English | 한국어

Second Claude Code

God Hands. Extra hands that find, analyze, plan, decompose, benchmark, and improve. The AI workflow that does not skip Check.

/scc:godhands finds, analyzes, plans, drafts, checks, and cuts. That is the whole product.

Second Claude Code is a knowledge-work plugin for Claude Code, Codex, and Grok. This tree is 16 skills, 19 command markdown files plus commands/version.mjs, 17 agents, 10 hook events, and 3 MCP servers (pdca-state, optional Playwright, optional MMBridge). The default auto-router is /scc:godhands, /scc:research, /scc:write, /scc:review, /scc:refine, and /scc:coach. Folded skills stay on disk; call them with /scc:. The one orchestrator is /scc:godhands. /scc:pdca is the slash-only compat name. /scc:workflow is named replay (slash) and /scc:batch is an independent parallel split (slash).

It does not auto-dispatch a skill from every prompt. Every skill sets user-invocable: false, so Claude Code routes a natural request from the commands/*.md frontmatter description, not from skills/*/SKILL.md. The routing eval suite in evals/ measures research, write, review, refine, coach, and analyze (24/24 on 3.1.2); a vague ask must not start a pass. Use a slash command when you need a predictable entry point. Codex exposes the 16 skills through .codex-plugin/plugin.json; Grok installs from .grok-plugin/ + walnut.manifest.yaml. Neither mirrors every Claude slash command.

This tree is 3.1.2. GitHub Latest Release is v3.1.2. Install from marketplace main or that release. Uninstall any leftover second-claude-code / old scc cache first.

Start here: User manual · Skill index · Command and document index · Architecture · Orchestrators

Install

Use the host that you actually run. All three pull main, which is 3.1.2. The GitHub Latest Release tag is v3.1.2.

Claude Code

claude plugin marketplace add unclejobs-ai/second-claude-code
claude plugin install scc

Start a new Claude Code session after installing. Check with:

claude plugin list

You should see plugin id scc. Commands look like /scc:godhands and /scc:write.

To update an existing v3 Claude install:

claude plugin update scc

Restart Claude Code after an update.

On Claude Code, both commands/ and skills/ register. 3.1.0 sets user-invocable: false on every skill so each /scc:* name appears once (the command).

Codex

codex plugin marketplace add unclejobs-ai/second-claude-code --ref main
codex plugin add scc@scc

--ref main is required so Codex follows current main (v3.1.2), not an older Release zip. Restart Codex after installing or refreshing the marketplace.

Codex uses .codex-plugin/plugin.json and plugin-relative MCP paths. It does not need CLAUDE_PLUGIN_ROOT. Request an installed skill by name, or describe the task; Codex does not mirror every /scc:* command.

To refresh an existing Codex install:

codex plugin marketplace upgrade scc
codex plugin add scc@scc

Grok

grok plugin install unclejobs-ai/second-claude-code --trust

--trust is required. This repository has .grok-plugin/plugin.json and walnut.manifest.yaml (Walnut listing, version 3.1.2). Grok installs from GitHub main or the v3.1.2 Release. After install, request skills by name the same way Codex does.

Migrating a pre-v3 Claude install

The v3 cache key is scc, not second-claude-code. If the old entry is listed:

claude plugin uninstall second-claude-code
claude plugin install scc

If second-claude-code is not listed, skip uninstall and run the normal scc install. Reopen the session before using /scc:*.

Durable data

Runtime data is runs, cycle memory, soul observations, workflows, and plugin preferences. Project decision standards live in the project (.scc/), not here.

`` is CLAUDE_PLUGIN_ROOT when the host sets it. Otherwise it is the plugin root associated with the current skill — the directory two levels above that skill’s SKILL.md. Codex and Grok often omit the Claude-only variables; the skills resolve this host-neutral fallback before any script or state path.

The default .data/ directory sits inside the plugin install. Replacing the plugin deletes it. Point CLAUDE_PLUGIN_DATA at a path outside the install if you want that state to survive updates:

export CLAUDE_PLUGIN_DATA="$HOME/.scc-data"

Standards are unaffected: they are written under the project as .scc/standards//STANDARD.md.

MCP servers

The plugin ships 3 MCP servers:

The public surface is these 3 servers, not a separate tool catalog to install. pdca-state also exposes many tools for PDCA state, cycle memory, soul data, project memory, session recall, and advisory cross-plugin plans; those tools inspect and advise. They do not execute an external skill for you.

Choose an entry point

/scc:* is the Claude Code surface. In Codex and Grok, use the skill name in the same row.

Default auto-router: /scc:godhands, /scc:research, /scc:write, /scc:review, /scc:refine, /scc:coach. analyze is God Hands Gather (/scc:analyze); it is not a top-level chooser. Folded skills stay on disk — call them with /scc:.

These are easy to mix up. They are not substitutes.

Tools

One orchestrator plus two slash utilities

/scc:godhands is the only orchestrator. /scc:workflow is named replay (slash). /scc:batch is an independent parallel split (slash). They are not three equal orchestrators. Autopilot is a named /scc:workflow preset, not a rival orchestrator. God Hands still slash-chains /scc:research, /scc:analyze, /scc:write, /scc:review, and /scc:refine. If Draft mentions /scc:workflow, that is an explicit slash, not an auto-route.

