A bootstrap wizard for Claude Code that generates a specialized agent workflow for your project — tailored to your stack, structure, and standards.
Summoner runs once on a new project. Through a guided conversation it analyzes your codebase and creates a set of prompt files (Claude Code slash commands) that form a complete agent workflow for tackling features:
/research → /design → /plan → /implement → /review-*
Each stage produces a document that becomes the input for the next. Code is written only at the final stage — everything before that is context preparation.
A generic "build me X" prompt produces unpredictable results because the agent makes all decisions at once: what to research, what architecture to use, how to split the work. Quality suffers.
Summoner creates project-specific prompts that guide agents through a structured process where each decision is made at the right time and reviewed by a human before moving forward.
Quality formula: (correctness + completeness) / (context size + noise)
| Stage | Agent does | Output | Human gate |
|---|---|---|---|
| Research | Scans the codebase for everything relevant to the task | Facts document with file/line references | — |
| Design | Creates architectural solution with diagrams (Mermaid), ADRs, test strategy | Design document | Review & approve |
| Plan | Breaks implementation into phases, each a committable unit with completion criteria | Phased plan | Review & approve |
| Implement | Writes code strictly following the plan, runs quality gates after each phase | Code + passing checks | — |
| Review | Three independent reviewers check quality, security, and design conformance | Issue lists | — |
Key principles (from context engineering:
- Research reports only facts — no opinions, no suggestions
- Design and plan are human-reviewed gates before any code is written
- Implementer follows the plan strictly — if it can't be done as written, it stops and reports
- Reviewers never fix code — they only point out issues
- Each prompt is specialized for your project's stack, patterns, and tooling
Summoner goes through 6 phases to generate the prompts:
-
Phase 0 — Understand the project. Subagents scan your docs and code independently. You describe the project in your own words. Summoner cross-references findings and resolves discrepancies.
-
Phases 1–4 — Generate stage prompts. For each stage (research, design, plan, implement), Summoner asks targeted questions about your project's specifics, validates answers against the codebase, and generates a specialized prompt.
-
Phase 5 — Generate review prompts. Creates three review prompts (quality, security, design conformance) tailored to your project's standards and requirements.
All generated files are saved to .claude/commands/ in your project.
.claude/commands/
├── research.md # Codebase investigation prompt
├── design.md # Architecture & diagrams prompt
├── plan.md # Implementation planning prompt
├── implement.md # Code writing prompt with quality gates
├── review-quality.md # Code quality review prompt
├── review-security.md # Security review prompt
└── review-design.md # Design conformance review prompt
/bootstrap
Answer the wizard's questions. It will scan your codebase, ask about your architecture, standards, and tooling, then generate all prompt files.
/research <task description>
# → creates research document
/design <task description>
# → creates design document, review it before proceeding
/plan <task description>
# → creates phased plan, review it before proceeding
/implement
# → implements phase by phase, quality gates after each
/review-quality
/review-security
/review-design
# → independent reviews after each implementation phase
- MANUAL.md — Project documentation. Explains the idea, how the wizard works phase by phase, how it differs from the agent system it generates, and what the output looks like.
- SKILL.md — The wizard prompt itself. This is the actual instruction that drives the bootstrap process — it tells the wizard what to do, what to ask, and how to generate each prompt.
- TODO.md — Backlog of potential improvements deferred in favor of a simple first version (context splitting, delta updates, task type specialization, subagent teams, etc.).
- Claude Code CLI
TBD