Design agent personas that actually follow instructions

The gap between a chatbot and a colleague is instruction design. Persona files — identity, values, operating rules — are what keep an agent consistent across long sessions and new tasks. Most people write them like bios; they work when written like contracts.

THE PAIN

• ✗ The agent drifts off-style twenty minutes into a session
• ✗ Rules get ignored the moment they're inconvenient
• ✗ Every new project means re-explaining how you work

THE PLAYBOOK

1. 01

   Separate identity from rules

   SOUL.md carries voice and values; AGENTS.md carries hard operating rules. Mixing them weakens both.

2. 02

   Write rules as gates, not vibes

   'Never commit without tests passing' beats 'be careful about quality.' Checkable beats aspirational.

3. 03

   Add anti-rationalization clauses

   Name the excuses in advance ('this is too simple to need a test') so the agent can't use them.

4. 04

   Version and iterate

   When the agent misbehaves, patch the persona file — it's a bug fix, and it compounds.