Prompt Builder Skill

Primary entry point for prompt-engineering work. Create or update prompt, instruction, agent, and skill artifacts through the full execution, evaluation, research, and modification loop, and route refactor or analyze requests to the specialized skills. references/orchestration.md is the canonical reference for the phase loop, sandbox contract, subagent dispatch matrix, artifact paths, and cleanup contract; the specialized skills reference it, mirroring only the parts they need.

Follow the shared .copilot-tracking conventions.

Goal

Create or update prompt-engineering artifacts through the full execution, evaluation, research, and modification loop until the evaluation log shows no remaining issues. Use this skill for new prompt artifacts, improvements, cleanup, and related instruction updates. Route scoped cleanup against requirements to /prompt-refactor and read-only review to /prompt-analyze.

Flow

1. Confirm the target prompt artifacts and any reference files. When the request does not already name them, decide which artifact types it needs using the Choosing the Right Artifact Type section in prompt-builder.instructions.md, propose the breakdown, and confirm scope with the user. Then derive the sandbox topic and next run number using the deterministic contract in references/orchestration.md.
2. When the target prompt files already exist, run the execution and evaluation phase (dispatch Prompt Tester, then Prompt Evaluator) to establish their current state and inspect the evaluation log; when that baseline shows no unresolved issues, skip to the final response. When the target files do not exist yet, skip to step 3.
3. Research: create or update the primary research artifact at .copilot-tracking/research/{{YYYY-MM-DD}}/{{topic}}-research.md, and delegate to Researcher Subagent when the topics are independent.
4. Modify: dispatch Prompt Updater to create or update the prompt files and related instruction files from the evaluation findings and research, then review the updater tracking.
5. Run the execution and evaluation phase, then repeat steps 3-5 until the evaluation log shows no unresolved issues or until the remaining issues are documented explicitly.

Routing