Free · MIT · Claude Code template

Folder architecture for the one-person business: 1 orchestrator + 4 worker sandboxes.

One orchestrator. Several worker sandboxes. Briefs as the handoff contract between them. ICM-aligned context layers. Hooks as guardrails. Routines as maintenance. The template ships the structure — you fill in your brand.

Fig. 01 — Dispatch architecture
What you get

Six layers. Day-one structure. Your brand goes in.

The template ships the architecture. You fill in your specific brand content and content cadence. Folder tree, brief templates, hooks, skills, routine specs, and context layers are all wired before you write a single line of your own.

L1 / orchestrator + workers

One root. Several sandboxes.

One human-in-the-loop session at the root. Specialized sandboxes underneath — content-claude/, design-claude/, research-claude/, va/. Each sandbox is isolated; the orchestrator dispatches.

L2 / briefs

Briefs as the handoff contract.

Every dispatch is a cold-start brief a worker can read without prior chat history. Templates included for content briefs, design briefs, research scouts.

L3 / ICM context layers

L0 / L1 / L3 / L4, mapped.

Root CLAUDE.md (L0), root CONTEXT.md (L1), _config/ + sandbox references (L3), per-brief working artifacts (L4). ARCHITECTURE.md spells out why this template skips L2.

L4 / routines

Weekly + monthly maintenance.

Workers purge. Orchestrator audits. Operator validates. Monthly adds heavier compression and the archive sweep. Specs included.

L5 / hooks

Four Python hooks out of the box.

safety-check.py (rm -rf / force-push prompt) · snapshot-before-edit.py (start-of-day snapshots) · auto-stamp-date.py (auto-bumps Last updated: on trackers) · handoff-files-guard.py (prompts before editing handed-off files).

L6 / memory + skills

Memory index + 2 skills.

MEMORY.md with explicit slots for user, feedback, project, reference memories. Skills: diagnose-recurring-edits (aggregate breadcrumbs into source-level fix candidates) and verify-handoff (cross-check shipped output against the dispatching brief).

Sample brief

What a dispatch looks like.

A brief is a plain markdown file the operator drops into a worker's briefs/ folder. The worker reads it cold-start, executes, ships, and breadcrumbs the result in STATUS.md. No prior chat history required.

briefs / 2026-05-14 / sample-content-brief.md

# Dispatch — Carousel for Week 19, Tuesday slot

Worker: content-claude
Due: 2026-05-14 EOD
Format: 9-slide Instagram carousel, 1080×1350 each

Topic

  • The single bias most solo founders miss when they're pricing their first offer

Source material

  • operator's 2024 transcript on pricing — `_config/transcripts/2024-pricing.md`
  • three reader DMs flagged in `MEMORY.md` (project memory: pricing-questions)

Voice gates

  • No "transformative", "journey", "unlock", "game-changer"
  • Em-dash ≤ 1 per paragraph in body prose
  • Zero emojis in slides 1, 2, 9. Optional ≤ 1 in middle slides.

Deliverables

  • 9 slide-copy blocks (text only; design-claude handles visuals separately)
  • 1 caption block (250–400 words)
  • 5 hashtag block (1 anchor + 4 discovery)
  • Breadcrumb into STATUS.md on ship

Out of scope

  • Visual rendering (separate brief to design-claude)
  • Scheduling (operator-side)
Proof block

What others have built with it.

Placeholder by design. Real-user receipts go here after the first cohort of solo founders runs the template against their own business. No invented testimonials, no synthetic engagement metrics.
Setup

Four steps. About 30 minutes.

Clone the template, fill in your brand layer, customize the worker sandboxes you'll use, drop a first brief. The architecture works on day one; the content you put into it improves with each dispatch.

  1. Clone the template

    gh repo clone NFTYoginis/agency-of-one your-business-name — or download the ZIP. Open the folder in your editor of choice.

  2. Fill in your brand layer

    Edit CONTEXT.md (your business at a glance) and _config/brand.md (your voice, your audience, your offers, your pricing). Each file has TODO markers showing what to replace.

  3. Customize the worker sandboxes you'll actually use

    Inside each sandbox (content-claude/, design-claude/, research-claude/, va/), there's a CLAUDE.md with worker-specific instructions. Edit those to match your actual content needs. Delete sandboxes you don't need.

  4. Drop a first brief

    Pick a worker. Open its briefs/ folder. Create a new markdown file describing what you want shipped. The worker reads it cold-start, executes, breadcrumbs STATUS.md. Refine the brief template as you go.

FAQ

What you're probably about to ask.

Is this a Claude Project, a Claude Code workflow, or both?
Claude Code, specifically. The orchestrator + worker sandboxes pattern uses Claude Code's working-directory + hooks model. You can use individual sandboxes as Claude Projects in isolation, but the orchestration loop assumes Claude Code at the root.
What does it cost?
Zero beyond your existing Claude subscription. MIT-licensed. Fork it, modify it, redistribute it.
Who's it for?
Solo founders running content + design + research + light VA-coordination through one chat. The "agency of one" — one operator, folder architecture handling the multi-worker shape. Not for: teams with multiple humans-in-the-loop (the orchestrator pattern assumes a single decision-maker), or pure-code projects (the template is content/design business-shaped, not codebase-shaped).
Why one orchestrator + multiple sandboxes — why not one big agent?
Context isolation. Each worker has its own CLAUDE.md, its own MEMORY.md, its own routines. When the design worker hits 100K tokens, it doesn't poison the content worker. When the research worker discovers a new pattern, you choose whether to propagate it to others. Briefs are the only contract across the boundary.
Why briefs instead of just chatting?
Briefs survive chat-history compression. They give the worker a cold-start grounding it can re-read on a fresh session. They produce an audit trail — six months in, you can grep briefs/ and see exactly what you dispatched. And they make the work portable: a brief is a thing you could hand to a freelancer or another AI.
What's "ICM" — and why does it matter?
Inherited Context Model. The four-layer model the template uses (L0 = global instructions, L1 = current-state context, L3 = scoped references, L4 = per-task working artifacts) keeps context fresh without re-feeding everything on every call. ARCHITECTURE.md walks the full rationale.
How do I know if a brief actually shipped what I asked for?
The verify-handoff skill. Run it from the orchestrator pointing at a brief + the artifact the worker produced; it cross-checks the brief's deliverables against what's actually on disk and flags gaps.
Will it work for non-content businesses?
The architecture transfers. The example briefs and worker shapes are tuned to content/design/research. If you're running a consulting practice, an indie SaaS, or a service business, the orchestrator + worker + briefs pattern still applies — you'll just want to rename and re-scope the sandboxes.
How do I contribute?
GitHub issues + PRs. Especially welcome: new worker-sandbox patterns, new hooks, new skills, refinements to the brief templates.