dream-loop

By @anshuc. Maintained fork: kvnloo/dream-loop.

An agent skill that builds a game, app, or scene with impressive visuals, by creating a closed loop:

1. AI "dreams" up a high-quality target screenshot using image generation
2. AI builds with this target in mind
3. A separate AI critic compares the live screenshot to the target and provides feedback
4. AI loops back to step 2 until the critic is satisfied
5. Optionally, AI loops back to step 1 and dreams up an even better target based on the current state.

Installation

Upstream: npx skills add achimala/dream-loop

This maintained fork: npx skills add kvnloo/dream-loop, or clone into your agent's skills directory, or paste the link into your agent and tell it to figure it out.

Drop into Hermes

1. Clone this repo into a Hermes skills dir, e.g. ~/.hermes/skills/software-development/dream-loop or a profile skills/software-development/dream-loop.
2. Confirm SKILL.md is at the skill root (Hermes loads that file).
3. Product work stays in that product’s gitignored .dream-loop/ (copy templates/meta.md there). Never merge other products’ worktrees or stills.
4. Lock one frame (a close-up or facility hero), one camera, one capture. Each round: overlay + scripts/crop_gate.py (exit ≠ 0 skips lighting/materials judging; continue with a camera or massing action). Then a fresh judge child with the verbatim prompt in references/hard-gates.md.
5. One change class per round. Stall if best score is not +1 in 2 rounds — ask the user; do not nibble.
6. Progress is MEDIA (target + capture + overlay), not a status paragraph.
7. Do not git-track sqlite under .dream-loop/. Do not TDD the render.

This fork’s gates live in references/hard-gates.md.

Prerequisites

You need an AI agent with:

• access to image generation, either built-in (e.g. Codex, Grok) or via API (e.g. give it a Gemini API key)
• vision input
• subagents (optional but strongly preferred)

Install Blender if you want custom 3D modeling. The Blender MCP or scripting interface is preferred to computer use; it produces better results.

At the moment this is only tested with GPT-6 Astra in Codex. Other strong models like Claude Fable 5.1 can likely work too.

Example

Prompt:

Build me a graphics demo: isometric camera, voxel-ish art style with realistic shading and reflective wet floors, a character in an interesting scene. Fantasy setting (think Elden Ring, Diablo). Three.js in browser, >60fps. Don't download assets. Time limit of 1 hour. Controls: click to move the character, camera lazy-follows; drag to rotate camera; scroll to zoom in/out. No gameplay for now. World should feel alive: motion, animations, subtle environmental behaviors. Area around player should look expansive, but only allow movement in a limited space. No need to confirm the art with me or ask questions, just go!

GPT-6 Astra on high effort in Codex:

Try the live demo

Classic workflow

The pre-simplify gated judge + asset-first loop lives under references/classic/ (restored from upstream git history). Use it when you want stricter iteration than the default Plus/Pro split. Say Dream Loop Classic in your prompt.

Contributing

This fork welcomes PRs at kvnloo/dream-loop. Please still credit Anshu as the original author.

See CONTRIBUTING.md. If you open a PR, please provide example results produced by the skill when the change could affect visual quality, to ensure the changes don't regress performance.