Founder Stack

Engineering workflow for non-technical founders shipping real products with Claude Code.

This is the workflow used to take a real SaaS from blank repo to deployed beta in 13 days, solo, without writing a line of code by hand. It's opinionated, gate-driven, and refuses to let you skip the parts that matter.

The five stages each list the commands and agents that run inside them. Diamonds are gates — shell-enforced checkpoints the agent cannot talk past. Full version with stage-by-stage detail in docs/workflow.md.

What's in here

• The Workflow (v0.1, stable) — Engineering-Playbook.md plus 15 slash commands, 8 subagents, and 5 shell hooks that enforce a build cycle: intake → (challenge) → (design) → plan → execute → verify → ship. The optional design layer (/ux-wireframe → /ux-mockup with human approval) lands designs as a source of truth before frontend code is written; /frontend-build then converts the approved design into a brief + scaffold the main agent implements against (hard-gated on the approval marker).
• Autonomous Missions (v1 preview) — /mission scopes a goal, writes a validation contract for your approval, then dispatches a feature-worker (sonnet), scrutiny-validator (sonnet), and v0.1 auditors on a /loop tick — designed for overnight runs. Each mission runs in its own git worktree for filesystem isolation. Orchestration logic runs in your main agent thread (Opus recommended) — sub-agent nesting is blocked by Claude Code, so the slash commands execute procedure files directly rather than dispatching a meta-orchestrator agent. See docs/missions.md.
• Skills — optional reasoning packs, modeled on mattpocock/skills. Planned for v0.2; the v0.1 release ships the core workflow only. (Note: /grill and /zoom-out are slash commands in workflow/commands/, not skills — they live with the workflow because they're user-invoked, not delegated sub-tasks.)
• Templates — parameterized CLAUDE.md and project.json for new projects.
• Init script — interactive setup that asks 8 questions and writes your project config.

Quickstart (60 seconds)

git clone https://github.com/rajconnects/founder-stack ~/founder-stack
cd your-new-project
git init   # if not already a repo
~/founder-stack/scripts/install.sh
~/founder-stack/scripts/init-project.sh

install.sh symlinks the workflow into your project's .claude/ and wires the framework's hooks into .claude/settings.json automatically. init-project.sh walks you through generating a project.json and a starter CLAUDE.md.

Open Claude Code in that directory and try /spec-intake to begin.

Installing into an existing repo, or one that already has another harness or global commands? Read docs/install.md — it covers the three deploy scenarios and what to consider for each.

Two ways to drive

Mode	Command	When
Control mode (v0.1, default)	/spec-intake → /test-gate → /design-gate → /handoff	You drive each gate. Hands-on, safety-net through process. Right when you're new to the framework, or when scope is exploratory.
Autonomous mode (v1 preview)	/mission "<goal>" → /loop /mission-tick <id>	The /mission* slash commands run an orchestration procedure in your main agent thread: scope the goal with you, write a validation contract for your approval, then dispatch workers and validators on a /loop tick until done. Right for overnight runs on crisp, well-scoped features. Run on Opus — orchestration logic inherits your session model, and the planning/retry passes are noticeably weaker on Sonnet.

Both modes coexist — install v1 additively on top of v0.1 with ~/founder-stack/scripts/install-v1.sh. v1 reuses v0.1's gates internally — the tick procedure dispatches design-auditor and schema-analyst in parallel with scrutiny-validator when applicable, so there's one source of truth per check.

Documentation

Start with Quickstart if you're new. Skim Workflow and Gates before your first phase. Read Install when going beyond a fresh repo. Read Missions when you're ready for autonomous mode.

Doc	When to read
docs/quickstart.md	Five minutes from clone to first spec. The fresh-repo path.
docs/workflow.md	The five-layer build cycle. Diagram + per-layer reference.
docs/gates.md	What each gate (/test-gate, /design-gate, /schema-gate, /deploy-gate, /publish-gate) actually checks.
docs/install.md	Three install scenarios (fresh / existing / harness-already-present), verification snippets, gotchas.
docs/missions.md	Autonomous mode (v1 preview) — orchestrator, workers, validators, the contract, overnight runs.
docs/decision-traces.md	What deserves a trace, the schema, lifecycle, anti-patterns.
docs/session-hygiene.md	When to /clear vs /compact between layers.
docs/coordination.md	Multi-terminal/multi-session protocol via coordination.json.
docs/for-non-tech-founders.md	The framing essay — what process buys you when you can't read the diff.
CONTEXT.md	Why this framework exists. The thesis in three principles.
CHANGELOG.md	Release log with rationale, tradeoffs, and what each release deliberately didn't do.
workflow/Engineering-Playbook.md	The portable spec that travels with .claude/ into installed projects. The reference.

Why "for non-technical founders"?

Most agentic-coding tutorials assume you can read the diff and tell when the agent is bullshitting. Non-technical founders can't — so they need process instead of judgment as the safety net. Gates, traces, and single-track days are how you ship without the codebase rotting.

Read the framing essay →

Editor support

V1 ships Claude Code-first. The primitives (slash commands, subagents, hooks) are Claude Code-native. Cursor/Copilot/Pi don't have first-class equivalents; ports lose the enforcement layer. A Cursor rules-only pack may follow if there's demand.

Inspiration & credit

Heavily inspired by mattpocock/skills. Founder Stack adds the workflow layer (gates, coordination, traces, handoffs) on top of the skills idea, framed for non-technical founders.

License

MIT