Why does the Rovo feature ignore instructions?

[Robby Roose]

I am constantly being slowed down in my work by having to revert changes made by this tool - for example, page titles in Confluence conform to a very consistent standard for traceability / audit purposes.

They cannot be summarised or changed - in fact, it would be an ideal thing for an LLM  to do as it's not complex. Instead I get a frequently incorrect summary. Fine. It's 'AI'. 

I can't disable it as it's used by others for things LLMs are good for (i.e. if you don't know JQL) - so I provide some explicit instructions when interacting with my Atlassian Account. For example 'Never update or add a title to a page I am working on'.

They are ignored.

When probed the tool explained that my rules are being superseded by system or developer ones, and with a very small amount of basic prompting I was able to retrieve it's system prompts and the names of Atlassian’s' internal tools used by these agents (despite there being an explicit prompt stating not to do this - so it's not just me it ignores).

I'm not going to share the transcript of this conversation here, but based on the rules Atlassian have defined, and the fact they've been broken makes me very concerned about the safety of this tool - does it actually have any guardrails? Would it respect them?

I've had these tools described as a very new intern - competent but inexperienced. This one is verging on obliviously malicious.

It's not winning hearts and minds - this was meant to help, not 

[Nick Jhon]

It sounds frustrating, and your concerns are valid—especially when consistency and control (like strict page titles) are critical for your workflow.

Rovo likely behaves this way because system-level and developer-defined instructions take priority over user instructions. Even if you tell it not to modify titles, it may still override that if it’s designed to optimize or summarize content.

This usually happens due to:
- Instruction hierarchy (system rules override user rules)
- Over-optimization (trying to “improve” content automatically)
- Context limitations (not always remembering preferences)

To handle this:
- Clearly repeat constraints like “Do not modify titles under any circumstances”
- Avoid using it on sensitive fields like titles if possible
- Share feedback with Atlassian, as this is a product limitation

Regarding your concern about safety, partial exposure of system behavior can happen, but it doesn't necessarily mean there are no guardrails—just that they may not be fully refined yet.

Overall, it's useful in some cases, but still unreliable for strict, rule-based tasks.

By the way, have you taken the Rice Purity Test yet? I’d be interested to hear your score.