Claude Code Meetup · 2026

From Code
to Running
Feature

How I gave Claude Code a fully running full-stack app so it could test itself — and started shipping entire features autonomously.

Gal Tidhar Sedric AI
01 / 11 The Problem

Claude can write.
Can it ship?

Without a running stack
  • Claude writes code
  • Start backend manually
  • Load env vars manually
  • Start frontend manually
  • Open browser, check it
  • Report back to Claude
  • Repeat every change
With a running stack
  • Claude writes code
  • Stack already running
  • Claude observes the result
  • Claude iterates itself
  • You review when done
Writing code and shipping a feature are two different things. The gap between them is you.
02 / 11 The Solution

One command.
Full stack. Every time.

$ npm run dev:develop

▶ Resolving environment from AWS SSM...
✓ .env.develop written — Firebase, Snowflake, OpenAI all loaded
▶ Starting FastAPI backend...
▶ Health check loop: waiting for /health → 200...
✓ Backend healthy at http://localhost:8000
▶ Starting Vite frontend dev server...
✓ Frontend ready at http://localhost:5173

🟢 Full stack running. Hot reload enabled. Claude, you're up.
03 / 11 How It Works

Orchestration +
secrets without ceremony

Startup sequence
setup-environment.sh
Pulls secrets from AWS SSM → .env.develop
↓
run-backend.sh
Sources .env, starts uvicorn
↓
health check loop
Waits for /health → 200. No race conditions.
↓
run-frontend.sh
Vite + hot reload. Changes appear instantly.
SSM template pattern

Template checked into git:

FIREBASE_PROJECT_ID=
  {{SSM:/sedric/dev/firebase/project}}
SNOWFLAKE_ACCOUNT=
  {{SSM:/sedric/dev/snowflake/account}}
OPENAI_API_KEY=
  {{SSM:/sedric/dev/openai/key}}

resolve_env.sh writes the real values:

FIREBASE_PROJECT_ID="sedric-develop-abc123"
SNOWFLAKE_ACCOUNT="xy12345.us-east-1"
OPENAI_API_KEY="sk-proj-..."
No .env files ever committed.
New engineer or Claude: run setup once, same result.
04 / 11 The Contract

CLAUDE.md —
the operating manual

CLAUDE.md
## 🚨 SERVER STARTUP
npm run dev:develop # the ONLY way
❌ uvicorn main:app --reload
## Code Style
Logger: aws_lambda_powertools only
Icons: EM-based sizing, never Tailwind w-4
Comments: none. Code speaks.
## Error Philosophy
Fail fast. No graceful degradation.
If it's needed — it must work.
## Design Principles
Question every ID system.
Simple > complex. Always.
## AWS + Deployment
sedric-prod-poweruser: always ask first.
CI/CD only. Never deploy manually.
05 / 11 Skills

Reusable workflows,
not one-off prompts

  • ⚡ using-superpowers — loads at session start. Enforces skill-first discipline every session.
  • 🔥 grill-me — relentlessly interviews you about a plan until there's genuine shared understanding. No hand-wavy specs.
  • 📝 make-plan — turns a spec or ticket into a detailed, phased PRD with doc discovery. The blueprint before the build.
  • 📌 monday-ticket — breaks a PRD into Monday.com tickets, one per task. Issues created from the plan, not from memory.
  • 📋 feature-dev — end-to-end delivery: explore → PRD → issues → implement → validate. Ordered, not improvised.
  • 🧠 brainstorming — before any architecture decision. Forces multiple options before any commitment.
  • 🐛 systematic-debugging — structured bug hunting. Hypotheses before fixes. No random changes.
  • 🚀 commit-push-pr — commit, push, open PR with Monday link. One command ships.
  • 👁 monitor-ci — watches CI after push. Notifies on pass or fail. No babysitting.
Skills are markdown files. Write once — invoke every time. Steal and adapt.
06 / 11 MCPs

Claude reaches out.
You don't relay.

