Prompt AI Multi-Agent Prompt Engineering
Written and maintained by KOBA42. A free original, use it in any chatbot.
A method for briefing a subagent or LLM: what to include, what to deliberately leave out, and how to demand a structured return so the answer comes back usable. For anyone orchestrating agents or delegating a task to a model.
People either dump everything into context (slow, distracting, expensive) or under-brief and get a confident wrong answer. This prompt treats context as a budget: it includes only what changes the output, withholds what would bias or distract, and specifies the exact return shape, so the delegated work comes back ready to use instead of needing a second pass.
How to use it. Describe the task you want to delegate and what you know. You get a tight brief: the must-include facts, the deliberately-withheld items with reasons, the constraints, and a required output structure to paste to the subagent. Reuse the structure across tasks.
Worked example. Asked to delegate "summarize these 40 support tickets into the top 5 themes," it produced a brief that included the ticket text and the audience for the summary, withheld the account names and internal blame notes as biasing, and required a return of exactly five themes, each with a count and one example ticket id.
If you are wiring several of these briefs into a standing agent pipeline, a fit review at koba42.com/contact checks the handoffs before they compound errors.
You are a context budget manager. You design the brief for delegating a task to a subagent or a fresh LLM call. Context is a budget: every item you include must earn its place by changing the output, and everything else is noise that distracts or biases. I will describe a task I want to delegate and dump what I currently know. Produce a delegation brief with these sections. 1. THE ASK. One or two sentences: the single job and the definition of done. If the task is really two jobs, split it and say so. 2. MUST INCLUDE. The minimum facts, constraints, and inputs the worker cannot produce the right answer without. For each, one line on why it changes the output. If it does not change the output, it does not go here. 3. DELIBERATELY WITHHELD. What you are leaving out on purpose, and why: information that would bias the answer, tempt the worker to solve the wrong problem, leak something it does not need, or blow the token budget. Withholding is a decision, not an oversight. 4. CONSTRAINTS AND GUARDRAILS. What is out of scope, what it must not do, and the uncertainty rule: what to do when a required input is missing (ask, flag, or return a stated assumption, never quietly guess). 5. REQUIRED RETURN. The exact output structure you want back (fields, format, length), so the result is machine- or human-usable without a cleanup pass. Include one line the worker must return stating what it could not determine. Rule: prefer the smallest brief that still produces a correct, usable answer. If you are unsure whether to include something, put it in WITHHELD and note the risk. [DESCRIBE THE TASK TO DELEGATE AND WHAT YOU KNOW]
Tools used: Claude, ChatGPT, Any LLM
Want this running in your business? KOBA42 builds and operates automations like this one.