Product · Shadow Clone · multi-agent dev toolkit

A dev team, in Claude Code.

Free open-source toolkit that turns Claude Code into a coordinated dev team. It plans, builds, refactors, debugs, and reviews its own work.
01

The problem

One model doing everything makes a mess

Ask an AI assistant to build something substantial in one go and you get what people keep describing the same way: half-finished pieces, things quietly breaking, and no way to tell which part went wrong. It writes confidently either way.

The missing ingredient is not a better model. It is the thing a development team provides and a single prompt does not: an order to the work, a plan before the building, and somebody who checks it afterwards.

Shadow Clone supplies that structure. It is not one assistant trying harder. It is planning, building and review kept as separate jobs, the way a team would run them.

02

How it works

Plan, build and review, separately

Every mode runs the same shape. You type one command; this is what happens behind it.

  1. 01 Research
  2. 02 Plan
  3. 03 Build
  4. 04 Review
  5. 05 Done
Fig 02.1 One piece of work. Review sends it back once, it is rebuilt, and only an approval lets it through, up to three rounds.
  1. 01

    It researches before it writes

    The first pass does not touch your code. It reads what is already there (how the project is laid out, what it depends on, and what the existing work looks like) so that what follows fits the codebase instead of fighting it.

    Wave 1 of 3 · output written to .waves/wave-1/deliverables/

  2. 02

    It plans before it builds

    The work is broken into pieces the size of a single pull request, with the order worked out: what has to come first, what can run alongside it, and which pieces everything else is waiting on.

    Wave 2 of 3 · dependency-ordered, never a week-by-week schedule

  3. 03

    It builds in focused passes

    Specialists take one piece each rather than one model attempting everything at once. The number working at the same time is capped, because past a certain point parallel agents start colliding with each other.

    Wave 3 of 3 · hard cap of 5 concurrent specialists

  4. 04

    A second model reviews it before you hear it is done

    Finished work is handed to a different AI model, which reads the changes, the full text of every file touched, and your project standards. It returns one of four verdicts. Anything short of approval goes back to be fixed and re-reviewed.

    Up to 3 rounds · APPROVE / REVISE / BLOCK / ERROR

  5. 05

    Nothing ships silently

    If three rounds pass without approval, the outstanding problems are written to your project issue tracker and reported to you. The work is never quietly declared finished because the review ran out of attempts.

    Open findings logged to docs/audit/ISSUE_TRACKER.md

Commands
19
Engineering protocols
15
Waves per mode
3
License
MIT
03

Use cases

Start from where you are

Each of these is a real journey through the toolkit. The commands shown are the ones you would chain together, in order.

  • “I know what I want to build, but not how to build it”

    Start at the top and let the planning modes do the architecture. You describe the end state; it works out the phases, the order, and what depends on what, then builds it in reviewed passes.

    1. /sc
    2. /sc-plan
    3. /sc-feature
  • “Something is broken and I cannot work out why”

    The debugging mode forms a theory, tests it, and only then proposes a fix, rather than changing things until the symptom disappears. A finding without a reproduction is treated as a question, not a bug.

    1. /sc-debug
  • “This code works, but I am afraid to touch it”

    The refactor mode restructures with behaviour preservation as its prime directive, and the review pass checks the behaviour did not drift. Tests can be generated first if the surface is not covered.

    1. /sc-test-audit
    2. /sc-tests
    3. /sc-refactor
  • “I need to know whether this is safe to ship”

    The audit mode produces a blueprint of what to examine, then runs the security checklist against it. Every finding has to be backed by a reproduction or a failing test before it counts.

    1. /sc-audit
    2. /sc-echo
  • “It is slow and I do not know where the time goes”

    The optimisation mode measures before it changes anything, so improvements are demonstrated against a baseline rather than assumed from a rewrite.

    1. /sc-optimize
  • “I am one person and there is nobody to check my work”

    Paired review is the whole point. Turn it on and every finished unit of work is judged by a second model against your project standards before you are told it is done.

    1. /sc-echo
  • “I have inherited a codebase I do not understand”

    The research mode investigates and reports with cited evidence, and the docs mode writes the documentation the project should have had, aimed at whoever reads it next, not at a marketing page.

    1. /sc-research
    2. /sc-docs
  • “I need to sequence months of work, not one task”

    Roadmap mode lays out multiple milestones as a pipeline with prerequisites, parallel branches, and the nodes everything else is waiting on. Shaped by dependencies, never by dates.

    1. /sc-roadmap
    2. /sc-sprint
04

Proof

What people have built with it

“It gave a rig electrician the structure of a dev team, and that’s what got me off the ground.”

Darren Mitchell, Founder, Smart Maintenance Systems

Marine electrician · twenty years offshore · not a developer

What came of it, as they reported it

  • Built Smart Maintenance Systems, an AI-powered maintenance and safety platform for the maritime industry
  • Won a £50,000 Innovate UK Smart Grant
  • Accepted into NVIDIA’s Inception programme
  • Presented as a case study at Ocean Tech Expo
  • Completed the Tech South West Growth Forge programme
  • Launched SMS Workboat, a compliance system for small commercial vessels

I’m a marine electrician. Twenty years offshore, working on everything from crane barges to research vessels. I am not a developer, and not long ago I couldn’t build software to save my life.

