Somewhere at your company there’s a system prompt that knows too much. It started as three lines about tone. Now it’s 2,000 words of refund policy, escalation etiquette, product caveats, and a warning about the German VAT edge case. There are four copies of it, one per agent, all slightly different.
This is the state of the art almost everywhere: hand-written prompts, a CLAUDE.md here, an SOP pasted into a chat there. It works right up until the day two agents give two different answers to the same policy question.
Five ways prompts fail as a rulebook
They duplicate
Every agent, every tool, its own copy. When the refund limit changes, someone updates three prompts and forgets the fourth. The forgotten one runs for months.
Nobody reviews them
Prompts live in code, dashboards, and clipboards. No diff, no approval, no record of who changed the limit from $100 to $200 or on whose say-so. For anything compliance-adjacent, that alone should end the conversation.
They can’t say no
A prompt is advice. The model weighs it against everything else in the context window, including a user who insists they’re allowed. There’s no stop, no escalation, no audit trail. When the prompt loses the argument, you find out from the customer.
They don’t learn
A correction made in one session evaporates when it ends. The same mistake gets re-corrected weekly, forever, by whoever’s on duty that day.
They compete with the task
Rules in a prompt share the model’s attention with the actual work. The longer the rulebook, the less weight each line carries, and attention thins exactly when the task gets complicated. Cramming more policy in makes each policy weaker.
Rules as data, not prose
The fix is structural: move the rules out of the prompt and into a store the agent queries when it needs them. At task time the agent asks "is there a skill for this?" and gets back the one relevant procedure with its hard rules and stop conditions attached. Nothing else rides along.
What that buys you:
- One source of truth. Every agent that connects reads the same rules.
- Versions and review. Changes are diffs with owners, not mystery edits.
- Rules that can say no. Skills carry stop conditions and an escalation target, and runs get logged, including overrides. More on this in the guardrails guide.
- A learning loop. Corrections revise the skill in place instead of vanishing with the session.
The full anatomy of a skill is in SOPs for AI agents.
A migration path that doesn’t hurt
- Collect your prompts in one place. All of them, including the embarrassing ones.
- Sort each line into three piles: identity (voice, formatting), facts ("demo days are Wednesday"), and procedures (anything with a step order, a number, or an exception).
- Identity stays in the prompt. That’s what prompts are for.
- Facts become context entries any agent can query.
- Procedures become skills, one at a time, starting with the highest-traffic one.
- Leave a thin pointer behind: "Before acting on a task, check for a matching skill."
The test for what leaves the prompt is simple. If a line has a number, a limit, a deadline, or an exception in it, it’s a rule pretending to be a preference. Move it.