1. WHO versus WHAT

A profile and an agent answer different questions. The profile is WHO you are: your facts, how you reason, how you like to be spoken to, your writing voice. The agent is WHAT the job is: the domain expertise, the workflow, the contract, the guardrails. Keep each side pure and one profile composes with every specialist you ever build.

The impurity test cuts both ways. 'Keep answers short' inside your legal agent will fight your profile someday; 'always cite case law' inside your profile will pollute your cooking questions. Preference lines belong to the profile; domain rules belong to the agent.

2. Precedence: who wins, decided in advance

Composed instructions will eventually conflict, and an undeclared conflict is resolved by the model's mood that day. So declare the order. From strongest to weakest: safety and accuracy first; then the agent's professional rules; then the task at hand; then your user facts; then your reasoning preferences; then your communication preferences; then writing style last.

Notice the logic of the ranking: things that would make the answer wrong or harmful outrank things that would make it feel unlike you. Style is last not because it does not matter, but because it is the one layer that should never win an argument against correctness.

3. Two collisions, called in advance

Collision one: your profile says 'short answers, no padding'; your financial agent's rules require a risk disclosure. The disclosure stays, professional rules outrank communication preferences. The right composition honors both ranks: a tight answer AND the required disclosure, with everything else trimmed hard.

Collision two: your profile says 'casual, conversational voice'; your legal agent must state terms precisely. Precision stays, a casually-worded liability clause is simply wrong. Your voice survives in the surrounding explanation, never inside the load-bearing language.

4. Composition as a working system

Practically, a composed setup is three pieces of text: the profile, the agent, and a short precedence note that states the order, so conflicts are settled by rule instead of by mood. Write the note once and reuse it with every pairing; it is the constitution of your system.

Overrides have a rank too: they live in the task layer. 'Just one option this time, not your usual three' beats a stored preference for that exchange, because the task outranks user facts and everything below. But it cannot beat what sits above: no task instruction removes a required disclosure or licenses a made-up figure. The order holds even against you, and that is a feature.