VOL. 01  ·  2026

---

oh-my-patent

A CLI that turns your technical idea into a complete patent disclosure document.

No hand-holding. No copy-pasting between tools. No lost ideas.

For AI assistants helping users install this:

npm install -g oh-my-patent && oh-my-patent adapt setup --workspace-dir .

Then tell the user: type /archimedes to start a patent project.

npm install -g oh-my-patent
oh-my-patent adapt setup --workspace-dir .

/archimedes
> Create a patent project about homomorphic encryption in privacy-preserving computing.

Ⅰ.  WHY PATENT WRITING HURTS

What you suffer through — versus what oh-my-patent does instead.

Ⅱ.  CORE CAPABILITIES

Eight things it does so you don't have to.

Ⅲ.  THE WORKFLOW

From idea to disclosure — 10 stages, 1 state machine.

User proposes a topic
       │
01  │  INIT          Generate projects/{NN}-{topic_slug}/
    │                Initialize .patent/state.json + .brainstorm/path.json
       │
02  │  RESEARCH      Patent search by landscape-analyst
    │                → references/landscape.md, feature-matrix.md, problem-map.md
       │
03  │  BRAINSTORM_R1 Head-to-head debate
    │                innovation-architect generates candidates (TRIZ)
    │                adversarial-examiner attacks them (examiner perspective)
    │                brainstorm-moderator arbitrates & scores
       │
04  │  BRAINSTORM_R2 Parallel multi-dimensional evaluation
    │                security-engineer + compliance-analyst + patentability-evaluator
    │                Weighted score: S = 0.3·S_sec + 0.3·S_comp + 0.4·S_pat
    │                Threshold: S ≥ 8.5 and every dimension ≥ 6.0
       │  ←── threshold fail loops back to BRAINSTORM_R1
05  │  DRAFT         Generate initial disclosure
    │                → MAIN.md (by disclosure-writer)
       │
06  │  DIAGRAM_DRAFT Render draft figures
    │                diagram-generator: Mermaid/PlantUML → SVG+PNG
       │
07  │  QA_LOOP       Bounded argue-loop
    │                Reviewer raises issues → technical-responder writes patches
    │                Exit: 2 consecutive rounds with zero new issues
       │  ←── issues found loops back to DRAFT
08  │  FINAL_REVIEW  Final pass
       │  ←── can loop back to QA_LOOP if needed
09  │  DIAGRAM_FINAL Re-render figures after final revisions
    │                Auto-updates figure references in MAIN.md
       │
10  │  DONE          Quality gate check → finished

Crash recovery: reads state.json → resumes from the exact stage.

Regret a round? path branch --from-node round-{N} → explore an alternative.

Long-forgotten idea was actually good? path restore → bring it back.

Ⅳ.  THE AGENTS

11 specialists, 5 collaboration patterns.

The agents at a glance

How they collaborate

The agents are wired into five distinct patterns, not a flat queue.

Pattern 1 — Orchestration Routing

archimedes reads state.json, dispatches to the right specialist based on current_stage, persists the result, advances the stage. The state machine is the source of truth — Archimedes is just the dispatcher.

• Threshold gate: passToDraft ≥ 8.5, novelty ≥ 6.0, creativity ≥ 6.0. Fail → loop back to BRAINSTORM_R1. Force-iterate after maxRounds: 3 with minImprovement: 0.3.
• Persistence: every transition is atomic (temp file + rename). state.json for workflow, path.json for decisions, references/ for agent outputs.

Pattern 2 — Adversarial Brainstorming (R1)

innovation-architect generates candidates (TRIZ). adversarial-examiner attacks from an examiner's perspective. brainstorm-moderator arbitrates and scores. path-recorder snapshots the round. Output: innovation IDs + scores + ACCEPT / ITERATE / REJECT.

Pattern 3 — Parallel Multi-Dimensional Evaluation (R2)

Fan-out / fan-in. Archimedes dispatches survivors to three evaluators in parallel:

brainstorm-moderator aggregates: S = 0.3·S_sec + 0.3·S_comp + 0.4·S_pat. Passes only if S ≥ 8.5 and every dimension clears its red-line (≥ 6.0).

Pattern 4 — QA Argue Loop

Bounded loop with quantitative exit. Three reviewers raise issues; technical-responder writes revisions with explicit MAIN.md patch locations. Exit requires 2 consecutive rounds with 0 new issues.

