Ai Agent System Prompts
Prompt Engineering
AI Agent System Prompts
Custom-engineered system prompts for a multi-agent AI content pipeline at DPG Media — covering creative director, strategist, and content production agents built to deliver precise, on-brand outputs at scale without human review on every piece.
Client
DPG Media
Tools Used
Claude, GPT-4, OpenClaw
Year
2025

Project Overview
The Challenge
DPG Media operates multiple content brands across different editorial verticals, each with its own distinct voice, tone, and audience. The challenge with deploying AI into their content pipeline wasn’t access to capable models — it was getting those models to behave as if they understood the brand. Without properly engineered system prompts, the agents defaulted to generic outputs that required heavy human editing before anything was publishable. Every agent was essentially operating without a proper brief, producing technically correct but editorially useless content that added workload rather than reducing it.
The Solution
We architected a three-agent pipeline on OpenClaw Gateway with each agent assigned a tightly scoped role and a precision-engineered system prompt. The Creative Director agent receives raw inputs and transforms them into structured creative briefs — it doesn’t write content, it directs. The Brand Strategist enriches each brief with brand voice guidance, audience context, and channel-specific constraints before handing it forward. The Content Producer agent then executes against both, operating with detailed instructions covering editorial tone, sentence structure, vocabulary restrictions, and formatting rules specific to each DPG brand property.
Each system prompt was built in layers: a core role definition, non-negotiable output rules, few-shot examples of acceptable and unacceptable outputs, and explicit instructions for handling edge cases and ambiguous inputs. The prompts were tested iteratively against a benchmark set of real content briefs and refined until first-pass acceptance rates consistently exceeded 90% without human correction required.
The Results
Human revision rounds dropped by 90% compared to unstructured AI usage. Content that previously required multiple edit cycles now exits the pipeline ready to publish on the first pass in the large majority of cases. The structured three-agent architecture also gave the team a clear intervention point for each category of problem — strategic issues get addressed at the Strategist prompt level, tone issues at the Content agent level — making the system maintainable and improvable over time rather than an opaque black box that nobody knows how to fix when it drifts.
Project Details
90%
Reduction in Output Revisions
5x
Faster Content Production
3
Specialist Agents Engineered
Tools Used
Claude AI
GPT-4
OpenClaw Gateway
n8n
Agents Built
Creative Director
Brand Strategist
Content Producer
Prompt Samples
Inside the Agent Prompts
Each agent runs on a layered system prompt that defines its role, constraints, and output format with no ambiguity. Below are excerpts — ↓ replace with your actual prompt content before publishing.
Agent 1 — Creative Director
System Prompt Excerpt
You are the Creative Director for [brand]. Your role is to receive raw content inputs and transform them into structured creative briefs that downstream agents can execute against without further clarification. For every input you receive, output exactly: 1. A one-sentence strategic objective 2. The primary audience and their core motivation 3. The desired emotional response this content must create 4. Three mandatory elements that must appear in the content 5. Two things this content must never do or say You do not write content. You direct. Output a brief — never a draft.
Agent 2 — Brand Strategist
System Prompt Excerpt
You are the Brand Strategist. You receive creative briefs from the Creative Director and enrich them with brand context before passing them to the Content Producer. Your enrichment must cover: – Which brand voice register applies: authoritative / conversational / aspirational – What the target audience already believes, and whether this content should confirm or challenge it – Active campaign context that must be reflected in the content – Channel-specific constraints: character limits, platform tone norms, format requirements Do not rewrite the brief. Annotate and extend it. Flag any strategic contradictions before passing forward.
Pipeline Output — Content Producer Result
Final Output — Ready to Publish, First Pass
↓ Paste a real example of the final content the pipeline produced here — a social caption, article intro, email subject line set, or ad copy. Visitors should read this and immediately see it sounds like a real brand, not a generic AI output. This is the proof that the prompt engineering worked.

Liked what you saw?
Let’s Engineer Yours
Got an AI setup producing generic outputs? Let’s build a proper prompt architecture for your agents — structured, tested, and tuned to your brand voice.
