DEV Community: milo zhang

Show DEV: Project Moon Life — a free browser-based urban life simulator

milo zhang — Sun, 18 Oct 2026 09:09:17 +0000

Project Moon Life is a free, unlimited, unofficial urban life simulator that runs entirely in your browser — no sign-up, no install, and the core game costs nothing.

What it does

You start a life in a fictional city — the Nest, the Backstreets, or the Outskirts — and steer it through choices that shape your attributes, career, relationships, and ending. Each run is seeded, so the free simulator is unlimited, and your story saves locally when you're done.

A one-time $10 USD AI Chapter Pass is optional; it unlocks AI-guided lives in the AI Chapter and Participating Life modes. The local simulator stays free either way.

Quick facts

Play in the browser: projectmoonlife.com
Free unlimited seeded runs, saves stay on your device
No sign-in required to play
An unofficial fan work inspired by the worlds of Lobotomy Corporation and Limbus Company

Feedback is welcome — what would you want from a text-driven life sim?

A Practical Checklist for Choosing a DeepSeek Harness Plugin

milo zhang — Sun, 11 Oct 2026 10:02:46 +0000

When adding a plugin to a coding workflow, the install command is only the beginning. A few quick checks can make the choice easier to understand and safer to maintain.

Start with the source

Find the public repository or package that the plugin uses. Check who maintains it, whether the project has recent activity, and whether the source matches the feature you expect. A directory entry is a useful starting point, but the project documentation remains the authority for installation and behavior.

Read the README before installing

Look for prerequisites, supported versions, permissions, configuration, and removal steps. If those details are missing, pause and investigate before adding the plugin to a profile you rely on. Treat commands as executable code: understand the package or repository they reference and the changes they make.

Match the install command to your setup

DeepSeek Harness plugins can use different installation formats. Confirm that a command targets the profile you intend to use, and check whether a path refers to a local checkout or a published package. Keep a note of the command and source so you can reproduce or undo the setup later.

Recheck after installation

Try one small task that exercises the plugin, then inspect the output and any logs it produces. Remove or disable the plugin if its behavior differs from its documentation.

For discovery, DSH Plugins is an independent directory that presents reviewed DeepSeek Harness extensions with public source and README notes, alongside installation commands. Use those details to find candidates, then verify each candidate against its maintainer documentation before installing.

A little source and documentation review up front makes plugin selection easier to explain, repeat, and troubleshoot.

AI Kenerate: A Practical AI Creative Studio for Video, Image, Music, and Voice Workflows

milo zhang — Sat, 10 Oct 2026 16:36:39 +0000

AI Kenerate is an AI creative studio for turning ideas into visual and audio content from one workspace.

It supports creating and transforming videos and images from text or reference media, as well as generating music and voiceovers. The workspace also includes transparent PNG generation, image prompts, a prompt library, and a community showcase.

A few workflows it covers:

Text-to-video and image-to-video experiments
Text-to-image creation and image transformation
Voiceover and music generation
AI-assisted creative workflows with the AI Kenerate Agent

The product offers monthly and yearly subscriptions plus one-time credit packs. Explore the studio at AI Kenerate.

A Practical Directory for DeepSeek Harness Plugins

milo zhang — Fri, 09 Oct 2026 14:50:34 +0000

DeepSeek Harness plugins can differ in installation paths, source visibility, and setup notes. DSH Plugins is an independent, human-reviewed directory for discovering these extensions and comparing the information that is publicly available.

The directory currently lists 36 reviewed plugins. Visitors can search summaries, compare package and local-checkout installation paths, inspect public source links and README notes, and copy profile-specific installation commands when they are documented. Maintainers can also submit canonical public GitHub repositories for review before publication.

DSH Plugins is a directory, not an installer. Its review is not a comprehensive source-code or security audit, so users should inspect the linked repository and documentation before installing anything.

Explore the directory: https://dshplugins.com/

A Practical Way to Compare Decision Models Before Shipping

milo zhang — Fri, 09 Oct 2026 14:02:25 +0000

When a product needs to classify, score, or route text, the hard part is often choosing a decision model that behaves well for the job. A useful workflow is to make the question structure explicit, compare several models on the same inputs, and only then move the best candidate into an application.

