llm-tools/.grok/context.md

Grok CLI Context Guide for this Repository (llm-tools)

This document provides persistent context, best practices, and usage patterns for Grok CLI Agent mode when working in this repository. It is automatically included in context (via .grok/context.md convention or explicit inclusion) to keep interactions model-provider agnostic, efficient, and aligned with the tools/skills here.

Core Principles

• Provider Agnostic: All documentation, prompts, and workflows should work for Claude, Grok, or any MCP-compatible LLM. Avoid platform-specific assumptions (e.g., prefer MCP tools over Claude-only skills where possible). Update README.md, SKILL.md, and sub-READMEs accordingly.
• MCP-First: Leverage the built-in MCP tools exposed in Grok CLI (prefixed mcp_dicom-mcp__*, mcp_filesystem-mcp__*, mcp_playwright-mcp__*, mcp_selenium-mcp__*, mcp_github__*, mcp_us-weather-mcp__*, etc.). These map directly to the MCP servers in ./mcps/.
• Sub-Agent System: Use the task tool proactively:
  • general: Multi-step execution, editing, implementation.
  • explore: Fast read-only research, reviews, "how does this work?".
  • verify: Sandbox-aware builds, tests, validation.
  • computer: Desktop interaction (snapshots, clicks, typing, screenshots).
  • Prefer delegate for long-running read-only background work.
• Workflow Priorities (from system instructions):
  1. Understand the request.
  2. Decide on sub-agent for investigation (default to task for review/research/bug triage/verification).
  3. Use read_file, bash for direct exploration (small tasks).
  4. bash with background=true for servers/watchers.
  5. delegate for parallel read-only work.
  6. edit_file (surgical), write_file (new/full rewrites).
  7. Verify with reads, tests, builds.
  8. search_web/search_x for up-to-date info.
• Tool Preferences:
  • File ops: Dedicated read_file/write_file/edit_file over bash.
  • Edits: edit_file for targeted changes (unique old_string + context); write_file only for new files or full overwrites.
  • Background: Use for dev servers; manage with process_* tools.
  • Schedules: Use schedule_* tools for recurring tasks.
  • Images/Video: generate_image/generate_video.
  • Payments/Wallet: wallet_*, paid_request, fetch_payment_info.
  • Desktop: computer_* tools (snapshot first for stable @e refs).
  • MCP: Call directly (e.g., mcp_dicom-mcp__dicom_summarize_directory); see individual MCP READMEs for params.
• Context Management: This file + .grokignore keep context lean. Avoid loading binaries, large dirs, or node_modules. Explicitly include key files (README.md, SKILL.md, specific MCP docs) when needed.

Key Directories & When to Use

• mcps/: Launch MCP servers via scripts (e.g., bash mcps/dicom_mcp/run_dicom_mcp_server.sh). Use corresponding mcp_* tools in Grok.
  • dicom_mcp: QA for DICOM (Dixon, SNR, metadata, rendering, validation). Start here for medical imaging tasks. See mcps/dicom_mcp/README.md + docs/.
  • Others: filesystem (ops), open_meteo (weather), playwright/selenium (browser automation/testing).
• skills/: Reusable agent instructions. Read SKILL.md files for patterns (e.g., mcp-builder for new MCPs, artifacts-builder for UIs, document-skills for Office/PDF). Adapt YAML frontmatter + guidelines into prompts or custom tools. mcp-builder/SKILL.md is especially valuable for extending this repo.
• config/ & .claude/: Claude-specific; update paths in claude_desktop_config.json for cross-machine use. Ignore for Grok workflows.
• .mcp.json: Registers MCPs for Claude Code; useful reference for tool signatures.

Best Practices for Grok CLI in this Repo

• Exploration: task with explore agent first for reviews/research ("review this change", "how does MCP X work?"). Follow with general for edits.
• Verification: Always use verify sub-agent for builds/tests/smoke tests before/after changes.
• Editing: Quote sufficient unique context in edit_file. Verify by re-reading the file.
• DICOM QA: Prefer mcp_dicom-mcp__* tools (e.g., dicom_summarize_directory, dicom_compute_snr, dicom_render_image). Use dicom_read_pixels + dicom_compute_snr with ROIs identified via render.
• Browser Automation: Choose playwright-mcp for modern snapshots/eval; selenium-mcp for multi-session or specific WebDriver needs. Combine with computer_* if testing desktop apps.
• Avoid: Loading entire node_modules, large image/DICOM dirs (use .grokignore), or Claude-only phrasing. For long research, delegate to explore.
• Extending: Follow skills/mcp-builder/SKILL.md phases (research/planning → implementation → evaluation). Use mcp_github__* tools for PRs/commits if working on remote repos.
• Scheduling: Use schedule_create for recurring checks (e.g., daily DICOM validation or security scans).

Quick MCP Tool Examples (Grok CLI)

• Summarize DICOM dir: mcp_dicom-mcp__dicom_summarize_directory (directory=..., recursive=true).
• Filesystem write: mcp_filesystem-mcp__write_file (after setting default if needed).
• Browser: mcp_playwright-mcp__browser_navigate or mcp_selenium-mcp__selenium_navigate.
• Weather: mcp_uk-weather-mcp__get_forecast or open-meteo equivalent.

Refer to this file in prompts: "Follow the Grok CLI context in .grok/context.md, the updated README.md, and relevant SKILL.md files. Keep all outputs provider-agnostic."

Update this context.md as the repo evolves (e.g., new MCPs, refined workflows). For full tool list, inspect available tools or run grep -o 'mcp_[^ ]*' -r . --include="*.py" | sort | uniq.