May 26, 2026built

Built a Free Claude Code Starter Kit — 5 Layers, 3 Personas, Drop-In Ready

Built a Free Claude Code Starter Kit — 5 Layers, 3 Personas, Drop-In Ready — explore insights on built, free and more.

builtfreeclaudecodestarterkit

Most teams use Claude Code like a smart autocomplete. They get maybe 20% of what it can do.

The other 80% lives in a 5-layer architecture that Anthropic published but almost no one has actually set up end-to-end. I spent the last week building a complete, opinionated starter kit for it — free, open-source, and ready to drop into any project in 30 seconds.

Here's what I learned.


The 5 Layers (Most People Stop at Layer 1)

Claude Code isn't just a coding assistant. It's an extensible engineering platform with five distinct layers:

Layer 1 — CLAUDE.md (Memory) A markdown file Claude reads at the start of every session. Your rules, your conventions, your team context. Most people either don't have this or they write three lines. A well-crafted CLAUDE.md is worth more than any prompt you'll write.

Layer 2 — Skills (Knowledge) On-demand context chunks loaded when relevant. Not always-on — they're modular. A security-audit skill, a database-migrations skill, a mvp-scoping skill. Each is a detailed .md file that Claude loads only when it needs it. This keeps the main context clean and focused.

Layer 3 — Hooks (Guardrails) Shell scripts that fire on events: PreToolUse, PostToolUse, SessionStart, Stop. Deterministic. Not AI-driven. Think git hooks, but for your AI agent. Block secrets before they get written to disk. Log every tool call for SOC 2. Remind yourself to commit when you have 5+ open files.

Layer 4 — Subagents (Delegation) Isolated agents with their own context window, scoped tools, and scoped permissions. A code-reviewer agent that can only read. A release-manager agent that can tag and push. A security-officer agent that scans diffs and blocks merges with critical findings. The main agent delegates — subagents can't spawn further subagents.

Layer 5 — Plugins (Distribution) A bundle of skills + hooks + agents with a manifest. Think npm packages for Claude capabilities. Install once across your team.


The Problem: Generic Doesn't Work

The catch is that these layers need to be opinionated to be useful. A generic CLAUDE.md helps no one. A code-review skill that doesn't know your stack is just noise.

Enterprise teams need compliance guardrails, SOC 2 audit logs, and two-reviewer gates on auth changes. Startups need a cost checker that blocks them from accidentally spinning up $200/month of infrastructure. SMBs need a hook that blocks DELETE FROM table without a WHERE clause and Slacks the team when a deploy fires.

So I built three personas — Enterprise, SMB, and Startup — each with all five layers fully populated. Real content, not placeholders.


What's Inside

Enterprise Persona

  • CLAUDE.md: SOC 2 rules, GDPR compliance mandates, JIRA ticket requirements, change calendar enforcement

  • Skills: compliance-checker, architecture-review, incident-response, change-management, security-audit

  • Hooks: Pre-tool audit log (immutable, for SOC 2), secrets scanner, post-tool SOC 2 log, session banner with branch + open PRs

  • Agents: security-officer (read-only, blocks on CVSS 7+), release-manager (version bump → changelog → tag → GitHub Release), documentation-agent (syncs API docs + env var registry)

  • Plugin: governance (ticket reference enforcement), SLA tracker, compliance gate, compliance bot (attestation documents)

SMB Persona

  • CLAUDE.md: Customer-first rules, ORM mandates, Stripe idempotency requirements

  • Skills: shipping-fast (80/20 filter), customer-support-triage, feature-scoping (three-tier breakdown), database-migrations (expand-contract pattern), writing-docs

  • Hooks: Destructive command blocker (catches DELETE FROM without WHERE), session work log, Slack deploy notifier

  • Agents: product-manager (prioritizes GitHub issues by customer impact), support-agent (read-only DB access, diagnoses customer issues), deploy-assistant (full pre/post deploy checklist)

  • Plugin: CRM integration (churn risk queries), analytics (MRR, cohort retention), onboarding-helper (generates help articles + tooltip copy)

Startup Persona

  • CLAUDE.md: Hypothesis-first mandate, cost discipline, "ship it" constitution

  • Skills: move-fast (30-minute rule + tech debt docs), mvp-scoping (YES/NO feature filter, 48-hour build target), cost-optimization (LLM cost calculator, free tier guide), api-design (tRPC internal, REST external), growth-hacking (PostHog instrumentation, A/B scaffolding)

  • Hooks: Cloud cost checker (flags new infra by estimated monthly cost), commit reminder (fires at 5+ uncommitted files), async standup (last commit + open issues at session start)

  • Agents: founding-engineer (full-stack, high-agency, documents decisions), growth-analyst (activation funnel from DB + PostHog), investor-update-writer (pulls real metrics, drafts Paul Graham-style update)

  • Plugin: pitch-deck-assist (slide-by-slide with real DB data), metrics-dashboard (generates dated report), velocity tracker, launch-coordinator (generates pre-launch checklist + GitHub issues)


The Install

bashgit clone https://github.com/sseshachala/claude-code-workspace-starter
cd claude-code-workspace-starter
bash install.sh

Pick your persona. The installer copies CLAUDE.md to your project root, drops skills/hooks/agents/plugins into .claude/, makes all hook scripts executable, and merges settings.json. Then edit CLAUDE.md to replace the placeholders with your actual project context.


What I'd Build Next

A few things that aren't in the kit yet:

  • Agency persona — client work, multi-project context switching, billing awareness

  • SaaS product persona — somewhere between SMB and enterprise, heavier on product analytics and feature flagging

  • Interactive CLAUDE.md generator — ask a few questions, generate a personalized memory layer instead of filling in placeholders

  • Plugin registry — a way to discover and install community-built skill/agent bundles


Why Free?

Because the Claude Code community is early. The patterns are still being figured out. The best way to accelerate that is to get opinionated, working examples in front of as many teams as possible and let people improve them.

GitHub: https://github.com/sseshachala/claude-code-workspace-starter

PRs welcome. Especially for new personas.

Turn your brand into content like this

Narratr reads your website and generates SEO-optimised blog posts that sound like you.

Try Narratr free →