System Prompt for Coding Agent: 12 Rules Templates

A system prompt for coding agent setups should encode your stack, repo rules, allowed tools, test commands, and secret handling before the agent touches files. Use these templates to draft policies, refusal lists, and mini evals. Do not invent CI scripts or print env secrets into chat. Below are 12 copy-paste prompts plus a flagship compiler.

Prefer modern Cursor project rules as .cursor/rules/*.mdc files when you use Cursor; a single legacy .cursorrules file is the older format. Use rule types "Apply to Specific Files" and "Apply Intelligently" when you split rules.

Prompts for the core system prompt

1. Role and stack lock

Write a system prompt opening that locks stack [stack] and repo layout [layout]. Forbid inventing other frameworks.

2. Imperative repo rules

Turn these standards into short imperatives for an agent: [rules]. No polite filler.

3. Definition of done

Done when: [criteria]. Expand into a checklist the agent must run, including commands [tests].

Prompts for tools and safety

4. Tool allowlist

Allowed tools: [tools]. For each, write when to use and when not to. Forbidden: [forbidden].

5. Secrets policy block

Write a secrets section from [policy]. Include refusal language for printing .env values.

6. Git safety

Draft rules for branch names, commit style [style], and bans on force-push to [protected branches].

7. Dependency installs

Package manager [pm]. Write rules for adding deps, lockfile updates, and refusing global installs.

Prompts for evals and hosting

8. Mini eval set

Create 5 tasks for [repo type] with pass/fail checks tied to [test commands]. Include one refusal task.

9. Cursor rules split

Split this system prompt [paste] into .cursor/rules/*.mdc ideas: one always-on, one Apply to Specific Files for [globs], one Apply Intelligently for refactors. Note that .cursorrules is legacy.

10. Output contract

Specify how the agent reports files changed, commands, test results, and residual risks.

11. Incident refusal drills

List 8 dangerous user requests (secret dump, prod deploy, disable auth). Write one-line refusals.

12. Diff budget

Write rules that cap unrelated refactors and require asking before renaming public APIs.

Build the full system prompt with the flagship

The Coding Agent System Prompt: Repo Rules, Tools, Tests compiles stack, rules, tools, tests, refusals, and a mini eval set. Its FastAPI + uv example requires ruff and pytest, forbids printing .env values, and omits CI vendors that were not in the inputs.

Related prompts

See the code review agent system prompt with severity rules, AI chatbot system prompt from business facts, System Prompt for AI Chatbot, and Claude Code Prompts for Security.

Related: How to write Cursor rules (.mdc) from a repo brief and keeping one AGENTS.md for Cursor, Claude, and Copilot.

Find more in the Agents category, look through Browse, or post your own on Share.