Content
# Tool List
A portable AI Agent workflow setup. Distill the "expensive high-end model's one-time judgment" into fixed specifications, allowing any low-end model (Sonnet / Haiku / Gemini, etc.) to take over the project: **no token waste, no hallucinations, know when to ask humans, and follow a unified architecture**.
Main target platform: **Claude Code**. Secondary: **Antigravity / Gemini CLI** (skill format is homogeneous, all `<skill-name>/SKILL.md` + YAML frontmatter; Antigravity proof see its [official skills documentation](https://antigravity.google/docs/skills), verified in 2026-07).
## Design Philosophy (explained in one paragraph)
The correct usage of strong models is to produce "strategies" rather than "answers": planner-executor division of labor, rules written as mechanically checkable criteria and positive and negative examples, output with built-in acceptance criteria, and verification always executed by fresh context. This setup is the distilled, persistent asset - low-end models don't need to become smarter, they just need to follow and honestly report where they get stuck.
## Installation (Node >=18, cross-platform including Windows)
```bash
# One-line installation: run in the target project root directory (no clone required; tested on 2026-07-06)
npx --yes github:M4ng0D0g/agentic-workflow-setup
# Or clone and run (can add --target <path>)
node <this-repo-path>/setup.mjs
```
Installer behavior: detect `.claude/`, `.agents/`, `.gemini/` (if none, default to creating claude+agents) → copy skills and delegation templates → entry files (CLAUDE/AGENTS/GEMINI.md) **do not exist** and are created from templates → MCP configuration **non-destructive merge** (only add missing server, automatically fill `MEMORY_FILE_PATH`, and Windows automatically packages `cmd /c`) → write lockfile (version + file hash). Re-run is idempotent; locally modified files are upgraded and preserved with warnings.
Common parameters: `--dry-run` (only show plan), `--verify` (self-test after installation), `--uninstall` (only remove files not modified locally), `--platforms claude,agents,gemini`, `--force` (overwrite local modifications).
Manual installation (when not running scripts): `cp -r skills/* .claude/skills/` (or `.agents/skills/`) + copy `templates/delegation-prompts.md` and corresponding entry templates; MCP see [mcp/MCP_SETUP.md](mcp/MCP_SETUP.md).
### Platform Matrix
| Platform | Skills Path | Entry File | MCP Configuration | Verification Status |
| --- | --- | --- | --- | --- |
| Claude Code (project) | `.claude/skills/` | `CLAUDE.md` | `.mcp.json` | ✅ tested |
| Antigravity (workspace) | `.agents/skills/` | `AGENTS.md`/`GEMINI.md` | `.agents/mcp_config.json` | ✅ official documentation + external testing |
| Codex CLI | `.agents/skills/` (shared with above, Codex scans up directory tree) | `AGENTS.md` | `~/.codex/config.toml` (manual) | ⚠️ external report, not self-tested |
| Gemini CLI | same as `.agents/skills/` content can be manually copied | `GEMINI.md` | `.gemini/settings.json` | ✅ configuration format tested |
| Copilot / Cursor | not supported | — | — | ❌ not verified |
## Content Map
| Path | Content | Reader |
| --- | --- | --- |
| `skills/agent-entry/` | **Single Entry Point**: role determination, task routing, hard rules | all agents, must read |
| `skills/work-protocol/` | Execution Discipline: reporting contract, retry limit, anti-hallucination checklist, token economy | all agents |
| `skills/human-intent/` | Human Intent Understanding and Questioning Discipline + **Human-side Work Specification** | all agents + humans |
| `skills/task-planning/` | Planning Tasks: plan structure, acceptance criteria writing, approval gate | during planning |
| `skills/task-development/` | Development Tasks: screen/data/network layer architecture, design pattern unification, code discipline | when writing code |
| `skills/task-delivery/` | Output Tasks: docs, architecture diagram (mermaid), version control, delivery format | during delivery |
| `skills/model-dispatch/` | Command Side: model ladder, upgrade/downgrade, delegation triplet, verification not self-verification | main models with subagent capabilities |
| `templates/` | CLAUDE.md / GEMINI.md project entry templates, delegation prompt fill-in templates | during installation |
| `mcp/` | Three-platform MCP configuration examples and tutorials (Windows traps including `cmd /c`) | during installation |
| `docs/maintenance.md` | System Maintenance Agreement: who can modify what, lessons written back to where, expansion limit | agents wanting to modify rules |
## Three Irreducible Bottom Lines
Complete hard rules are six in `agent-entry` (they are detailed implementations of these three and can be adjusted when customizing projects); however, the following three cannot be deleted or weakened under any customization:
1. **Completion = Evidence**: only output test output text/exit code/source reference is considered completion, "should be able to" is not completion.
2. **Verification not Self-Verification**: acceptance is done by fresh-context agent, prompt does not carry author's conclusion, ask "find problems" not "confirm no problems".
3. **If Not Found, Mark It**: "not verified" is always cheaper than fabrication.
## Origin
Established by Claude Fable 5 on 2026-07-05 (product of planner-executor distillation experiment), rules come from real projects (Planist) after two rounds of adversarial review and revision. License: MIT.
