RoundTable

Autonomous AI team workflow for Claude Code, Codex, Gemini, Cursor, VS Code, and long-running tmux sessions.

RoundTable turns an AI coding session into a project-local product team. You give a brief, the team forms roles, challenges assumptions, locks one sprint, implements through agents/tools, reviews the output, verifies edge cases, and stops when the work is Ready For Human Test.

It is built for solo founders and builders who want an AI team that can work while they sleep, without losing visibility, memory, review discipline, or audit history.

Why RoundTable Exists

Most AI coding workflows fail in the same places:

• the model rushes from brief to code
• assumptions stay hidden
• security, QA, design, and product concerns are skipped
• long sessions lose context
• autonomous work happens silently
• "done" means "code was written", not "work was reviewed and verified"

RoundTable fixes that by making the AI operate as a visible team with logs, roles, challenge, signoff, and final human-test handoff.

The Trigger

The human says:

roundtable

The AI asks which project/folder to use, then initializes or loads:

<ProjectRoot>/RoundTable/

After that, meaningful work runs through the RoundTable protocol.

What It Does

• Forms a project-specific AI team before scoping.
• Uses roles like [CEO], [CTO], [Security], [QA], [Product], [Design], [Infra], and [AI Operator].
• Forces roles to challenge each other instead of silently agreeing.
• Locks one sprint at a time to avoid drift.
• Allows workers/subagents to implement, but requires RoundTable review before acceptance.
• Prevents the AI Operator from self-approving its own work.
• Makes QA a hard gate for Ready For Human Test.
• Makes Security a hard gate for security-relevant work.
• Keeps a live terminal waterfall of role discussion.
• Keeps timestamped audit logs for historical review.
• Produces simple human-facing handoff and test instructions at the end.

Install

Clone this repo, then install RoundTable into any project:

git clone https://github.com/naif824/RoundTable.git
cd RoundTable
bash install.sh /path/to/project

The installer creates:

/path/to/project/RoundTable/

It also adds managed RoundTable sections to:

• CLAUDE.md
• AGENTS.md
• GEMINI.md

Existing content is preserved.

Project Layout

RoundTable/
  README.md              # project-local protocol
  project.md             # selected project root
  environment.md         # BUILD_SANDBOX / STAGING / PRODUCTION_ASSISTED
  team.md                # active team and rationale
  scope.md               # locked scope and assumptions
  sprint.md              # current sprint
  tasks.md               # task queue
  backlog.md             # ideas outside the sprint
  live.md                # visible terminal waterfall
  current-status.md      # compact status
  resume.md              # recovery state after restart/compaction
  handoff.md             # simple human-facing final summary
  human-test.md          # simple human-facing test checklist
  audit/                 # timestamped chronological audit
  roles/                 # optional role notes
  bin/
    rt-log
    rt-watch
    rt-preflight
    rt-handoff

Live Terminal Waterfall

RoundTable is designed for tmux.

Run:

RoundTable/bin/rt-watch

The visible pane shows clean role dialogue without timestamps:

[CEO] We are forming the team before scope.
[Security] I need auth, secrets, and public exposure reviewed before signoff.
[QA] Ready for human test requires edge-case verification for the locked scope.
[AI Operator] Action: running preflight.

Every log line can be written with:

RoundTable/bin/rt-log CEO "We are forming the team before scope."
RoundTable/bin/rt-log Security "I need auth, secrets, and public exposure reviewed before signoff."

rt-log writes:

• clean visible text to RoundTable/live.md
• timestamped audit text to RoundTable/audit/YYYY-MM-DD-session.md

Core Rules

• Quality beats speed.
• Team formation happens before scope.
• One active sprint at a time.
• No role bypass.
• The AI Operator executes but cannot self-approve.
• RoundTable reviews every meaningful output before acceptance.
• QA controls whether a sprint is verified.
• Security controls signoff for security-relevant work.
• New ideas go to backlog.md unless needed for the locked scope.
• Human test is the final autonomous endpoint.

Ready For Human Test

RoundTable does not claim market launch.

The autonomous team can reach:

Ready For Human Test

That means:

• locked scope implemented
• RoundTable reviewed the output
• objections resolved or logged
• QA verification signed off
• Security signed off when relevant
• human-facing handoff.md and human-test.md are ready

Only the human can move the project to:

• Human Accepted
• Market Ready

Who This Is For

RoundTable is useful if you:

• use Claude Code, Codex, Gemini, Cursor, or VS Code Remote SSH
• run long AI sessions in tmux
• want autonomous AI work while you are away
• care about product quality, QA, security, and auditability
• want a solo-founder version of an AI product team

What This Is Not

RoundTable is not:

• an AI model
• a replacement for Claude/Codex/Gemini
• a project management SaaS
• a magic memory system
• a production-launch approval system

It is a project-local operating protocol and toolkit for AI-assisted autonomous delivery.

Test

bash tests/test_install.sh

Current Status

RoundTable is early but usable. The first version focuses on the installable project template, AI instruction files, visible live log, timestamped audit log, and core methodology.