99 lines at a phase boundary
Only skill metadata is normally discoverable. The router loads when a meaningful phase starts, classifies the work and picks one primary specialist.
superforge
One thin router. Fourteen specialists. Less context.
Say superforge once at the start of a phase and describe the result you want. The router reads the task, inspects the project, sizes the work and loads only the specialist it needs — strategy, design, implementation, proof or release.
>/plugin marketplace add takaoumehara/superforge-skill
01Why a thin router
A front door, not a suite that loads all at once
You do not need to know the specialist names. superforge is the concierge; the fourteen skills are the workshop behind it. Each request loads the smallest useful route, and nothing more.
Fourteen, loaded on demand
Medium work loads one specialist. Large work loads them in sequence rather than all at once. A second specialist joins only at the boundary where it becomes necessary.
Work, proof, release
The specialist does the work. superforge-verify checks evidence before “done”. superforge-ship independently decides whether it may reach users.
Why it uses less context
- 01
Small work stays inline
A one-file fix loads no extra skill unless the task genuinely needs one.
- 02
Follow-ups reuse the active phase
Ordinary feedback continues the current route instead of re-reading the router.
- 03
Model guidance only for real dispatch
Per-subagent model tiers are read only when an agent is actually dispatched.
- 04
Files only when they matter later
Project state and run logs are written for corrections, failures and releases — not for small edits.
The saving comes from structure, not a promise about a vendor's price. Actual billing varies by tool, model, prompt caching and task.
02Specialists
Fourteen routes across five phases
Direct invocation still works for experts, but it is optional. Describe the outcome; the router picks the route.
| Phase | Specialist | Responsibility |
|---|---|---|
| Think | superforge-brain | Explore and judge product ideas |
| Think | superforge-biz | Market, pricing, business model and value case |
| Think | superforge-brand | Brand direction, voice and visual system |
| Think | superforge-roast | Expose weak assumptions before they reach users |
| Build | superforge-dev | Plan and implement multi-component features |
| Build | superforge-ui | Design and build Web, iOS and Android interfaces |
| Build | superforge-scroll | Cinematic scroll-driven experiences |
| Prove | superforge-a11y | WCAG, assistive technology and platform accessibility |
| Prove | superforge-test | Choose and implement valuable tests |
| Prove | superforge-debug | Root-cause debugging and failure memory |
| Prove | superforge-secure | Practical security review for small teams |
| Prove | superforge-verify | Attach evidence before any completion claim |
| Ship | superforge-ship | Decide whether a product may be released |
| Continue | superforge-handoff | Preserve session state before switching tools or threads |
03Install
Two lines in Claude Code
The plugin installs the router, all fourteen specialists and the optional dynamic workflows. Use one install path per tool, not both.
claude code
/plugin marketplace add takaoumehara/superforge-skill
/plugin install superforge-skills@superforgeRecommended. Open a new session afterwards and start a phase with superforge.
bash
git clone https://github.com/takaoumehara/superforge-skill.git
cd superforge-skill
./install.sh # --dry-run to preview, --update to refreshLinks every skill into the tool directories already on the machine, including ~/.agents/skills and ~/.codex/skills. On Windows use .\install.ps1.
bash
python3 scripts/package_skills.py --claude-web
# upload dist/superforge-claude-web.zip in Customize → SkillsOne ZIP holds the router and all fourteen guides. Dynamic workflows are not available in the browser; the bundle uses the prose fallback.
Where it runs
| Environment | Support |
|---|---|
| Claude Code | Skills plus optional dynamic workflows |
| Codex CLI / app | Skills through ~/.agents/skills or ~/.codex/skills |
| Gemini CLI / Antigravity | Skills through their discovered skill directories |
| Claude.ai | Standalone ZIP; no dynamic workflow runtime |
04How it works
Say it once per phase, then talk normally
Start a meaningful phase with superforge. Keep giving ordinary feedback after that. Say it again only when you begin a different feature, a release review or another substantial phase.
Start the phase
prompt
superforge — Redesign onboarding so a first-time
user can publish in three minutes.Then just give feedback
prompt
Keep the previous conditions. Make the empty
state clearer, then show one preview.Small, Medium, Large
Work inline
One issue, usually a few files. No extra skill is loaded unless the task genuinely needs one.
One primary specialist
One UI flow, an investigation or a coordinated multi-file change. The router loads exactly one specialist.
The minimum sequence
A new feature, broad redesign, security review or release. Specialists run in order, and verification comes before completion.
Three optional modes
| Mode | Use it for | Default behaviour |
|---|---|---|
superforge quick | One bounded correction | Handle inline; no orchestration or log by default |
superforge build | A feature or multi-file change | Plan, choose one primary specialist, implement, verify |
superforge ship | Release readiness | Verify evidence first, then run the independent release gate |
05FAQ
Frequently asked questions
Do I need to learn the fourteen skill names?
No. Say
superforge and describe the outcome. The router chooses the specialist. Calling a specialist directly still works if you prefer.Should I repeat “superforge” on every message?
No. Once per phase is enough. Follow-up feedback reuses the active phase. Start it again for a different feature, a release review or another substantial phase.
Does it make my model cheaper?
It does not change the price of the active model. It loads less context per request and assigns an appropriate model tier when it dispatches subagents. Actual billing depends on your tool, model, caching and task.
Will it pick a library for me?
In a substantial UI phase where the technique is still open, it checks your existing stack, scans current options from official sources once, explains the choice and implements it. If the native platform is the better answer, it adds no dependency.
Can it declare my app secure or correct?
No. It makes routing and evidence requirements explicit. Results are still limited by the available tools, credentials, source material and human decisions.
I installed both the plugin and install.sh. Why do skills appear twice?
The installer also links into
~/.claude/skills when it exists. Use one path for Claude Code — the plugin is recommended.