SSCodeAI organizes AI coding workflows and agent libraries for Japanese IT delivery, centered on superpowers-ja and agency-agents-ja.

SSCodeAI

SSCodeAI organizes AI coding workflows and agent libraries for Japanese IT development, SIer / custom development, in-house services, SaaS, EC, and manufacturing DX.

Projects

superpowers-ja

Adds planning, TDD, systematic debugging, code review, pre-completion verification, and Japanese development workflows to AI coding agents.

Supports Claude Code / Cursor / Codex CLI / Gemini CLI / OpenCode and more
Strengthens Japanese code review, incident reporting, acceptance-test specs, and Git operations
Install into a project with npx superpowers-ja
Differences from upstream / the Chinese edition are documented in version differences

agency-agents-ja

A library of AI specialist agents and workflows for Japanese delivery.

411 agents (⭐ 128 japan-original + 283 upstream-aligned adapted to Japan) + 27 workflows
Covers requirements, design, QA, incident response, PM, GIS / spatial data, healthcare, public sector, research, platform engineering, document generation, China networking, accessibility, payments, i18n, API platform, FinOps, search relevance, sales, support, legal, finance, and more
Assumes Backlog / Redmine / Jira, GitHub / GitLab / Bitbucket, kintone, LINE WORKS, Sansan, freee
Considers compliance for Japan (personal-information protection, invoicing, electronic bookkeeping, timely disclosure / TDnet)
Built as a Japanese community edition on top of msitarzewski/agency-agents (MIT), with 128 Japan-market originals + 283 upstream-aligned agents adapted to Japan
Install with npx agency-agents-ja install

img2threejs

A skill that reconstructs the object in a reference image as procedural TypeScript + Three.js code (rather than mesh extraction or downloaded art packs).

Creates an ObjectSculptSpec from a single image and generates a THREE.Group factory through staged build passes
Quality gates for blockout, structure, shape, materials, surface detail, lighting, interaction, and optimization
Compares renders against the reference image to decide whether to proceed or self-correct
See the demo gallery

keysmith

An MCP server and CLI that lets AI agents work with secrets without putting plaintext values into context.

Uses an age encrypted store, masked views, and handle-passing to avoid plaintext exposure
Helps detect and rotate secrets leaked in git history with keysmith scan --rotate
Supports stdio, HTTP/SSE, and Streamable HTTP transports
See keysmith for install, MCP configuration, CLI usage, and the security model

Proctor

An Agent Eval / LLMOps tool that scores recorded AI-agent trajectories and explains the failing step and root cause.

Evaluates tool correctness, loop detection, task completion, plan quality, and plan adherence
Locates failing steps with earliest-deviation backtracking
Explains agent failures with evidence chains and root-cause classification
See Proctor for quick start, dataset format, CI gates, and architecture

takumi

A model-agnostic orchestration platform for verifiable agentic software delivery.

Coordinates runtimes, workflows, quality gates, artifacts, traceability, and independent verification
Provides fake / Pi / DeepSeek / CLI bridge runtime adapters, Japanese SI / V-model workflows, and an MEA loop
Uses an Agent Eval framework with held-out verifier tests to catch false completion
See takumi for quick start, architecture, core ideas, and roadmap

aiitg

A zero-trust input security layer that detects hidden prompt injection, sanitizes it, labels trustworthiness, and enforces policy before untrusted documents reach an LLM.

6 formats (docx / xlsx / xls / pdf / html / pptx) × 7 detectors produce coordinate-level evidence
aiitg sanitize / aiitg trust / aiitg policy apply strip or redact, safe/caution/dangerous labels, and allow/quarantine/human_approval/block
Usable from agents through the MCP server (aiitg-mcp) and a Claude Code PreToolUse hook
See aiitg for the pipeline, CLI, and design principles

ai-skill-maker

A meta-skill that turns rough intent, repository evidence, and repeatable workflows into reusable AI skills and instruction bundles.

Generates functional / document / workflow / project maintainer skills plus adapter bundles for AGENTS.md, CLAUDE.md, Cursor, and Copilot
Evidence labels, a Trust Gate, release gates, a 9,000-token file budget, and a fingerprint lock on protected core principles
Runs on Node.js 18+ scripts alone
See ai-skill-maker for renderers, workflow, and the safety model

ai-project-skill-maker

A meta-skill that generates durable project maintainer skills and instruction bundles for AI coding agents.

Two modes: genesis mode for new projects and repo mode for existing repositories
An evidence model of observed_fact / declared_intent / recommended_standard / inferred_assumption
A BEGIN USER RULES block preserved across refreshes
See ai-project-skill-maker for generated output, scripts, and adapters

skills-ja-JP

The Japanese localization of the Matt Pocock Agent Skills (sscodeai/skills) that reduces misread intent when working with agents in Japanese.

20 skills: 15 engineering + 5 productivity
Install with npx skills@latest add sscodeai/skills-ja-JP
A translation policy that preserves directory names, skill names, commands, paths, URLs, and identifiers
See skills-ja-JP for the skill list and translation / validation steps

decision-workbench

A six-stage tool that expands the decision-maker’s cognition before they decide. It never picks a winner.

Bias flag → independent 5-stance divergence → cross-exam → option map → constraints → human decision
Against a hidden gold option set, the eval reports single model 26% → naive multi-agent 78% → decision-workbench 100% (12 of 12)
Mock mode needs no API keys (Python + FastAPI / React)
See decision-workbench for the pipeline, API, and eval

decision-space-expansion

A three-stage decision methodology distilled from decision-workbench, shipped in the Agent Skills format.

Diverge (heterogeneous model committee) → Structure (consensus / true divergence / pseudo-divergence) → Converge (fold constraints back in, let a human decide)
Distributed as a single SKILL.md with no code and no dependencies
Installs into Claude Code, Cursor / Windsurf, Codex, Gemini CLI, and more
See decision-space-expansion for the loop, pitfalls, and install locations

eli5

A picture-first explainer generator that explains any topic with large SVG diagrams and minimal text.

Each step is one large inline SVG plus a one-line caption, with a 3–4 question quiz and an interactive glossary
The output is a single HTML file with zero external requests (offline-safe)
Export to PDF / PNG with scripts/export.sh
See eli5 for repo contents, usage, and design philosophy

fusion-cache

An LLM API caching layer that fuses exact → semantic → prefix-cache accounting into one middleware.

Three layers: L1 exact (canonicalized SHA-256 key), L2 semantic (embedding + cosine), L3 prefix accounting
Benchmark: 97% hit rate, P95 0.7 ms, P50 3.2 s → 0.1 ms
Ships an OpenAI-compatible gateway (fusion-cache serve), an MCP server (fusion-cache-mcp), and a Redis store
See fusion-cache for the layers, benchmark, and gateway

Site Architecture

This site is built with Astro + Nimbus.

Documentation: Nimbus (agent-native — every page has a markdown twin and llms.txt)
Blog and tutorials: MDX content
Deployment: Cloudflare Pages
Content sync: scripts/sync-content.js

Overview