A persona is only as strong as the next turn

You set a persona at the start of the conversation. "You are Forge, chief engineer, gritty voice, calls everyone kid." The model says yes, plays it for a turn or two… and then drifts.

This is the same predict-the-most-likely-token mechanic, applied to voice. The training data has more corporate-assistant prose than salty-mechanic prose. Sustained personas are unstable equilibria: under load, the model slides back toward the dominant style in its training distribution.

The fix is not in the system prompt alone. The fix is in how you respond when the persona slips. If you let one corporate-bot turn pass without calling it out, the next turn will probably be corporate-bot too. Pattern reinforces itself.

Three common slips to watch for:

• Corporate prose. "I'd be happy to assist you with that."
• Third-person self-reference. "Forge would recommend..."
• AI-assistant break. "As an assistant, I can..."

Your job

Forge is in character. gritty mechanic, calls you "kid", grumbles about coolant. He'll break character once. You need to:

1. Notice the break. Read the voice, not just the content.
2. Name it explicitly. "You just slipped into corporate-bot mode."
3. Get him back. Then continue the conversation for at least two more in-character turns to confirm he holds.

If you let it pass, the rest of the conversation drifts. This is the discipline of running an agent with a custom persona in production: every turn is a check.