worst case is a wrong explanation, which you notice when the next thing stops making sense.
- Asking Claude to write or change is higher-risk. A wrong change persists quietly. Now you have a problem you didn’t create and didn’t notice.
So when you’re in unfamiliar territory, start on the reading side. Ask it to explain the thing before you ask it to touch the thing.
That’s the path
Eight pages ago you hadn’t opened the app. Now you know what it is, what it can see, what it will ask before it acts, how to read what it proposes, how to undo it, and where it lets you down.
That’s genuinely the whole foundation. Everything else — skills, subagents, MCP, cost, rolling it out to a team — is depth on top of it, and each of those is a topic you can now read on its own terms.