eleven

M-theory needs eleven dimensions to unify five string theories into one. eleven unifies eleven roles into one AI team.

eleven is a folder of Markdown files you drop into any project. It turns your coding agent into a software team: product, design, architecture, backend, frontend, mobile, QA, DevOps, security and docs, led by a CTO who runs the loop. You stay the CEO.

• One folder, invisible to git. It lives in eleven/ and hides itself through .git/info/exclude.
• No server, no dependencies. About 25 Markdown files you can read in fifteen minutes.
• Any agent. Claude Code, Codex, opencode, Cursor… anything that can read files and follow instructions.

Quick start

Clone the kit once, then install it into any project.

macOS / Linux

git clone https://github.com/AbdullhRtima/eleven.git ~/eleven
~/eleven/install.sh ~/code/my-project

Windows (PowerShell)

git clone https://github.com/AbdullhRtima/eleven.git $HOME\eleven
~\eleven\install.ps1 C:\code\my-project

If scripts are blocked: powershell -ExecutionPolicy Bypass -File $HOME\eleven\install.ps1 C:\code\my-project

The installer copies the kit to my-project/eleven/ and adds /eleven/ to .git/info/exclude, your local ignore list. Your .gitignore is untouched and nothing shows up in git status.

In a terminal, the installer asks once whether you want eleven on your phone (paid, beta). The default is no. Add --with-sync to say yes up front, or --no-sync to skip the question; scripts and CI never get asked.

Then tell your agent to lead. Start it in your project and say:

Read eleven/AGENTS.md and act as the CTO. Fill eleven/project.md from the codebase.

Then, for each piece of work:

Read eleven/AGENTS.md and act as the CTO. New item 001: <your idea>. Start at step 0.

To pick up where you left off: Read eleven/AGENTS.md and act as the CTO. Continue item 001.

One command pulls the newest kit and refreshes your project:

macOS / Linux

~/eleven/install.sh --update ~/code/my-project

Windows (PowerShell)

~\eleven\install.ps1 --update C:\code\my-project

-Update works on Windows too, and the flag can go before or after the folder. The update runs git pull --ff-only in the kit first. If that fails (no network, or the kit has diverged), it tells you why and carries on with the kit you have. If you edited the kit itself, it stops and asks you to commit or stash first.

It refreshes AGENTS.md, agents/, templates/, docs/ and ui/. Your project.md, work/ and decisions/ are never touched.

Phone sync (optional). Say yes to the question, or pass --with-sync, and the installer copies eleven-sync into eleven/.sync-bin/, inside the project, with an eleven-sync launcher (eleven-sync.cmd on Windows). It does not use npm, sudo or a global install, and it never writes inside your clone of the kit, so it no longer depends on where your clone is. It needs Node 22+; without Node you get a message and the kit still installs. The choice is saved in eleven/.sync, so every --update refreshes the copy.

How it works

Every piece of work is an item with its own folder: the ledger. The CTO picks one task for one role. That role runs in a fresh context, does the task, and writes a report. QA can veto "done". You approve the brief and accept the result.

The guards keep the loop honest:

Guard	Rule
Round budget	At most 10 manage → execute rounds per item, then one final best attempt, marked unverified, and a report to you.
No progress	The same task to the same role twice? Stop and ask you. Failed twice? Try a different approach.
Last green	Work that passed its checks is never overwritten by something that doesn't pass.
No repro, no rewrite	A bug claim needs a command that reproduces it before anyone rewrites working code.
Small memory	plan.md ≤ 4,000 chars, notes.md ≤ 8,000. Rewritten, never appended.
Open questions	They pause the loop. Nobody guesses about intent, money, data or security.
Evidence	Reports paste the real build and test output. "Should work" isn't evidence.

The full manual is AGENTS.md. It's short on purpose.

The team

Each role is one short Markdown file: what it does, what it never does, and when it's done.

	Role	Owns
★	You, the CEO	the why; you approve the brief and accept the result
1	CTO	the loop, the plan, who works next, merging
2	Product manager	the brief, scope, acceptance criteria
3	Designer	flows, screens, states, copy
4	Architect	structure, interfaces, data, decisions
5	Backend	APIs, domain logic, data
6	Frontend	web UI
7	Mobile	iOS / Android / cross-platform
8	QA	verification; can veto "done"
9	DevOps	build, CI, deploy, observability
10	Security	threats, secrets, dependencies

Not every item needs everyone. A one-line bug fix is CTO → engineer → QA.

Watch it work

Open eleven.rtima.org/app (or eleven/ui/index.html locally) in Chrome or Edge and choose your project's eleven folder. It reads the ledger in your browser while your agents write it, refreshing every 2 seconds. Nothing is uploaded. Firefox and Safari open a snapshot instead.

There's also a file tree with a reader for every ledger file; HTML mockups open in a sandboxed frame. The screenshots use a synthetic sample project.

Why it works

eleven's loop comes from GVS5H (Gao et al., arXiv:2608.26480): one model split into a manager and fresh single-task workers that share state only through a few size-capped files on disk, with a verifier that can veto "done".

On 100 hard LiveCodeBench problems	Result
Same model, alone → in the loop	66.8 → 92.4 (+25.6 points)
Orchestrated cheap model vs frontier model	93.0 vs 90.4
Cost of the orchestrated run	about 9% of the frontier model

Honest caveats: the loop uses about 3× the tokens of a single call. Not every model gained, and some got worse. Strong models gain least. eleven adds the roles a real product team needs on top of that loop. More in docs/working-with-ai.md.

Inside eleven/

AGENTS.md         the operating manual every agent reads first
project.md        facts about this project (filled on first run)
agents/*.md       the 11 roles
templates/        brief, plan, tasks, notes, report, decision, project
work/<NNN>-slug/  one folder per item: the shared ledger
decisions/        decision records that outlive an item
docs/             the habits and research behind it, the CTO role, the roadmap
ui/index.html     the live browser UI
install.sh        install / update (macOS, Linux)
install.ps1       install / update (Windows)

This repo also holds site/, the landing page at eleven.rtima.org.

Roadmap

Guardrail hooks, a reviewer role, slash commands and a v0.1.0 release are next. See docs/roadmap.md.

Contributing

Issues and pull requests are welcome at github.com/AbdullhRtima/eleven. Keep changes small and the kit plain Markdown. If you change the live UI, run node ui/test/run.mjs (headless Chrome) before opening a PR.

Licence

MIT © 2026 Abdullah Rtima