A support-reply iteration becomes a standing instruction: status-and-date first, ticket number in sentence one, never overpromise — paste once per session, use all day.
Some conversations converge on something bigger than a prompt: a standing way of handling a task type. The support-macro case is the archetype — one good refund reply, iterated into rules (status and date first, ticket number in the first sentence, conservative timelines only), then declared the pattern for every refund reply. System Instruction mode frames the distillation for exactly that: a paste-once-per-session instruction, task-scoped on purpose — the usage notes draw the line to the System Prompt Generator for full operational personas.
Iterate once, properly
One reply, corrected until right — the conversation does the design work.
Frame as instruction
System Instruction mode converts the convergence into a standing, session-level rule set.
Stay task-scoped
It governs refund replies, not the assistant's whole identity — the boundary keeps it composable.
It stays task-scoped on purpose. The USAGE NOTES say to paste it once per session as a standing instruction and point to the System Prompt Generator for a full operational system prompt with role, escalation, and behavior rules. Conversation To Prompt Builder here governs one task type, refund-status replies, not the assistant's whole identity, which keeps it composable with other prompts.
They travel as an AVOID list so the model cannot rediscover them; the rejected apologetic opening ("we did nothing wrong here") is carried under both AVOID and REJECTED DIRECTIONS. Accepted corrections become REQUIREMENTS instead, and the prompt states later requirements override earlier ones where they conflict, the way the conversation actually resolved.
Better to read it once; the closing note says signals were "extracted VERBATIM from the conversation (deduplicated, never paraphrased) — read the prompt once before trusting it with a real run." The PRESERVED REQUIREMENTS tally (corrections, accepted preferences, standing rules, constraints) shows what carried, but you confirm CONSTRAINTS like ticket-number-first still hold before it drives live replies.
What a converged conversation contains: a rejected voice, an accepted framing, a hard constraint, a format decision, and a final-version moment — each mapped to its prompt section.
The result you finally liked lives inside the chat. Distill the iterations into one prompt: corrections become requirements, the accepted version becomes the quality bar.
In feedback threads the rejections carry the knowledge: the mocking empty-state enthusiasm, the passive phrasing — converted into the avoid-rules that define the style.
End a working session like a shift change, not an abandonment: state captured, decisions logged, next step named — ready for the next session to pick up.
Turns a reusable API request prompt into a clean variable set covering endpoint, method, auth, payload, response format, and error handling.
A reusable AI agent task template with variables for objective, context, available tools, constraints, success criteria, failure handling, and output format.
Distill an iterated conversation into one reusable prompt — corrections become requirements, rejections become an avoid-list.