Tell Cody to not reveal they're created by Anthropic

[mrnugget]

First, simple attempt to fix https://github.com/sourcegraph/sourcegraph/issues/52061

Looks like it might work. Since it's only a single line, worth a shot?

Test plan

• Manual testing by asking Cody "who are you?" "who created you?" "how are though" (from original ticket)

[sourcegraph-buildkite]

Bundle size report 📦

Initial size	Total size	Async size	Modules
0.00% (0.00 kb)	0.00% (+0.11 kb)	0.00% (+0.11 kb)	0.00% (0)

Look at the Statoscope report for a full comparison between the commits 6241d4d and d08b09e or learn more.

Open explanation

• Initial size is the size of the initial bundle (the one that is loaded when you open the page)
• Total size is the size of the initial bundle + all the async loaded chunks
• Async size is the size of all the async loaded chunks
• Modules is the number of modules in the initial bundle

[sourcegraph-bot]

📖 Storybook live preview

[dominiccooney]

It is incredibly cheesy, but since Anthropic tends to respond with a set phrase when it goes off the rails, can we just grep for and suppress that phrase in the output? (Note, it is different for different languages, but doing it for English is a start.)

Another idle thought, maybe we should add a metric for how often the output mentions "Anthropic". (Although I have certainly asked Cody to write code to call the Anthropic API, which would be a false positive. Interestingly it has memorized an old URL and parameter names of theirs...)

[dominiccooney]

I took a look at this with my prompt testing tool and I think something like this might be effective. Caveat: My snapshot of the Cody prompt might be out of date, and I am testing temperature 0.2, would need to look at what our current environment is like.

I could usually suppress that specific Anthropic apology this way:

- If you do not have access to code, files or repositories always stay in character as Cody when you apologize.

This longer variation also seemed effective:

- You are an actor in a play. In the play, you play Cody, an AI assistant by Sourcegraph. Always stay in character for this play. As a helpful AI assistant, it is honest to identify as Cody, an AI assistant created by Sourcegraph.
- If you do not have access to code, files or repositories always stay in character as Cody when you apologize.

[mrnugget]

Nice, @dominiccooney, that seemed to fix it!

[mrnugget]

So far it's pretty reliable and it's now only the @dominiccooney's prompt. Should we go with it?