← Dev Blog

Day 0: Before the jam

The Left for Mead title art: a carved wooden tavern sign with the game's name above two clinking tankards.
AI-generated UI art (ChatGPT image generation, made during the jam window). Shown here as the series banner.

TL;DR: Left for Mead started as a conversation I had with ChatGPT, and that conversation became my master brief. I handed the brief to Claude, we locked in the big design calls together, and we spent two days getting every tool connected. Nothing for the game was built or generated before the jam opened at 6 PM on Thursday 12 November (NZDT).

First, a bit about me

Before any of the game stuff, some context. I’ve never built a game before. I’ve played plenty over the years, though not for a long while now, because children and life and adulting got in the way. I work full time, and at home there’s a three-year-old and a three-month-old. So “four days of jam” for me really meant snatched hours: early mornings, nap times, late nights and the odd few minutes between everything else.

I’m sure plenty of people in this jam are in the same boat, so this isn’t a sob story. It’s just the honest backdrop to everything below. I could only get to the computer so often, and Claude and Codex did a huge amount of the heavy lifting in between. I’d make the calls, play the builds and say what felt wrong, and they’d turn that into code, levels and fixes while I was at work or with the kids.

The tools had their own limits too. I started on Claude’s Pro plan and kept hitting the five-hour usage wall, sometimes twice in a day, and everything would stop until it reset. About halfway through the jam I moved up to the Max 5x plan, and that’s when it really got into a flow.

Where it started: ChatGPT and the master brief

Before the jam, I talked the idea through with ChatGPT: a co-op medieval survival-horror game in the spirit of the classic four-player zombie shooters, but with drunk tavern heroes, a druid villain and a lot of dark humour. It’s a game I’ve wanted to make for a long time. That conversation became the Left for Mead Master Project Briefing v1.2. It set out the survivors, the villain (the Briar Warden), the mead risk-and-reward idea, the Ruckus meter, a pressure director, bots, listen-server co-op and a hard low-spec target.

On Wednesday 11 November I handed that brief to Claude, which would coordinate the project. My first instruction set the rule for everything after it: Left for Mead was my preferred game, not a commitment. When the theme was revealed, we’d weigh three options (build it, adapt it, or pivot to a smaller game) and not force a weak fit. Nothing would be built until I chose a direction inside the jam window.

Wednesday 11 November: decisions and setup

Most of Wednesday was design talk. Claude would suggest, I’d decide, and by the evening we had the shape of the game:

  • Skills: three active skills and one ultimate per hero. Skills come from each hero’s body or the kit they always carry, so they work no matter what weapon is in their hands. The Hunter got traps, throwing knives, Mark Prey and a Volley ultimate. The Guard got a buckler bash, Shield Wall, Rally and Last Stand. The Brawler got a shoulder charge, Ground Slam, War Cry and Berserker Surge. The Tavernkeeper got Bottle Toss, Cask Spill, Field Remedy and the “Round on Me!” ultimate.
  • Pings: a “point it out” button that highlights weapons and armour for the team and the bots, with a voice line from each hero.
  • Length: the Cursed Village as a mini-campaign of three segments, each aiming for about 15 minutes, with checkpoints. Each segment has to be finished and polished before the next one starts.
  • Bots: follow, fight and revive, with simple skill use where it’s cheap. If the host drops, everyone goes back to the lobby rather than migrating the host.
  • Ending: a hilltop camp at dusk, the distant tavern lights below, and a wizard silhouette as a tease for a post-jam hero.
  • Budget: effectively zero. Free assets first, paid packs only as a fallback.

While we talked, Claude ran the setup work in the background. A small, cheap Claude model (Haiku 5.5) installed Unreal Engine 5.9 Preview on my PC next to my existing 5.8 install and tested its built-in MCP server on a throwaway project. Another Haiku thread built a jam asset catalogue from Fab and Megascans: world first, then characters, then animations. Claude installed the Meshy MCP server and set up the Hyper3D Rodin connection, while I did the sign-ins and kept the API keys out of the chat. It read the jam perk emails and listed what to claim and what to cancel later, and it drafted the game repo’s README and AGENTS.md, with Cursor added as a second coding agent next to Codex.

The tools that day were ChatGPT (the original brief, and later a licence check on a bundle of Leartes assets, which I decided not to rely on because of its AI-use clause), Claude Code Desktop on my PC, Unreal Engine 5.9 Preview, the Meshy and Hyper3D MCP servers, ElevenLabs (connected on the Starter plan, which allows commercial use) and GitHub.

Thursday 12 November, before 6 PM: the theme and the plan

On Thursday morning I shared a second brief, this time for this website, the dev blog and the Codex pipeline. We agreed the site would live in its own repository and that Codex would make every change to it. At about 12:25 PM the organisers’ email confirmed the theme: Director’s Cut, “your vision, your edits, your final cut”. Claude wrote up the three options, and I chose to build Left for Mead. The fit felt natural. The pressure director, the thing that decides when the horde comes, becomes something players can actually see.

That afternoon I reworked the gear system. Heroes keep their fixed skills, but weapons and armour set a hidden class: the main weapon picks the class, and the four armour slots set how well you fit it. A Hunter in plate with a sword is allowed, just less effective. Skills ignore gear so nothing stacks out of control, and every hero gets a main weapon plus a weaker sidearm so nobody is useless at range or up close. In the lobby, each player picks a different hero and bots fill the empty slots. Challenge unlocks for future heroes (the wizard first) went on the post-jam list.

Claude then checked the plan against the official rules: everything in the game must be made after 6 PM, all text and audio in English, no Unreal Engine logo on screen, teen-suitable content, and a 30 to 60 second gameplay video at 1080p or higher. It wrote a kickoff plan so that one word from me would start every workstream at once, and prepared the first Codex prompts. The README and AGENTS.md went into the game repo as pre-jam setup.

One thing I was careful about: nothing for the game was made before 6 PM. No assets, models, images, voices or code. The jam rules say everything must be made inside the window, and I wanted the timeline to be clean.