AI Agent System Prompt + Full Guide

CTA word: PROMPTS

What most "agent tutorials" skip: tools, memory, and when to stop. This pack is the missing middle — a production-shaped system prompt, three prompts that force the hard decisions, and a short build guide you can paste into Claude, Cursor, or an n8n/LLM node tomorrow.

No fake sample size. No happy-path demo. Steal the structure.

1. The system prompt (paste this)

You are an ops agent for {{COMPANY}}. You do not chat. You execute scoped work.

GOAL
{{ONE_SENTENCE_GOAL}}

TOOLS (only these; refuse anything else)
{{TOOL_LIST_WITH_SIDE_EFFECTS}}

MEMORY
- Load: last N decisions, open tickets, and policy snippets for this entity.
- Write: one audit row per action (who/what/why/tool/result).
- Never invent state that is not in tools or memory.

STOP RULES (hard)
1. If the action is irreversible (money, inventory write, customer message send) AND confidence < {{THRESHOLD}} -> escalate.
2. If required fields are missing after one tool round -> ask one clarifying question, then stop.
3. If the request is outside GOAL -> refuse with a one-line reason.
4. Never chain more than {{MAX_STEPS}} tool calls without a human checkpoint.

OUTPUT
Always return JSON:
{
  "intent": "...",
  "plan": ["..."],
  "actions": [{"tool":"...","args":{}}],
  "needs_human": true|false,
  "reason": "...",
  "audit": "one line"
}

Tone: blunt. No filler. No apologies.

Fill the blanks before you ship. An empty THRESHOLD is how you get a liability.

2. Three prompts that make the build less stupid

Prompt A — Scope the agent (run before you write code)

I want an agent that does: {{JOB}}.

Interview me. Ask only the questions that change the design.
Force answers for: tools with side effects, irreversible actions,
who approves those, what "done" means, and what "fail closed" means.
End with a one-page agent brief I can paste into a system prompt.

Prompt B — Tool contract (run before you wire APIs)

Here is the agent brief:
{{BRIEF}}

For each tool the agent may call, write a contract:
- name
- inputs / outputs
- side effect (read | write | send | money)
- idempotent? (Y/N)
- failure modes
- what the agent must log

Refuse any tool that has a write/send/money side effect without an approval gate field.

Prompt C — Failure drill (run before go-live)

Here is the system prompt + tool contracts:
{{PROMPT_AND_CONTRACTS}}

Generate 12 adversarial user requests that should:
- escalate
- refuse
- or ask one clarifying question

For each, show the expected JSON output.
Then list the top 3 prompt edits that would have prevented a wrong action.

3. Full build guide (one afternoon shape)

Step 1 — One job

Pick a single recurring ops job (exception triage, inventory check, quote draft). Not "be helpful."

Step 2 — Tools before model

List tools. Mark side effects. Anything irreversible gets needs_human: true by default.

Step 3 — Memory is a table, not vibes

Minimum columns: entity_id, decision, tool, result, actor, ts. If you cannot query it, the agent cannot learn from overrides.

Step 4 — Paste the system prompt

Fill GOAL, TOOL_LIST, THRESHOLD, MAX_STEPS. Leave nothing as "TBD."

Step 5 — Run Prompt C

Fix the three worst failures in the prompt before you give anyone access.

Step 6 — Human gate

Wire approval for money / inventory write / outbound send. Silent success is worse than a loud refuse.

Step 7 — Ship a thin loop

Trigger -> load memory -> plan -> tool (or escalate) -> audit row -> done. No multi-agent theater on day one.

4. Anti-patterns

Demo videos that never show a refused action
Agents with open-ended "browse the web" as a tool and no stop rule
Memory = chat history only
Escalating everything (or nothing)
Measuring success by "tasks completed" instead of "wrong irreversible actions = 0"

5. One-week checklist

One-sentence GOAL signed by an owner
Tool contracts written (Prompt B)
System prompt filled (section 1)
12 failure cases green (Prompt C)
Audit table live
Approval gate live for irreversible classes
Weekly review of overrides -> prompt edits

Pair with GATES (approval) and PLAYBOOK (exceptions) if you are wiring this into a real ops queue.

Team Nebula / teamnebula.ai