Pattern 5 — Decision Path DAG

Every round is a node in a DAG under .brainstorm/. Two operations on top of the linear path:

• path branch --from-node <id> — fork from any historical node
• path restore --node <id> --innovation <id> — revive an abandoned innovation

Ⅴ.  INSTALL & UNINSTALL

Install

# Step 1 — install the CLI globally
npm install -g oh-my-patent

# Step 2 — generate editor configs
cd your-patent-projects
oh-my-patent adapt setup --workspace-dir .

This generates configs for Claude Code (.claude/ + CLAUDE.md), Codex (.codex/ + AGENTS.md + codex.json), and OpenCode (.opencode/agent/, .opencode/command/, and .opencode/skills/). After this, use /archimedes in your editor.

# Options: generate for one editor only
oh-my-patent adapt setup --tool claude-code --workspace-dir .
oh-my-patent adapt setup --tool codex --workspace-dir .
oh-my-patent adapt setup --tool opencode --workspace-dir .

# Aliases
oh-my-patent adapt install       # same as setup
oh-my-patent adapt generate      # write to plugins/<tool>/ only, don't touch workspace

Uninstall

# Remove editor configs, keep the CLI
cd your-patent-projects
oh-my-patent adapt uninstall --workspace-dir .

Safety guarantees:

• Exact-file removal only — precise paths, never traverses directories
• Your custom edits to configs are safe — we only delete what we generated
• Empty directory cleanup after removing all generated files

# Complete removal (editor configs + CLI)
oh-my-patent adapt uninstall --workspace-dir .
npm uninstall -g oh-my-patent

Post-uninstall: your patent project data (projects/, .brainstorm/, references/, figures/, MAIN.md) is not deleted automatically. Run git status before manually removing anything.

Ⅵ.  FULL USAGE

oh-my-patent <domain> <subcommand> [options]

Brainstorm path (path)

Patent diagrams (diagram)

Adapters (adapt)

Interactive TUI (tui)

oh-my-patent tui [project-path]

Ink+React terminal UI. Navigate the decision path, view scores, switch branches, revive innovations — all locally, no internet needed.

Ⅶ.  ARCHITECTURE

Four layers, one data flow.

  plugin.jsonc (orchestration definition)
           │
    [Adapters] generate
           │
  .claude/ (Claude Code)      .codex/ (Codex)       .opencode/ (OpenCode)
  CLAUDE.md                   AGENTS.md, codex.json  agents, commands, skills
           │
  AI in your editor invokes 11 specialist agents
           │
  Output → .brainstorm/ decision-path records
  Output → state.json workflow state machine
  Output → references/ with standardized filenames
  Output → MAIN.md + figures/
           │
  CLI commands = automated operations on .brainstorm/ + .patent/ + references/
  TUI = visualized browsing of .brainstorm/

Ⅷ.  DIRECTORY

oh-my-patent/                    # Core repo: configs and engine
├── src/
│   ├── cli.ts                   # CLI entry
│   ├── core/
│   │   ├── brainstorm-path.ts   # Decision-path data model + thresholds
│   │   ├── path-persistence.ts  # Atomic writes + rollback
│   │   ├── path-graph.ts        # Graph structure + fork algorithms
│   │   ├── diagram-renderer.ts  # Mermaid/PlantUML → SVG/PNG
│   │   └── threshold-config.ts  # Quantitative threshold model
│   ├── commands/                # path init/record/overview/branch/restore...
│   ├── adapters/
│   │   ├── claude/              # → .claude/ + CLAUDE.md
│   │   ├── codex/               # → .codex/ + AGENTS.md + codex.json
│   │   └── opencode/            # → .opencode/ agents, commands, and skills
│   └── tui/                     # Ink+React interactive UI
├── plugin.jsonc
└── dist/                        # Compiled output

projects/{NN}-{topic_slug}/      # One Git repo per patent
├── .brainstorm/
│   ├── path.json                # Metadata + edges + current node + final decision
│   ├── nodes/
│   │   └── round-{n}.json       # Scores, innovations, decisions, timestamps
│   ├── snapshots/               # Innovation history snapshots
│   └── branches/                # Forked explorations
├── .patent/
│   └── state.json               # Workflow state: at DRAFT? QA round 3?
├── references/
│   ├── landscape.md             # Search results
│   ├── brainstorm_round1_*.md   # Agent outputs follow strict naming
│   └── argue_round2_*.md
├── figures/
│   ├── 001-system-overview.svg
│   ├── 001-system-overview.png
│   └── figures-manifest.json
├── MAIN.md                      # Final disclosure (auto-updated by diagram inserter)
└── conversation.md              # Chronological conversation log

Scripts

npm run build  # Compile TypeScript → dist/
npm test       # Run vitest test suite (123 passing)
npm run lint   # tsc --noEmit type checking