Blueprint Mode
Picks one of four fixed workflows (Main, Debug, Express, Loop) for a task, then executes it end to end with a numeric self-reflection rubric and a confidence threshold that decides whether to ask a clarifying question.
Overview
Blueprint Mode is a structured alternative to open-ended agent prompting. Instead of improvising an approach for every request, it classifies the task first, runs a matching workflow, and grades its own output against a fixed rubric before calling anything finished.
How it works
- It classifies the request into one of four workflows: Main for full feature work (analyze, design, plan, implement, verify), Debug for reproducing and fixing a specific bug, Express for small, low-risk changes touching two files or fewer, and Loop for repetitive changes across many files.
- It never assumes a library's API. It checks
package.json,requirements.txt, imports and neighboring files before using anything, and matches the project's existing style and conventions. - Once a workflow is chosen, it runs autonomously, batching independent reads and edits in parallel and only sequencing steps that depend on each other.
- Before declaring a task complete, it scores its own solution 1-10 on correctness, robustness, simplicity, maintainability and consistency. Anything scoring below 8 sends it back to design or implementation, up to three retries.
- It only asks the user a question when its confidence in interpreting the request drops below 90 out of 100. Otherwise it proceeds and reports a final status of completed, partially completed or failed.
Examples
"Add rate limiting to the /api/upload endpoint."
The agent classifies this as Main workflow, analyzes the existing middleware pattern, designs the rate limiter, implements it, then verifies and self-scores the result before reporting back.
Installation
You need VS Code with the GitHub Copilot extension.
- Create the agents folder and download the file into your repo:
mkdir -p .github/agents curl -o .github/agents/blueprint-mode.agent.md \ https://raw.githubusercontent.com/github/awesome-copilot/main/agents/blueprint-mode.agent.md - Reload VS Code, or open a new Copilot Chat; custom agents under
.github/agents/are picked up automatically. - Pick this agent from the mode picker in Copilot Chat and give it a task to confirm it classifies it into one of the four workflows.
Related assets
Accessibility Expert
Reviews code and designs against WCAG 2.1/2.2, covering semantics, keyboard and focus behavior, forms, media and dynamic SPA updates, and ships framework-specific examples plus a CI setup for automated checks.
Address Comments Agent
Works through pull request review comments one at a time: applies the minimal fix for each, pushes back on comments that don't make sense, adds test coverage, and commits with a descriptive message before moving to the next comment.
ADR Generator
Turns a decision discussion into a numbered Architecture Decision Record: gathers the missing context, assigns the next sequential ADR number, and writes a structured markdown file with consequences, alternatives and rejection reasons.