Home
Back to Portfolio

AGENTS.md / Operating Prompt

Agent Prompt System

A structured operating prompt that tells an AI agent how to interpret requests, control scope, make minimal changes, verify results, and stop at the correct point.

CONTROL CONSISTENCY TRACEABILITY SAFETY
AGENTS.md

01priorityuser_request > host_rules > project_style

02scopeexplicit_task_only

03inspectiondirectly_related_files

04changeminimum_safe_diff

05verificationproportional_to_risk

06stoprequested_result_complete

07reportfiles + changes + manual_check

POLICY LOADEDREADY

Why It Is Used

Reliable AI work needs operating boundaries

Without clear operating instructions, an agent may inspect too much, change unrelated files, add unnecessary complexity, or continue after the requested result is complete. This prompt turns broad expectations into explicit execution rules.

01 / SCOPE

Prevents scope creep

Limits work to the latest explicit request and the files directly required to complete it.

02 / CHANGE

Reduces unnecessary diffs

Prioritizes surgical edits instead of broad rewrites, refactoring, or unrelated cleanup.

03 / REVIEW

Makes decisions traceable

Connects every changed line to a specific requirement, making review faster and clearer.

04 / DELIVERY

Creates predictable outcomes

Defines how to verify the result, report the change, and stop once the task is complete.

Control Parameters

Each rule controls a specific execution decision

These are operational parameters rather than model-generation settings. They determine how the agent behaves around the work.

ParameterConfigured valueMeaning and function
priority_orderuser_request_first

Resolves conflicts by placing the latest user request above host guidance, project conventions, and general judgment.

scope_boundaryexplicit_task_only

Prevents the agent from adding features, fixing adjacent issues, or expanding the task through assumptions.

inspection_depthdirectly_related

Allows only the files and sections required to understand and complete the requested change.

change_sizeminimum_required

Keeps the implementation simple and preserves existing structure, naming, style, and functionality.

dependency_policyno_new_by_default

Avoids introducing packages or configuration changes unless the task explicitly requires them.

verification_levelproportional

Matches the check to the change: manual review for copy, targeted checks for code, and focused proof for bugs.

stop_conditionresult_complete

Ends execution immediately after the requested outcome is achieved, preventing extra work.

report_schemastructured_summary

Reports changed files, the exact change, its purpose, and what still requires manual review.

Execution Logic

From request to verified result

The prompt creates a repeatable decision path instead of leaving each task to unstructured interpretation.

01Request

Identify the latest explicit outcome.

02Interpret

Resolve priority and choose the smallest safe reading.

03Inspect

Read only the directly related context.

04Change

Make the minimum required implementation.

05Verify

Use one proportional, targeted check.

06Report

State what changed and what to review.

07Stop

End when the requested result is complete.

Strengths

What the prompt improves

  • More focused implementation
  • Smaller and safer changes
  • Faster human review
  • Clearer delivery reports
  • Consistent behavior across tasks

Tradeoffs

What it does not replace

  • Clear requirements from the user
  • Project-specific technical documentation
  • Tests for high-risk application logic
  • Human judgment for ambiguous decisions
  • Exploratory work when broad discovery is required

Practical Result

A prompt that behaves like an operating contract

The value comes from translating general expectations into rules that can guide each inspection, edit, verification step, and final report.