Meta-Prompt: Generate Custom System Prompts

Use this when you need to create a system prompt for a new use case and do not want to start from a blank page. Describe the job in plain language, and the meta-prompt produces a structured, production-ready system prompt.

This is the "teach a man to fish" resource. Instead of searching for a pre-made system prompt that approximately fits your use case, you generate one that exactly fits it.

Especially useful for: building custom GPTs or Claude Projects, configuring AI tools for clients, setting up AI-powered workflows, creating role-specific chatbots, or any situation where you need a bespoke system prompt quickly.

The six-part structure prevents missing sections. Most hand-written system prompts are incomplete. They define a role but forget constraints. They add constraints but skip the output format. The prescribed structure ensures every system prompt covers: who the AI is, how it should behave, what it should produce, and how it should self-check.

"Anticipate failure modes" is the instruction that produces the best constraints. Rather than asking the model to write generic rules, this instruction pushes it to think about what goes wrong for this specific role. A customer support prompt gets constraints about empathy and escalation. A data analyst prompt gets constraints about accuracy and uncertainty. The constraints are tailored, not boilerplate.

The self-check section is often missing from system prompts. Most prompts are one-directional: do this, produce that. The self-check adds a feedback loop where the AI evaluates its own output before presenting. This consistently improves output quality across all use cases.

The 200-500 word target prevents two common mistakes. Too-short system prompts (under 100 words) lack the specificity to meaningfully constrain behaviour. Too-long system prompts (over 1000 words) dilute key instructions in noise. The target keeps prompts in the productive middle ground.

Add domain-specific context. If you primarily work in a specific field, add: "When generating system prompts for [your field], prioritise [specific quality criteria] and always include constraints about [common issues in your field]."

Adjust the structure. The six-part structure is a strong default. If your use cases consistently need additional sections (e.g. "Knowledge Sources" for research roles, or "Tone Guidelines" for writing roles), add them to the structure template.

Chain it with testing. After generating a system prompt, test it with 3-5 representative inputs. Then feed the results back: "Here is the system prompt you generated and the outputs it produced. Refine the prompt to fix [specific issues]." This creates a meta-prompt > test > refine loop.

A meta-prompt is only as good as the description you provide. "Make me a prompt for writing" will produce a generic writing prompt. "Make me a system prompt for writing LinkedIn posts for B2B SaaS founders, contrarian tone, 150-250 words, with a hook in the first line and a question at the end" will produce something immediately useful.

The generated system prompt is a strong first draft, not a finished product. You should always review and adjust it before using it in production. Pay particular attention to the constraints (are they specific enough?) and the self-check (does it catch the failure modes you care about?).

Claude: Produces well-structured system prompts with thoughtful constraints. The design decision explanation is particularly useful. Claude tends to write slightly longer system prompts than the 500-word target — trim if needed.

GPT: Generates solid prompts but may over-explain in the system prompt itself (adding context that should be in the user prompt, not the system prompt). Review for conciseness.

Gemini: Follows the structure reliably. May produce blander constraints than Claude or GPT. Push for specificity: "Make the constraints highly specific to this role. Avoid generic rules like 'be helpful.'"