Oct 22 edition/Reporting & analysis
CodingAgentsBusiness

CodingSoftware & developer tools

French fork 'Panache' reskins Simon Willison's AI-built Battle Drones RTS a day after its Apache 2.0 release

A day after Simon Willison released Battle Drones, a browser RTS written almost entirely by Claude Code agents, under Apache 2.0, a French fork called Panache moved it to steampunk Belle Époque France with new 3D models, music and an animated intro.

From the Simon Willison article of October 21, 2026 and cited sources: Willison published Battle Drones on 20 October 2026 and says he added an Apache 2.0 licence that same day. [1] [2] [5] [7]

SOURCES[1] Simon Willison: Panache — RTS Belle Époque pour mobile simonwillison.net[2] When I slapped an Apache 2.0 License on the code yesterday… x.com[5] simonw/battledrones (GitHub) github.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Willison published Battle Drones on 20 October 2026 and says he added an Apache 2.0 licence that same day. [1] [2] [5] [7] [8]
Original source
Simon Willison article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
October 21, 2026 article published
Evidence
12 sources across 9 sites

Reach: not measured yet

THE CORE IDEAS3 TAKEAWAYS
01

Willison published Battle Drones on 20 October 2026 and says he added an Apache 2.0 licence that same day. By 21 October, GitHub user menufactory43 had a playable French adaptation called Panache online, which shows how quickly permissively licensed, agent-generated code can be forked and reused. [1] [2] [5] [7] [8]

02

Panache mainly changes the look and feel. It renames units (for example, dirigibles and the walking fortress Gargantua) and adds new models, effects, music, a title screen and a redesigned mobile interface. The game's art is made by code or an image API, and its audio is synthesised in code, so reskinning it is mostly a matter of editing code. [1] [7]

03

Willison says the original was vibe-coded with almost no human code review, and its balance changes were checked with the agent's own simulated playtests. As ExplainX points out, simulated playtests can miss hidden bugs such as memory leaks, so anyone who forks the code also inherits its unknown defects. [4] [6]

WHY IT MATTERS

What the sources show: permissively licensed, agent-written code was forked and themed within a day.

Read the full assessment

What follows from that: when art and audio are generated in code, localising or remixing a game becomes cheap, but each fork inherits defects nobody reviewed.

Executive brief

Simon Willison released Battle Drones, his "vibe coded" browser RTS, on 20 October 2026. He says he added an Apache 2.0 licence the same day. By the next day a French fork called Panache was live. It moves the game to steampunk Belle Époque France with new 3D models, French-flavoured music and an animated intro (Willison). Willison wrote almost none of the original code himself, and so far the fork mostly changes the look rather than the gameplay. For practitioners, the case shows that code an AI agent generated, once released under a permissive licence, can be forked and reskinned within a day. No independent analysis of the fork exists.

What changed and event timeline

  1. Battle Drones released

    Willison published a Command & Conquer-style browser RTS built with Claude Code across 321 agent-written commits. Its two-device multiplayer sends game data as sound through each device's speaker and microphone ().

  2. Licence applied

    Willison later said he put an Apache 2.0 licence on the code that day. The repository now shows Apache-2.0 (;).

  3. Secondary coverage

    ExplainX summarised the project. It pointed to undisclosed token costs and to multiplayer that had only been tested lightly ().

  4. Panache fork surfaces

    GitHub user menufactory43 published Panache, which the repo calls "adaptation française de Battle Drones" (a French adaptation of Battle Drones), with a playable build (;).

  5. Willison highlights the fork

    He described the fork as mostly a visual reskin and published a side-by-side model page. On Hacker News, one commenter called projects like this low-effort "slop" (;).

Capabilities and access

  • Battle Drones: free in mobile and desktop browsers. Modes are Skirmish, Last Stand and an 8-mission campaign.
  • Built with: Claude Code in "Ultracode" mode, using Claude Fable 5, Claude Opus 5.5 and Claude Haiku 5.5 (post).
  • Panache: Apache 2.0.
Read the full section
  • Battle Drones: free in mobile and desktop browsers. Modes are Skirmish, Last Stand and an 8-mission campaign. Multiplayer works either over audio between nearby devices or through a WebSocket relay (post).
  • Built with: Claude Code in "Ultracode" mode, using Claude Fable 5, Claude Opus 5.5 and Claude Haiku 5.5 (post).
  • Panache: Apache 2.0. It renames units (dirigibles, steam "Fardiers", the walking fortress Gargantua) and adds new models, effects, sound, music, a title screen and a redesigned mobile interface (repo).
  • Model comparison: original units and Panache units.