Start with a structured question

A decision task is easier to evaluate when its expected output is clear. For example, a Choice question can select one label, a Score question can return a rating, and a yes-or-no question can answer a binary check. Keeping the input and output shape consistent makes model comparisons more meaningful.

Compare before you route

Decisions API is an independent workbench and API for exploring hosted decision models. Its playground supports structured Choice, Score, and Noul questions over text or JSON, so a team can compare models before sending requests to a shared API endpoint. Supported models can also accept image inputs. See the Decisions API playground to explore the workflow.

The same approach helps with routing: use a small evaluation set, record where models disagree, and keep the examples that expose edge cases. The goal is not to pick a winner once, but to make the trade-offs visible before they become production behavior.

Decisions API currently lists 13 hosted models. New accounts receive 100 credits shared across the playground, API, and evaluations; one-time credit packs start at $10 and credits do not expire. Model rates vary, so measuring a representative set of requests is still important.

A compact checklist

Define the output shape before comparing models.
Test the same text or JSON examples across candidates.
Inspect disagreements and borderline scores.
Route only after the evaluation set reflects real traffic.

This keeps model selection tied to observable behavior rather than a vague preference for a particular model.

A Practical Workflow for Turning Flat Images into Editable Layers

milo zhang — Fri, 09 Oct 2026 11:47:30 +0000

Working with a finished PNG or JPEG can be frustrating when the original design file is unavailable. You may want to move a subject, remove a decoration, or reuse text, but the image is still one flattened surface.

A useful workflow is to treat the image as a set of visual components and reconstruct those components as transparent layers. Magic Layers is an AI image tool built around that idea: it separates a finished PNG or JPEG into a base image and individual transparent PNG layers, with a name, description, and position for each result.

A simple layer-separation workflow

Start with the best source image available. A larger, cleaner image gives the separation model more visual information. Avoid adding compression artifacts if you can export the source again.
Decide what you actually need to separate. A target of two to fifteen images is supported, and an optional description can call out the elements that matter most—for example, the subject, headline, background, or decorative shapes.
Review the generated layers as raster assets. Each output is a transparent PNG, so it can be moved and composited in Photoshop, Figma, Canva, a presentation, or a game engine. The original positions are retained so the layers can be aligned with the base image.
Check edges and small details. Hair, shadows, text, and partially occluded objects can be difficult to recover perfectly. Treat the result as an editable starting point and clean up masks or pixels where the project requires precision.
Keep the source and outputs organized. The source image and generated layers are stored privately until they are deleted, which makes it easier to return to a separation later.

What this workflow does—and does not do

Layer separation recovers useful raster components from a finished image; it does not recreate a Photoshop, Figma, or other project file. Results can vary by image, especially when several elements share similar colors or overlap heavily.

Pricing is usage-based: each output image costs 2 Credits, with monthly plans and one-time Credit Packs available. That model can work well when you only need to rescue a few assets from an archive or adapt an existing visual for a new format.

The main benefit is speed. Instead of redrawing every visible component from scratch, you get named, positioned layers that can be inspected and refined in the tool you already use.

Designing a Keepsake-First Digital Letter Experience

milo zhang — Mon, 05 Oct 2026 10:27:57 +0000

Most messaging products optimize for speed. Garden Letters started with the opposite question: what if sending a digital note felt more like giving a small gift?

The product flow is built around four deliberate choices:

Composition: floral paper, fonts, and decorations make the writing surface feel personal.
Mood: an optional watercolor background is generated from the letter’s mood.
Delivery: recipients open a sealed-envelope experience, with optional preset music.
Sharing: a letter can be published publicly or shared privately with a code.

A few product constraints shape the experience:

Composing and previewing are free.
Publishing costs $4.99 USD per letter as a one-time payment. That unlocks five successful AI background generations, sharing, long-term storage, and an HD image download with a watermark.
Private letters do not appear in the public garden and can be protected with a share code.

