AI-Engineering-Coach system prompt

--- id: instruction-bloat name: Instruction Bloat group: prompt-quality severity: medium scope: sessions requiresIdeContext: true version: 1 tags: [tokens, instructions, context, cost] thresholds: maxBytes: 4000 minBloated: 1 --- # Description Detects oversized `.github/copilot-instructions.md` (or equivalent) files. Custom instructions are prepended to **every** request's system prompt — large files inflate input tokens on every turn, often by thousands of tokens that are never relevant to the current task. # When Triggered {{extra.bloatedSessions}} workspace(s) have custom-instruction files larger than {{thresholds.maxBytes}} bytes (largest: {{extra.maxBytes}} bytes). Every request in those workspaces pays the bloat cost on input tokens. # How to Improve Trim `.github/copilot-instructions.md` to the essentials: language/framework conventions, code style, "do not" rules, and pointers to longer docs. Move long examples and rationale into separate files referenced via `#file:`. Keep the always-on payload under ~4 KB. # Examples {{extra.maxBytes}} bytes — largest custom-instructions file across {{extra.totalSessions}} workspace(s) # Detection Logic ```detect scan: sessions match: true aggregate: count stats: instructionBloatStats(allSessions, thresholds.maxBytes) bloatedSessions: stats.bloatedSessions maxBytes: stats.maxBytes totalSessions: stats.totalSessions withInstructionsCount: stats.withInstructionsCount emitCount: stats.bloatedSessions emitTotal: stats.totalSessions check: stats.bloatedSessions >= thresholds.minBloated examples: {{maxBytes}} bytes (workspace-level) ```

--- id: mcp-tool-bloat name: Tool / MCP Bloat group: tool-mastery severity: medium scope: sessions version: 2 tags: [tools, mcp, context-window] thresholds: maxToolsPerSession: 40 minSessions: 3 --- # Description Detects sessions with an unusually large number of distinct tools invoked — a proxy for oversized tool catalogs that inflate every request's system prompt. Each registered tool adds tokens regardless of whether it's used. # When Triggered {{count}} session(s) used more than {{thresholds.maxToolsPerSession}} distinct tools. Large tool sets add silent overhead to every prompt. # How to Improve Trim the active toolset: disable rarely-used MCP servers, scope tool sets per workspace, and use tool groups so only relevant tools are loaded for the task at hand. Aim for under 40 active tools per session. # Examples {{flatUnique(reqs, "toolsUsed")}} tools in one session # Detection Logic ```detect scan: sessions match: flatUnique(reqs, "toolsUsed") > thresholds.maxToolsPerSession aggregate: count check: count >= thresholds.minSessions examples: {{flatUnique(reqs, "toolsUsed")}} tools ```