agent-blindspot-questions

---
name: agent-blindspot-questions
description: "Use two questions to surface what an AI agent skipped or is unsure about before you trust its work — its low-confidence list and its 'what am I missing?' blind spots."
source: https://x.com
generated: 2026-07-07T18:04:34.442Z
category: concept
audience: general
---

## When to use

- An agent hands back a confident answer and you want to stress-test it before acting
- Before merging, shipping, or deciding on something the agent produced
- You suspect the model glossed over details or made unstated assumptions
- You want the agent to self-audit instead of you manually hunting for gaps
- A plan or analysis feels too clean and you want the risks it didn't volunteer

## Key concepts

### The confidence question

Ask: 'What are you least confident about right now?' The model will list 6-7 things it never properly investigated. Roughly 1 in 4 times, one of them turns out to be a big deal — a wrong assumption or an unchecked path it silently glossed over.

### Investigate the list exhaustively

Once it surfaces the low-confidence items, follow up with 'investigate each one exhaustively, find the root cause.' This converts a vague hedge into concrete verification work instead of leaving the uncertainty buried.

### The blind-spot question (Sam Altman's)

Ask: 'What's the biggest thing I'm missing about this situation? What don't I realize?' It pulls out framing errors and unstated assumptions the model saw but didn't consider worth raising unprompted.

### The model knows what it skipped

Models generally track what they didn't investigate or were unsure about — they just don't volunteer it unless asked. These questions make the skipped work explicit rather than letting a confident tone hide it.

## API reference

```
Least-confident prompt
```

Surface the things the agent never properly investigated.

```
What are you least confident about right now?
```

```
Exhaustive-investigation follow-up
```

Turn the low-confidence list into real root-cause verification.

```
Investigate each one exhaustively, find the root cause.
```

```
Blind-spot prompt
```

Expose framing errors and unstated assumptions the model didn't raise.

```
What's the biggest thing I'm missing about this situation? What don't I realize?
```

## Gotchas

- Not every low-confidence item is a real problem — expect noise, and triage the list rather than chasing all 6-7 blindly
- 'Investigate exhaustively' can burn tokens; scope it if the list is long
- These prompts reduce, but don't eliminate, missed issues — treat them as a fast self-audit, not a guarantee
- Ask before you act on the agent's work, not after you've already shipped it

---
Generated by SkillMake from https://x.com on 2026-07-07T18:04:34.442Z.
Verify against source before relying on details.

File: ~/.claude/skills/agent-blindspot-questions/SKILL.md