The broader product lesson is simple: if the emotional intent matters, it should be visible in the interface and in the privacy model. A “send” action can be technically successful while still feeling wrong if the recipient experience is treated as an afterthought.

Garden Letters is available at https://gardenletters.net/create. This is a small example of designing for meaning and anticipation, rather than only optimizing for message throughput.

QwQ AI: source-grounded web research for technical and financial questions

milo zhang — Sun, 04 Oct 2026 19:39:23 +0000

I’m sharing QwQ AI, a browser-based AI search and research assistant for questions where following sources matters.

Three focused tools

Answer provides quick responses grounded in live web sources.
Research investigates topics across multiple sources with Lite, Standard, Deep, Exhaustive, and Frontier levels.
Finance Research focuses on company filings, earnings, and market information.

The workflow is straightforward: ask a question, choose the level that fits the task, inspect the cited sources, and download the result as Markdown. The site supports English, Chinese, Vietnamese, Japanese, Spanish, and Portuguese.

Pricing and fit

QwQ AI uses a freemium model. Answer queries cost $0.02 each; Research ranges from $0.04 to $3.60; and Finance Research ranges from $0.35 to $1.50 per query. There are no subscriptions or automatic renewals. It may be useful for developers, researchers, analysts, founders, and other knowledge workers who want source-backed research in one browser workflow.

A note on data handling

The product configuration says QwQ AI does not save questions or report content in the application. Long-running research task references are available for up to 24 hours, and the upstream provider may temporarily retain task content under its own policies.

This is a product introduction, not a benchmark or endorsement. Check the current pricing and policies on the QwQ AI website before using it.

Exploring Structured Decision Workflows with Decisions API

milo zhang — Sun, 04 Oct 2026 15:48:44 +0000

Many AI applications need an answer that can be inspected, compared, and routed—not only a block of generated text. Decisions API is an independent playground and API for working with decision models from multiple providers.

One workflow for multiple models

The service brings several decision models into one place. You can choose from Jev 1.13, Liquid d1, Solar Decide, Tev1 4B Experimental, Kev 4B, Span-01, and Span-01 Lite, then compare how they respond to the same input.

The workflow accepts text or JSON state and supports structured Choice, Score, or Noul questions. The returned answers are typed and include probability or confidence information. That makes it easier to inspect a result as a decision signal instead of treating every response as unstructured prose.

Examples that make the workflow concrete

The playground includes examples for support and agent routing, citation verification, content moderation, lead scoring, and LLM guardrails. These examples show how a decision model can be used to classify, score, or route information while keeping the output structured.

From playground to API integration

For server-side use, Decisions API provides a SystemOne endpoint that accepts POST requests with a Bearer API key. The documentation includes cURL, JavaScript, and Python examples, so a developer can move from a quick model comparison to an application integration with the same general workflow.

You can explore the Decisions API playground and browse the available decision models. The service describes its workflow simply: “Text in. Decisions out.”

Credits and pricing

New accounts receive 100 welcome credits. One-time credit packs start at $10, credits do not expire, and $1 buys 10,000 credits. Billing is based on model input usage, output tokens are not charged, and successful requests have a one-credit minimum.

For developers evaluating model behavior, structured outputs, or routing ideas, Decisions API offers a focused place to compare models and inspect typed answers before wiring the workflow into a larger system.

Huobidex: a curated directory for tools and products, reviewed before listing

milo zhang — Mon, 28 Sep 2026 07:40:01 +0000

Huobidex is a curated navigation directory for tools and products. Its catalog covers 5,000+ handpicked tools across 10+ use-case categories, and every listing goes through a review before it gains visibility.

What you can do on Huobidex

Search, sort, and compare tools in one place instead of scrolling through endless tabs
Browse 15 categories: AI, Analytics, Automation, Design, Developer Tool, E-commerce, Education, Entertainment, Finance, Health, Marketing, No-code, Productivity, Social, and Utility
Read why each product exists — every listed product states its user, the problem it solves, and its value

How ranking works

Rankings combine editorial quality judgment, freshness, and measurable engagement. Low-signal submissions are deprioritized, and there are no hidden pay-to-rank shortcuts.