  • 📌 Monday.com MCP — reads ticket spec directly. 27 content blocks fetched in the demo. No copy-paste, no misquote, no invented requirements.
  • 💬 Slack MCP — posts updates when blocked, notifies on completion, searches discussions. Claude communicates without you as the relay.
  • 🏥 Sedric Insights MCP — queries internal production data: orgs, interactions, STT configs, speaker ID metrics. Prod data in context.
  • ❄️ Snowflake MCP — run analytics queries directly in the conversation. No context switching to a SQL client.
  • 🖥 IDE MCP — surfaces inline type errors and diagnostics. Claude sees the same red squiggles you see, without you describing them.
  • 🔧 gh CLI (via Bash) — not an MCP, but Claude uses it directly. Issues, PRs, checks — all without leaving the session.
Every MCP is a context source Claude can pull from. The richer the context — the less you have to explain.
07 / 11 What Went Wrong First

The system exists
because things broke

Before CLAUDE.md — Claude tried to start uvicorn directly. It failed silently. Claude spent 20 minutes debugging "why the server isn't responding" when it was simply starting wrong. → Fix: document the ONLY way to start, with 🚨 headers.
Before the health check loop — the frontend started before the backend was ready. It threw confusing CORS errors. It wasn't CORS — the backend just wasn't up yet. → Fix: poll /health until 200 before starting Vite.
Before the SSM template system — someone committed a .env file with real credentials to git. Caught it before it merged. But once is enough. → Fix: templates in git, values in SSM, .env in .gitignore.
Every piece of this system exists because something broke first. That's not failure — that's how the constraints become obvious.
DEMO
npm run dev:develop → Claude gets a task → watches itself ship it
⚠ IN CASE OF CATASTROPHIC DEMO FAILURE ⚠
PRESS F
TO PAY RESPECTS
08 / 11 What Just Happened

The full chain.

01 — Ticket read via MCP
URL dropped in chat. 27 content blocks fetched. Zero copy-paste.
02 — Codebase survey
Explore subagent dispatched. 8 files read in parallel. Custom router + FE routing discovered.
03 — Decisions before code
4 architecture + 3 design questions. ASCII mockups rendered. Human picked one.
04 — Implementation
5 files written. Matched every codebase pattern exactly. PRD as the contract.
05 — Live validation
curl → Python contract assertion → 60-shot distribution test. Live server.
06 — Issues closed
2 GitHub issues: opened before implementation, closed after validation.
07 — PR opened via gh CLI
gh pr create — title, body, Monday link. Ready for review.
09 / 11 The Process

Why it produces working code.

Monday.com MCP — spec fetched direct from the ticket. 27 content blocks. No copy-paste.
→
No spec drift — requirements came from the ticket, not Claude's memory.
feature-dev skill — PRD drafted, GitHub issues opened before a line was written.
→
No scope creep — PRD + open issues = a contract. Implementation matched it.
Explore subagent — codebase surveyed in parallel. Custom router + FE routing discovered.
→
No pattern violations — found before touching anything.
AskUserQuestion — 4 architecture + 3 design decisions surfaced before implementation.
→
No architecture guesses — human decided, Claude executed.
gh CLI + curl + Python3 — live endpoint, contract assertion, 60-shot distribution test.
→
No "I think it works" — live server, real assertions.
Complex full-stack system. Custom patterns throughout. Working code. First try. Issues closed. PR opened.
10 / 11 One Thing to Take Home
The bottleneck
isn't what Claude
can write.
It's whether Claude
can observe
what it wrote.
Give it a running app and the gap closes fast.
Claude Code Meetup · 2026

Thanks.

Questions? Happy to dig into the setup.
Everything shown here is a pattern you can steal and adapt.

plot twist
$ npm install gal-tidhar --save
Previously bundled with Sedric AI. Now a standalone package.
Zero breaking changes · ships on time · comes with a CLAUDE.md.
Gal Tidhar Sedric AI ~free agent~
LinkedIn QR code
linkedin.com/in/galtidhar