Case Study · In Progress
One codebase, many branded events sites for dispensaries, brands and manufacturers. Built with Claude Code and a decision graph, from an empty repo to production in about nine days.

Cannabis businesses run a lot of events: vendor days, product drops, tastings, grower Q&As, pop-ups and staff-only nights. Most of them get promoted on a single flyer and a few social posts that are gone by the weekend. The mainstream event tools aren't built for this industry, and every dispensary, brand and manufacturer that wants a proper events page ends up paying for its own one-off website.
getstonedwith.me flips that around. It's one Laravel codebase serving many branded events sites. Each client gets their own subsite, either at yourname.getstonedwith.me or at getstonedwith.me/yourname, with their own logo, colors, calendar, event pages and RSVPs. Adding a new client is a config file plus an assets folder, not a new project. At the root sits a hub with a combined calendar, a client directory and a lead form.
The repo was empty on November 3, 2026. Production and staging were both live on November 4. By November 11 the platform had seven feature milestones behind it and more than 230 commits. A few things made that pace possible without cutting corners.
The thing that makes AI-assisted development hold together over days instead of minutes is memory. Every session of an AI coding agent starts cold. We use deciduous, a decision graph that lives in the repo, to give it that memory.
Every piece of work follows the same chain: a goal leads to the options considered, the options lead to a decision, the decision leads to actions, and actions lead to outcomes. Observations (a gotcha, a surprise, a production quirk) attach wherever they're relevant. Commits are linked to the nodes that produced them, and the owner's request is stored verbatim on the goal node, so the graph records why, not just what.
At the start of every session the agent runs a recovery step that reads the recent graph, the open goals and the git history before touching code. A session that picks up a half-finished task knows what was tried, what was rejected and why. When the agent spotted that njlegalize.me showed the wrong organizer on synced events, it logged that observation, and it became the next goal.
Over the first nine days the graph grew to more than 650 nodes and 700 links: every option weighed, every decision made and every outcome verified.
None of this runs unattended. Agents can only push to their own claude/* branches; a git hook blocks anything else, and deny rules keep them away from production secrets and destructive commands. Brian reviews and merges every pull request, decides what gets promoted to production, and owns every product call, from pricing to which networks to launch with. The AI does the typing and keeps the notes. The judgment stays human.
Each client site is a config file plus an assets folder, served at a subdomain or a path. Colors flow from config into CSS variables, so one set of templates renders every brand.
Every record belongs to a site and is filtered automatically. Client staff log in as owners, editors or viewers, and Legalize Media has a super-admin console across all sites.
AI-Assisted Development
Built with Claude Code working in parallel sessions, each on its own branch, with every pull request tested in CI and merged by a person. Here's what that pace looked like.
Day 1
A full design doc was written and logged as decisions before any code. Then the Laravel 12 scaffold, multi-site core, CI on MySQL 8, the Manage console with roles, the hub and the deploy config all landed the same day.
Day 2
The Decision Graph
Every piece of work is logged in a decision graph that lives in the repo: the goal, the options weighed, the decision, the actions taken and what came of them. Each session starts by reading it, so nothing is re-decided and nothing gets lost. Here's a real chain from the build.
1. Goal
Turn on live event publishing to njlegalize.me, staging first
2. Options
Staging only; production waits for the first real client
✓ Chosen
Switch staging and production together
Create a demo dispensary on njlegalize.me staging
The same Laravel stack as njlegalize.me, so the events code ported cleanly, plus an AI-assisted workflow with guardrails at every step.
Laravel 12
One app serving every subsite, with site resolution from the host or path
MySQL 8
Production-parity database in development, CI and production
Blade + Alpine.js
Server-rendered templates themed per site through CSS variables
Tailwind CSS
One design system, recolored per client from config
FullCalendar
Month and list calendar views on every subsite and the hub
Leaflet
Venue maps on event and venue pages
Redis Queues
Retrying background jobs for njlegalize.me publishing and social posts
Pest 4
270+ feature and browser tests, including five edge-case QA sites
GitHub Actions
CI on MySQL 8; develop deploys staging, main deploys production
DigitalOcean
Isolated PHP-FPM pool, worker and Redis database on a shared droplet




We're building getstonedwith.me end-to-end — product, design, code, CI/CD, and deployment — with AI doing the typing and people making the calls. If you want a platform built this way for your cannabis business, let's talk.
FullCalendar month and list views, Leaflet venue maps, category filters, and RSVPs with guest counts and a waitlist.
VIP nights, vendor days and staff events that only invited guests can see, with personal invite links.
A per-site iCal feed for Apple and Outlook, one-click Google Calendar links on every event, and a subscribe button on the calendar page.
A script tag or an iframe drops a client's upcoming events into the website they already have, and it stays in sync automatically.
Events publish to the client's njlegalize.me listing over a signed API, and edits and cancellations sync. The subsite stays canonical, and njlegalize.me links back to it.
Events post to the client's own Bluesky, Mastodon, Telegram and Discord accounts, with branded card images, free-form posts, and a share kit for networks that aren't connected.
A 21+ age gate on every site, social copy written to stay within cannabis advertising rules, plus JSON-LD event markup and sitemaps for search.
A develop → staging, main → production pipeline matching njlegalize.me. Security hardening, and live event publishing to njlegalize.me verified end to end on staging.
Days 3–4
A "Post to" network picker on the event form, five edge-case QA sites, a 21+ age gate on every site, Google Calendar links and role-based dashboards.
Days 5–7
Free-form social posts, live email, public pricing and concierge pages, and leads flowing straight into our CRM.
Days 8–9
A new palette and type, sharper homepage and pricing copy, and Founding Partner pricing, reviewed by a person and shipped through the same pipeline.
✓ Chosen
Point demo1 at an existing staging business
3. Decision
Switch staging only, and create a demo dispensary to publish to
4. Actions
Code: resync command, paused-site warning, signature test, docs
Ops runbook for the owner: shared secret, API token, environment
5. Outcomes
Code merged: 207 tests pass, linter clean
Verified on staging: 10 events sent, 10 created, 0 failed
njlegalize.me shows itself as organizer, not the dispensary. Became goal #170, fixed the same day
651 nodes and 711 links, with commits tied to the nodes that produced them and the owner's requests stored verbatim on each goal.
Claude Code
AI pair programmer working in parallel sessions on claude/* branches
deciduous
Decision graph that gives every AI session the full history of what was decided and why