Technical analysis for researchers and developers

  • Stack: TypeScript, Vite, Three.js and Vitest, plus a Python art pipeline.
  • Determinism: the simulation replaces the browser's built-in math functions, because Chrome and Safari returned slightly different values and players' games would drift out of sync.
  • Assets: art is made by code or by an image API, and Panache's README credits OpenAI's gpt-image-2.5-sunburst.
Read the full section
  • Stack: TypeScript, Vite, Three.js and Vitest, plus a Python art pipeline. The game began in Phaser 2D and moved to Three.js in one overnight agent run (repo; post).
  • Determinism: the simulation replaces the browser's built-in math functions, because Chrome and Safari returned slightly different values and players' games would drift out of sync. The same deterministic design makes replays and AI-vs-AI matches possible (ExplainX).
  • Assets: art is made by code or by an image API, and Panache's README credits OpenAI's gpt-image-2.5-sunburst. Sound and music are synthesised live in code. Because there are few binary asset files, a reskin is mostly a matter of editing code (repo).
  • Fork size: at retrieval, Panache had 374 commits and upstream had 334 (repo; upstream). That gap suggests a modest amount of new work, but this is an inference. The fork's coding tools are not documented.

Claims and evidence

  • The game was vibe-coded with almost no human code review. This is Willison's own account (post).
  • The "Make it a LOT more fun" update helped losing sides. That figure comes from the agent's own simulated playtests, not human players (post).
  • The relay costs about $2.19/month. This is his estimate for the Fly.io server (post).
Read the full section
  • The game was vibe-coded with almost no human code review. This is Willison's own account (post).
  • The "Make it a LOT more fun" update helped losing sides. Willison says comebacks from 800+ scrap behind rose from 24% to 46% after the update. That figure comes from the agent's own simulated playtests, not human players (post).
  • The relay costs about $2.19/month. This is his estimate for the Fly.io server (post).
  • Panache is mostly aesthetic. This rests on Willison's assessment and the fork's README. No independent review exists (Willison; repo).

Context and prior work

  • The audio multiplayer grew out of Willison's earlier "Chirp Pong" experiment.
  • Willison defines vibe coding as building with an LLM without reviewing the code it writes (Cinevva guide).
  • The story also spread through aggregators (daily.dev; Vibe Leaderboard).
Read the full section
  • The audio multiplayer grew out of Willison's earlier "Chirp Pong" experiment. He first used a "Radio Probe" tool to measure how much data speakers and microphones could carry (post).
  • Willison defines vibe coding as building with an LLM without reviewing the code it writes (Cinevva guide).
  • The story also spread through aggregators (daily.dev; Vibe Leaderboard).

Limitations, safety and contested findings

  • Willison calls Battle Drones "way too low effort" to count as a legitimate indie project (post).
  • The high-pitched audio link caused headaches, which led to the WebSocket alternative (post).
  • ExplainX notes that simulated playtests catch gameplay problems but not hidden bugs such as memory leaks (ExplainX).
Read the full section
  • Willison calls Battle Drones "way too low effort" to count as a legitimate indie project (post).
  • The high-pitched audio link caused headaches, which led to the WebSocket alternative (post).
  • ExplainX notes that simulated playtests catch gameplay problems but not hidden bugs such as memory leaks (ExplainX).
  • On Hacker News, one commenter dismissed such projects as "slop" (HN).
  • Willison's launch post as retrieved does not mention a licence. The licence was confirmed separately on X and in the repo (X).

Business and practitioner implications

  • A permissive licence on agent-generated code made a themed derivative possible within a day.
  • Generated assets are cheap to swap. When art and audio come from code or an image API, localisation and reskinning get cheap.
  • Review obligations remain. Teams that fork unreviewed AI-written code inherit its unknown defects, so they still need tests and review.
Read the full section
  • Licensing matters. A permissive licence on agent-generated code made a themed derivative possible within a day.
  • Generated assets are cheap to swap. When art and audio come from code or an image API, localisation and reskinning get cheap. That favours remixing over protecting a distinctive look.
  • Review obligations remain. Teams that fork unreviewed AI-written code inherit its unknown defects, so they still need tests and review.
  • Determinism pays off. Building the simulation to be deterministic gave the project replays and AI-vs-AI evaluation at little extra cost.
FOLLOW THE EVIDENCE

The source trail.

Sources (12)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.