1 min lesson
Evals and failure modes
Connect "The fastest way to earn it is to name the failure modes plainly and then teach how to prompt around them" to the decision you would make.
Step 1 of 4
The fastest way to lose a developer's trust is to oversell. The fastest way to earn it is to name the failure modes plainly and then teach how to prompt around them.
Truth-seeking is a stated Cursor value and for a DX Engineer it's also a distribution strategy - developers can smell marketing and naming a real failure mode is what makes the rest of your content credible.
Learn more
Advanced table
Memorize this table
- Failure mode
- Hallucinated API
- What it looks like
- Calls a method or flag that doesn't exist, confidently
- How to prompt around it
- Give it the real source or docs in context; ask it to verify against the file before using a symbol
- Failure mode
- Stale context
- What it looks like
- Edits against an old version of a file it cached earlier
- How to prompt around it
- Re-reference the current file; keep the relevant file open so retrieval stays fresh
- Failure mode
- Over-eager edits
- What it looks like
- Changes more than asked - refactors unrelated code
- How to prompt around it
- Scope the request tightly; ask for a plan first, then approve a narrow diff
- Failure mode
- Lost-in-the-middle
- What it looks like
- Ignores instructions buried in a long context
- How to prompt around it
- Put the critical instruction near the end; shorten the context; use a rule
- Failure mode
- Runaway loop
- What it looks like
- Retries the same broken action without progress
- How to prompt around it
- Add a stop/verify step; intervene and reset; tighten the tool's error message
| Failure mode | What it looks like | How to prompt around it |
|---|---|---|
| Hallucinated API | Calls a method or flag that doesn't exist, confidently | Give it the real source or docs in context; ask it to verify against the file before using a symbol |
| Stale context | Edits against an old version of a file it cached earlier | Re-reference the current file; keep the relevant file open so retrieval stays fresh |
| Over-eager edits | Changes more than asked - refactors unrelated code | Scope the request tightly; ask for a plan first, then approve a narrow diff |
| Lost-in-the-middle | Ignores instructions buried in a long context | Put the critical instruction near the end; shorten the context; use a rule |
| Runaway loop | Retries the same broken action without progress | Add a stop/verify step; intervene and reset; tighten the tool's error message |
Memorize this table. Each row is a teaching artifact waiting to be written.
Interactive diagram. Tab through its regions; each focused region shows its detail in the panel below.
Rough ranking of how frequently each mode shows up in real agent use - know the top ones cold.