CourseGuidesArticlesPodcastGitHubAboutStart free course
Course/When Things Go Wrong

6 min · +100 XP

Typical mistakes with agents

Five patterns that waste time, and the fix for each.

By the end of this lesson

You will be able to recognise the five common failure patterns early and apply the fix.

Most problems with agents come from a few repeating habits rather than from the model itself. The official best-practices guide lists five of them. Each has a simple fix.

Five mistakes and the fix

  1. 1Everything in one conversation. You start with the budget tool, ask about a holiday letter, then return to the budget. Claude's context is now full of unrelated things. Fix: type /clear when you move to an unrelated task.
  2. 2Correcting over and over. Claude does it wrong, you correct, it is still wrong, you correct again. The conversation fills with failed attempts. Fix: after two failed corrections, type /clear and write a better first request that includes what you learned.
  3. 3A very long CLAUDE.md. The project memory file grows to hundreds of rules, and the important ones get lost. Fix: keep it short; delete anything Claude already does correctly without being told.
  4. 4Trusting without checking. The result looks right but fails in unusual cases. Fix: always ask for a check, such as a test, a script or a screenshot, and look at the evidence.
  5. 5Endless investigation. "Look into why it is slow" with no limits, and Claude reads hundreds of files. Fix: name where to look and what the answer should be, for example "name the cause in three sentences".
The second mistake, and the fix
you ▸ the chart is still wrong
you ▸ no, still wrong, the weeks are shifted
you ▸ /clear
you ▸ In @app.js the weekly chart must start on Monday and include today. Last attempts shifted the weeks by one day because of time zones. Fix it, add a test for a week that crosses a month, and run it.
For beginners
›Why a fresh start helps

Claude keeps the whole conversation in its context window, which has a limited size. When the window is full of wrong attempts and unrelated topics, Claude works worse. /clear empties the context but leaves your files as they are, so a clear new request starts from a clean page.

For experienced users
✦If you are experienced

The underlying constraint is that performance degrades as the context window fills. Check it with /context, compact with /compact plus instructions ("focus on the API changes"), or summarise part of the session from the /rewind menu. Ask quick questions with /btw so they never enter the history. Prune CLAUDE.md hard and move rules that must always hold into hooks. For unattended work, add an independent review step: a subagent that sees only the diff and the criteria, told to flag only gaps that affect correctness.

Quick check

You have corrected Claude twice on the same bug and it is still wrong. What do the docs suggest?

Quick check

Which request avoids the "endless investigation" pattern?

Facts checked against the official Claude Code docs October 24, 2026 · source

EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Chief Executive Officer, digital health