Slash-only (still on disk): collect, discover, translate, batch, workflow, soul, loop, evolve.

write --skip-review vs God Hands Check

write runs an internal /scc:review (default quick preset) unless you pass --skip-review. That pass is part of the write skill, not a hook.

Hands Check is a separate review. The Draft phase is supposed to call write with --skip-research --skip-review so Check owns quality. If you invoke /scc:write and then wrap the same draft in /scc:godhands without skipping, you review twice.

The same rule applies when you chain them yourself:

• Direct /scc:write — leave internal review on, or skip it only when another step will review.
• /scc:godhands — Draft skips write’s internal review; Check reviews.
• /scc:workflow autopilot — a named /scc:workflow preset, not a rival orchestrator. It already passes --skip-research --skip-review on write, then runs /scc:review and /scc:refine. Do not add a God Hands Check on top unless you want a third pass.

--skip-research is appropriate only when real source material is already supplied (or Plan/workflow already produced it).

/scc:research "AI agent frameworks" --depth medium
/scc:write --format report --skip-research --input report-notes.md
/scc:write --format report --skip-research --skip-review --input draft.md
/scc:review draft.md --preset content
/scc:refine "raise this to 4.5/5" --file draft.md --max 3
/scc:godhands "AI agent market report" --depth deep
/scc:workflow run autopilot --topic "edge computing"
/scc:batch --topic "10-part series on AI infrastructure" --skill write --parallel 3

Research depth

research uses Jina Search as its primary search path. If Jina is unavailable or a page cannot be read, it falls back through WebSearch/WebFetch and the unblock or Playwright paths when available. Credentials, rate limits, and optional integrations affect which path is available; no particular provider or result count is guaranteed beyond the selected depth contract.

When the legacy WebSearch/WebFetch engine is explicitly selected or Jina is unavailable, the equivalent search/read calls use that fallback engine. See the research guide for source validation and gap handling.

The 16 skills and 3 tool-only commands

The skills make judgments or produce work. The three tool-only commands execute a utility and do not occupy a skill slot. skills/unblock/ still ships the fetch engine; it has no SKILL.md.

Auto-router: coach, godhands, refine, research, review, write. God Hands Gather (not a chooser): analyze. Slash-only: batch, collect, discover, evolve, loop, soul, translate, workflow. Slash-only compat: pdca.

The skill guides contain command syntax, options, examples, and limitations. Slash-only skills set disable-model-invocation: true; their command descriptions stay visible, but call them with /scc:. Maintainer-only loop and evolve stay in that slash-only set. The public /scc:loop command optimizes prompt assets against a fixed benchmark suite; it is a direct maintainer entry point.

Review presets

Review panels contain 2–5 specialized reviewers. The preset controls both reviewer coverage and the vote threshold; a Critical finding blocks the result regardless of votes.

The built-in thresholds are quick 2/2, the 3-reviewer presets 2/3, academic 3/4, and full 3/5. These are consensus rules, not a promise that every review will finish or pass.

Example: /scc:review draft.md --preset academic.

Hands and its gates

Hands is an explicit orchestration path, not a requirement for every request. Runtime state and MCP tools stay pdca_*.

Gather → research and analyze a brief
Draft  → produce the artifact from the approved plan
         (write --skip-research --skip-review)
Check  → review with the selected 2–5 reviewer preset
Cut    → route gaps to Gather, Draft, or Refine

The runtime-enforced gate subset is the source of truth for whether a phase may transition:

Skill guidance is broader than this runtime subset. Research methodology, source quality, reviewer roles, format instructions, and writing style are instructions to the skill and its agents; they are not a prompt hook that dispatches work, and not every narrative condition is independently enforced by the state MCP server. Read the Hands guide for the public name and the PDCA guide for the engine contracts.

For code work, the Code Engineering Lane adds stricter acceptance, isolation, validator evidence, and handoff guidance on top of these same gates. It is a specialization of PDCA, not a second runtime.

Standards that survive sessions

/scc:coach can record a settled choice in your project:

.scc/standards//STANDARD.md

The record keeps the chosen direction, rejected alternatives, reopening conditions, and checks. Sessions can read active standards later, and /scc:standard-check artifact.md reports whether one artifact meets them. Checks print FAIL, UNPROVEN, or UNCHECKED; a clean run prints ok — N standard(s) checked. The checker does not invent a passing judgment for an adversarial check.

State, integrations, and background work

Lifecycle hooks restore state, report active-standard literal triggers, collect reviewer results, and save session summaries. They do not choose or invoke a skill from a user prompt.

Cross-plugin discovery is advisory: orchestrator tools inspect installed capabilities and return a ranked plan. They do not execute an external skill or command. Invoke any suggested external capability explicitly and review its data boundary first.

Background handoff returns a Claude Code command; it does not execute queued work itself:

claude --bg "/scc:workflow run weekly-digest"
claude agents

Configuration

There is no config file to place. Defaults live in each skill’s Options table (skills//SKILL.md) and are overridden per call with flags such as --depth, --preset, or --max.

Further reading

• User manual — step-by-step entry and examples
• Skill guides — one page per skill
• Command and document index
• Architecture — runtime boundaries and state model
• Orchestrator architecture — advisory routing
• Changelog · Issues

Issues and pull requests are welcome. Built by Unclejobs.

Version 3.1.2 | GitHub Latest Release v3.1.2 | MIT License