agent-guidance

Files (1)

• SKILL.md

SKILL.md (first lines)

## When to use me
- Run this skill at the beginning of EVERY task, repository lookup, or codebase refactoring.
- Run this skill to check project conventions and avoid raw file reading/search operations.
- Re-run this skill at EACH phase transition (plan → implement → test → review).

## How to use me
You must invoke the 6 core `agent-guidance` MCP tools in this priority order:
1. Call `task_pipeline(task="...", project_path="<path>", phase="plan")` at the start of any task or phase to initialize workspace context, tree metrics, and architecture patterns.
2. Immediately call `guidance(operation="search", query="<user request>")` to discover and propose required domain skills. When skills are proposed, ask user via `ask_question(questions=[{question: "...", options: [...], is_multi_select: true}])` (format options as `"{name} - {short_desc}"` with short `{short_desc}` <= 38 chars) before loading them via `select_skills(skills=[...], user_confirmed=true)`. Do NOT inspect codebase files or start planning before completing skill discovery.
3. Call `project_context(operation="read", relative_paths=[...])` (or `relative_path="..."`, `target_symbol="..."`) for reading files (capped at 300 lines). Clustered read allows inspecting multiple related files in 1 turn. NEVER use native IDE file reading (VS Code `readFile`, Cursor `read_file`, Antigravity `view_file`) or shell commands (`Get-Content`, `cat`, `type`, python file reads).
4. Call `project_context(operation="search", query="...")` for searching codebase. NEVER use native IDE searching (VS Code `textSearch`/`findFiles`, Cursor `codebase_search`/`grep`, Antigravity `grep_search`/`find_by_name`) or shell search commands.
5. Call `workflow_gate(action="authorize_edit", project_path="...", relative_paths=[...], risk_level="LOW", justification="...")` in a single batch call (or directly via `workflow_gate(action="approve_plan", user_confirmed=true, relative_paths=[...])`). Calling `authorize_edit` file-by-file across multiple turns is prohibited. Individual `relative_path="..."` is reserved for ad-hoc discovered files.
6. Call `guidance(operation="search", query="...")` for coding standards or `guidance(operation="ui_ux", query="...")` for UI/UX design rules.
7. Call `guidance(operation="verify", verification_command="...", expected_output_keyword="...")` for empirical post-code testing.

## Critical Behavioral Rules
- When unsure about anything, trigger the IDE/CLI `ask_question` tool! DO NOT GUESS.
- Propose an implementation plan before making any big or complex changes.
- DRY & Reusable Code Mandate: Always reuse existing shared/common utilities, formatters, and UI components instead of creating duplicate logic.
- Every created or modified file MUST be < 300 LOC (aim for < 150 LOC per sub-module/component).
- For each new work phase, re-call `task_pipeline` with that phase's goal and request user confirmation for stage transitions.

From JunMystery/Agent-Guidance, MIT.

More from JunMystery/Agent-Guidance