Overview
Compose a production-grade system prompt from five guided blocks - role, mission, rules, output format and verification - and export it as markdown you can paste into an agent config, AGENTS.md or CLAUDE.md file. The page assembles a consistent skeleton so nothing important is forgotten, counts its tokens, and never sends your prompt text anywhere.
How It Works
Fill four fields: the role (the page prefixes it with 'You are a'), the mission, the rules (one per line, rendered as bullet points), and the output format. The assembled markdown appears live in a copyable pane under fixed headings: Role, Mission, Rules, Output format and Verification. The Verification section is added automatically - it instructs the model to ground every claim and to say exactly what is missing rather than guessing. A token counter runs on the finished prompt so you can see its footprint against your context budget.
Step-by-Step Usage Guide
- Describe the role in a short phrase.
- State the mission: what the model should accomplish.
- Enter rules one per line; each becomes a bullet.
- Copy the assembled markdown and paste it into your agent config or AGENTS.md.
Technical Specifications & Standards
The value here is the fixed skeleton, which mirrors the structure experienced prompt engineers converge on: identity first, task second, constraints third, format fourth, and a self-check clause last. The builder is deliberately opinionated about the last block - it appends a verification instruction verbatim, because models that are told to verify and to admit missing information fail less often than models told only what to do. Rules are split on newlines, so blank lines are dropped and each non-empty line becomes one bullet; that means a rule cannot span two lines. The output is a markdown fragment, not a full document, which is exactly what AGENTS.md-style files expect at their core. Token counting uses the same in-browser GPT byte-pair tokenizer as the other AI tools on this site. Because everything is local, you can iterate on proprietary system prompts without them touching a server.
Targeted Use Cases
- Bootstrapping a code-review agent prompt with role, rules and output contract.
- Writing an AGENTS.md or CLAUDE.md skeleton for a repository.
- Standardising system prompts across a team with the same five blocks.
- Checking a draft prompt's token footprint before adding it to a context-heavy pipeline.
Notes & Gotchas
- Keep the role to one sentence; long personas dilute the mission.
- Make each rule testable - 'cite line numbers' beats 'be accurate'.
- Specify the output format concretely, naming sections or schema fields.
- Read the Verification block in the preview and edit your rules to complement it, not repeat it.
Frequently Asked Questions
Can I edit the assembled markdown after copying?
Yes. The output is plain markdown - paste it anywhere and edit freely. The builder just guarantees the skeleton starts consistent.
Why is there a fixed Verification section?
It encodes two behaviours that measurably reduce hallucination: grounding claims in provided material, and naming what is missing instead of guessing. The builder treats it as non-negotiable.
How are rules formatted?
One rule per line, each rendered as a markdown bullet. Blank lines are dropped, so spacing in your input does not affect the output.
What is the token count for?
System prompts sit in every request's context, so their size multiplies across thousands of calls. The counter shows the real GPT-token cost of the assembled prompt.