Brandon Welner

Problem

Most RPGs hand you a class and put you on its rails. MythReal does the opposite — a character is defined entirely by which of 20 tech trees they invest in. Removing classes removes the scaffolding that keeps a system legible, so the real work is replacing that scaffolding with cleaner math and a tighter economy. I've been designing it for about four years.

My role

Sole designer, developer, and art director. The ruleset, worldbuilding, visual identity, and digital tooling are all mine — and they have to agree with each other.

Craft

Twenty curated tech trees instead of classes, modifier-only attributes, and a fixed action-point economy — 5 AP a turn, a 2 AP bank cap, and a spike mechanic where banking AP unlocks decisive abilities. Every tuning choice traces to one unit: 1 AP ≈ a second of focused effort ≈ a d6 of effect. A two-register visual language carries the world: hand-drawn parchment for maps, diagrams, and seals; cinematic painting for scenes and characters. Maps are generated as a base, then finished by hand in Affinity. A repeatable prompt pipeline drives both the art and the tooling — including a Claude Code–built auto-battler that generates characters from the real rules and runs automated combat in a battle log.

The digital tooling is where the ruleset comes alive. The auto-battler, Raw Essence, is a seeded gacha: every run is deterministic and shareable as a copyable code, with pack-opening, a pity system, a persistent collection, and a draft-your-party-of-three loop feeding an auto-battle simulator that runs from half-speed to instant. Both it and the Build Your Legend character forge generate characters from the genuine v0.5 rules — the rules are data, with a single source of truth driving game, tooling, and balance alike.

Outcome

A locked v0.5 spec, iterated through public alpha across several versioned releases; a coherent identity that scales to future modules; and live, playable tooling — a seeded auto-battler and a character forge — running on that genuine ruleset.

the design output

Primary wordmark

Key art — Burny's chamber

Character study — Blob Marley

Field study — dragon anatomy

Location map — Teaminton

Location map — Raefund

Problem

An end-to-end tool for the software development lifecycle — carrying a piece of work from its initial requirements writeup all the way through to individual pull-request creation and review, on top of the team's existing systems.

My role

Two-person build. I owned the requirements-writeup (RPE), spike, epic, and ticket-creation skills — the front half of the pipeline that turns an intent into structured, ready-to-build work.

Craft

An agentic system built on Claude Code and MCP. Discrete skills handle each stage — RPE, spike, epic, and ticket creation, then dev work, PR review, and addressing PR comments — backed by MCP servers for Confluence, Jira, Bitbucket, and context management. The pattern is the point: small, composable agents that each own one SDLC step and hand off cleanly.

Outcome

Shipped as a unified SDLC now used across the engineering domain at Choice Hotels.

Problem

I hypothesized a cheap epistemic signal that could flag a model's ungrounded claims and gate verification inside an agent loop.

My role

I designed and built the eval to hold my own idea to an empirical standard.

Craft

A harness with independent measurement pathways and a preregistered effect-size bar — built so the hypothesis could fail cleanly. When the first run looked like a win, I traced it to a directionless baseline and replaced it with the model's own signed verdict.

Outcome

Against the honest baseline the signal added nothing, and on the decisive cases — confident, consistent, and wrong — it flagged none of seven. The mechanism explained why: it measured whether an entity exists, not whether a claim is true. A clean negative, reported as one.