Skip to content

server_src_prompts_md.instructions.md

Latest commit

 

History

History
44 lines (32 loc) · 2.62 KB

server_src_prompts_md.instructions.md

File metadata and controls

44 lines (32 loc) · 2.62 KB
applyTo server/src/prompts/*.prompt.md
description Instructions for MCP-server-hosted workflow prompt files.

Copilot Instructions for server/src/prompts/*.prompt.md prompt files

PURPOSE

This file contains instructions for working with workflow prompt files in the server/src/prompts/ directory. These prompts are registered by the MCP server (via workflow-prompts.ts) and exposed as slash commands in VS Code Copilot Chat.

REQUIREMENTS

  • ALWAYS start each prompt file with the following YAML frontmatter block:

    ---
agent: agent
---

    Note: VS Code has deprecated mode: agent in favor of agent: agent.

  • ALWAYS ensure each prompt has a corresponding registration in server/src/prompts/workflow-prompts.ts, including a Zod parameter schema and a server.prompt() call.

  • ALWAYS ensure the prompt name in WORKFLOW_PROMPT_NAMES matches the registration.

  • ALWAYS write prompts that work in any environment where codeql and the MCP server tools are available, including terminal-only environments without an IDE.

  • ALWAYS use explicit, numeric tool parameters (e.g., file_path, line, character) instead of IDE-centric language like "position the cursor" or "click on".

  • ALWAYS document when MCP tools use 0-based positions (all codeql_lsp_* tools) versus 1-based positions (find_predicate_position, find_class_position, read_file).

  • ALWAYS note the workspace_uri requirement for LSP tools: it must be a plain directory path to the pack root containing codeql-pack.yml, not a file:// URI.

  • ALWAYS run npm run tidy from the repo root directory to apply (markdown) linting for all prompt files.

PREFERENCES

  • PREFER referencing other prompts by their registered MCP name (e.g., codeql://prompts/ql_lsp_iterative_development) to enable cross-prompt navigation.
  • PREFER including a "Worked Example" section showing concrete tool invocations with realistic parameter values.
  • PREFER a validation tools comparison table when multiple tools can validate queries at different fidelity levels (e.g., validate_codeql_query vs codeql_lsp_diagnostics vs codeql_query_compile).
  • PREFER documenting tool limitations discovered through actual usage (e.g., codeql_lsp_diagnostics cannot resolve imports; find_class_position finds class only, not module).

CONSTRAINTS

  • NEVER leave any trailing whitespace on any line.
  • NEVER use four backticks to nest one code block inside another.
  • NEVER assume the calling LLM has access to an IDE, cursor, or editor UI.
  • NEVER reference file:// URIs for the workspace_uri parameter — use plain directory paths.