Who it is for

Makers: submit your product, get reviewed, and reach people actively looking for tools
Teams: compare alternatives and make decisions faster

The site is updated daily. Listings are free; one-time paid plans are available for placement and faster review (from $6.9).

Site: https://huobidex.com/

Disclosure: this post was submitted by the Huobidex team as part of announcing the directory to the developer community.

Turning flat images into editable layers with AI layer decomposition

milo zhang — Sun, 27 Sep 2026 15:15:12 +0000

Flat images are easy to make and painful to reuse. Change one element — swap a product, recolor a shape, move a character — and you're either hunting for the original PSD or masking pixels by hand.

Layer decomposition attacks the problem at the source: instead of treating an image as a grid of pixels, an AI model splits it into semantic, editable pieces.

What the model does

Qwen-Image-Layered (released by Alibaba's Qwen team in December 2025, arXiv 2512.12299) decomposes a single image into multiple RGBA layers. Each layer is a full-resolution, transparent-background component — object, text, background — that you can move, recolor, or replace independently, then recompose back into the original image.

The interesting part is the training objective: the model learns to produce layers whose recomposition reproduces the input, so decomposition quality is directly measurable (reconstruction fidelity + semantic separation).

What that unlocks in practice

Design reuse: pull a product shot out of a banner without a clipping path
Transparent PNG export: every layer exports with alpha, ready for Photoshop, Figma, or Canva
Batch asset work: e-commerce and ad teams can regenerate variations from one master image
Downstream editing pipelines: layers are structured data, so they feed cleanly into editors and automation

Try it

We run a hosted implementation at ImageLayered: upload a flat image, get named, editable layers back. Quick mode starts at \$0.05 per layer, Precision mode decomposes up to 16 layers, and new accounts get free credits to test it. If you want to go deeper, the core model (Qwen-Image-Layered) is available for self-hosting.

If you build anything with image layering — or you have opinions on where decomposition beats segmentation — I'd love to hear about it in the comments.

I Built PSL Scale: AI Facial Attractiveness Scoring with a Free On-Device Scan

milo zhang — Sun, 27 Sep 2026 00:22:53 +0000

Facial attractiveness scoring sounds like a toy idea, but it touches real computer-vision problems: facial landmark detection, geometric ratios, and the tricky part — turning measurements into a score that holds up. We built PSL Scale to explore exactly that space.

What PSL Scale does

Upload a clear, front-facing photo (JPG, PNG, WebP, up to 10MB) or use your live camera, and you get a PSL score on a 0–8 scale — the community-standard perceived-attractiveness scale used in looksmaxxing circles — broken down across eight dimensions:

Symmetry
Harmony
Proportions
Skin quality
Facial structure
Averageness
Sexual dimorphism
Memorable features

Scores map to seven tiers from VERY LOW RANGE up to TRUE ADAM/EVE, with CHAD/STACY, CHADLITE/STACYLITE, HTN/HTB and MTN/MTB in between.

Free on-device scan, credit-based AI reports

The geometric scan runs entirely on your device — free and unlimited, no upload needed for the measurement step. The full AI analysis (cross-verified score, PSL tier and personalized improvement advice) is unlocked with credits: new accounts get 10 welcome credits, packs are one-time payments (Starter 20 / Standard 80 / Professional 300 credits) that never expire, and failed AI requests are never charged.

More than a scorer

The site grew into a small suite: True Adam Test, True Eve Test, Face Report (three-photo aggregate), Face Rating Deep, AI Attractiveness Test, Hunter Eyes, Golden Ratio Face, Face Shape Detector and PSL Rating Compare — plus calculators for canthal tilt, facial ratios, eyebrow mapping, jawline analysis and a true mirror, alongside a wiki explaining the concepts.

Privacy first

Photos are not stored. Evaluation history and accounts can be permanently deleted anytime.

Natural improvement over surgery

Everything on the site promotes natural, non-surgical approaches — mewing, posture correction, facial exercises, body composition changes and grooming — rather than medical procedures.

Try the free scan at pslscale.com — happy to answer questions about the pipeline in the comments.