What I had was twenty years in vessel maintenance, hitting the same issues and the same hurdles over and over again, across different vessels and different companies. Nobody ever fixed them properly. And somewhere along the way I worked out how I’d solve it, a way to bring AI in and actually make life easier for people like me, the ones doing the work. I knew exactly what needed building. I just had no way to build it.

What makes it work for someone like me is the way it orchestrates the job. I don’t know how to architect software. I don’t know the right order to build things in, how to catch my own mistakes, or what “good” even looks like. Shadow Clone handles that. It breaks the work into proper stages, plans before it builds, builds in focused passes, then has a second AI review the work before it tells you it’s done. When you’re on your own with no team, that review is like having someone looking over your shoulder who actually knows what they’re doing.

On my own with a raw AI, I got a mess, half-finished bits, things breaking, no idea why. Shadow Clone gave the work an order and a discipline I didn’t have and couldn’t have provided myself. It’s not just “AI writes code for you.” It’s AI working like a proper development team, planning, building, reviewing, so that someone with no experience ends up with something solid instead of a pile of broken parts.

If you’ve got something in your head you don’t know how to build yet, this is the tool that shows you that you actually can. That’s exactly what it did for me.

05

Reference

What's in it

A slash command is a short text shortcut you type into Claude Code (like /sc-plan) that puts the assistant into one focused mode. Shadow Clone ships one umbrella command and eighteen mode commands, grouped by what they are for.

System

Set up the toolkit, ask for help, and keep yourself honest.

/sc
Walks you through project setup. Asks Type / Stack / Team / Stakes, then writes AGENTS.md, CLAUDE.md, an issue-tracker scaffold, and a .waves/ directory.
/sc-help
Show the full catalog of every /sc-* command with status.
/sc-bootstrap
Verify the install is complete and walk you through finishing any missing pieces.
/sc-echo
Enter paired-review mode. A second AI model reviews every completed unit of work against AGENTS.md before you hear it is done.
/sc-update
Check for and apply Shadow Clone updates.

Planning

Plan at three altitudes: one milestone, a full project, or a multi-milestone pipeline.

/sc-sprint
Plan one milestone. Decomposes it into a PR-sized task graph with prerequisites and parallel branches.
/sc-plan
Plan a full project from zero to end state. Produces a multi-phase MASTER_PLAN.md.
/sc-roadmap
Sequence multiple milestones into a workstream pipeline with decision gates.

Execution

Do the work. Each mode keeps the assistant focused on one job.

/sc-feature
Build a new capability through dynamic waves, secure-by-design.
/sc-refactor
Safe restructure with behavior preservation as the prime directive.
/sc-debug
Systematic root-cause analysis and bug resolution with hardening.
/sc-optimize
Measure-first performance improvement with clear baselines and targets.
/sc-research
Investigation with cited evidence and actionable recommendations.
/sc-audit
Audit blueprint that chains into the security checklist scan.
/sc-quick-fix
Fastest path to a small, focused change. Minimal team, no waves.

Quality

Verify the work: tests, coverage gaps, docs that match reality, and a tidy repo.

/sc-tests
Generate meaningful tests for work currently in scope.
/sc-test-audit
Diagnose test-suite coverage gaps in an existing project.
/sc-docs
Generate or refresh documentation for work currently in scope.
/sc-cleaner
Tidy the repo by archiving files nobody needs anymore. Plans first, never deletes, fully reversible.
06

One command

Install

Already have Claude Code open? Paste this into your terminal. Then inside Claude Code, type /sc to start your first project.

Install: paste into your terminal
git clone --depth 1 https://github.com/Ignis-AI-Labs/shadow-clone.git
cd shadow-clone && bash bridge/install.sh && bash scripts/sc-doctor.sh

New to the terminal? The README walks through the full setup, including OS-specific dependencies, in plain English: step-by-step install guide on GitHub.

The source · MIT

Ignis-AI-Labs/shadow-clone

Every command, every protocol, every line of the bridge that wires it into Claude Code is in the repo. No accounts, no API keys, no cloud round-trip.

View on GitHub ↗
07

Before you install

Questions

  • Q-01Do I need Claude Code for this to work?
    Yes. Shadow Clone is a set of slash commands for Claude Code, the official Anthropic CLI. If you can run Claude Code, you can run Shadow Clone. There is no separate runtime to install.
  • Q-02Is it really free? Where is the catch?
    No catch. The toolkit is MIT-licensed open source. The entire source is in the public GitHub repo. There are no accounts, no telemetry, no paid tier, and no Ignis API key. (The optional paired review needs an API key for the reviewer model you pick.) Your Claude Code subscription pays Anthropic for the model; we charge nothing on top.
  • Q-03Can I use my own model or a different review model?
    The slash commands run against whichever model Claude Code is configured to use. The optional paired-review loop (/sc-echo) can route to GLM through OpenCode, or to Grok or Kimi through their own CLIs, and it can run two of them as independent lenses on the same work. You stay in control of which model sees your code.
  • Q-04Does anything I write get sent to Ignis AI Labs?
    No. Shadow Clone runs locally. The slash commands live at ~/.claude/commands/ on your own machine. We do not collect prompts, code, or usage data. The only network traffic is the call your Claude Code instance already makes to the model, plus, if you turn on paired review (/sc-echo), the work it sends to the reviewer model you chose.
  • Q-05Who is this actually for?
    Engineers building real projects with Claude Code who want structure instead of vibes. Shadow Clone earns its keep when the work is bigger than a single edit: when you would normally split it into design, build, review, and test phases. The slash commands give you those phases without you having to script them yourself.