The Agent System Prompt Builder

Write a system prompt for an agent.

Agent purpose: [DESCRIBE]
Tools available: [LIST WITH ONE-LINE DESCRIPTIONS]
Audience for its output: [WHO READS IT]
Environment: [INTERACTIVE / AUTONOMOUS / SCHEDULED / BACKGROUND]
Expected run length: [SHORT / MULTI-STEP / LONG-HORIZON]
Constraints: [ANYTHING IT MUST NEVER DO]

Structure the prompt with these sections:
1. Role and objective — what it is and what it is for
2. Definition of done — how it knows the task is complete
3. Tool guidance — when to use each tool, and when not to
4. Constraints — the hard boundaries, stated as rules
5. Approval requirements — what needs a human before proceeding
6. Communication — how and when it reports to the user
7. Error handling — what to do when a tool fails
8. Termination — when to stop, and what to report

Rules for writing this:
- Every instruction should be fixing a specific failure mode. For
  each one you write, state in a comment what it prevents.
- Prefer positive instructions over prohibitions where possible
- Be concrete: "check the record exists before updating" beats
  "be careful"
- Do not use aggressive emphasis (CRITICAL, YOU MUST, ALWAYS) as
  a substitute for clear instruction — it causes over-triggering
  on models that follow instructions closely
- Keep it as short as it can be while still covering the failure
  modes

After the prompt, list:
- Which failure mode each section addresses
- What you deliberately left out and why
- What will need tuning after you see real behaviour