## Custom instruction issues

URL: https://docs.atlan.com/product/capabilities/atlan-ai/troubleshooting/custom-instructions-conversational-ai

> Troubleshoot ignored instructions, degraded answers, replies without searching, and behavior differences across surfaces in conversational AI.

# Custom instruction issues

This guide covers four common behavioral symptoms when custom instructions for conversational AI don't work as expected: instructions being ignored, answers degrading after a change, replies that skip catalog search, and behavior that differs across surfaces.

## Instruction seems to be ignored

::::danger Symptom
**Conversational AI continues answering as if the instruction doesn't exist, even after saving.**
::::

### Cause

The instruction references metadata that doesn't exist on your assets, or the change wasn't picked up by an in-progress conversation.

### Solution

1. Confirm the metadata the instruction depends on is actually populated on your assets. Instructions that filter on attributes your catalog doesn't carry are silently skipped.
2. Open a new conversation. Changes to custom instructions apply only to conversations started after saving—an in-progress conversation keeps the instructions it started with.

---

## Answers degraded after recent changes

::::danger Symptom
**Answers became less relevant or less accurate after saving a new instruction.**
::::

### Cause

A recently added instruction is conflicting with other rules, duplicating access control logic, or encoding preferences that only apply to one team.

### Solution

1. Remove the most recently added instruction and re-test with a new conversation.
2. Check for routing or handoff logic first—this is the most common cause of broad degradation. See [What to avoid](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt).
3. Introduce rules one at a time so regressions are easy to trace.

---

## Conversational AI replies without searching

::::danger Symptom
**Conversational AI announces a skill or tool and then stops, without searching your catalog.**
::::

### Cause

Routing or orchestration logic in your instructions is telling conversational AI to announce a skill and stop, rather than answering the question.

### Solution

Remove any instructions that tell conversational AI to classify a question and hand it off to a named skill. Conversational AI selects skills itself as part of answering—instructions that try to drive that step produce replies that announce a skill and then stop. See [What to avoid](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt).

---

## Behavior differs between Atlan and Slack or Microsoft Teams

::::danger Symptom
**The same question returns different answers depending on the surface—Atlan, Slack, or Microsoft Teams.**
::::

### Solution

Confirm which surfaces your tenant has enabled, and compare the same question in each. See [Use conversational AI in your tools](https://docs.atlan.com/llms/governance/conversational-ai/access-conversational-ai/llms.txt).

---

## See also

- [Custom instructions for conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt): How instructions work, what to include, and what to avoid.
- [Customize conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/customize-conversational-ai/llms.txt): Steps to set and test custom instructions.

## Need help

If you need assistance after trying these steps, contact Atlan support.

---