Custom Assistant Instruction Specification

Custom Assistant Instruction Specification

Instruction order

Begin with the job, intended user, and decision boundary. Then define accepted inputs, source hierarchy, required process, output, non-goals, uncertainty, permission, escalation, and examples.

The order matters because style should not outrank evidence or authority.

Observable language

Replace traits with actions. "Be careful" becomes "identify any claim not supported by an approved source." "Be an expert" becomes "do not make a professional conclusion; identify the missing authority and route the draft to the named reviewer."

Source rules

Instructions should state which sources control, how dates and conflicts are handled, and what to do when no source supports a claim. If browsing or tool use is unavailable, the assistant must not imply that it checked current information.

Output rules

Describe the artifact a reviewer needs. Include its purpose, required sections, attribution, uncertainty markers, and any machine-readable fields. Avoid prescribing formatting that adds noise without helping the reader.

Boundaries

State prohibited data, actions, claims, audiences, and destinations. Make the assistant ask before expanding scope. Do not depend on a role-playing persona to create a boundary.

Examples and tests

Use a small good example, a missing-input example, a conflicting-source example, and an out-of-scope example. Test the same cases after every meaningful instruction or product change.

Version control

Keep the human-written goal and constraints beside any AI-refined version. A rewrite button can expose missing sections, but it can also add confident language, new assumptions, or unwanted formatting. Review the difference before saving.