Talking to Nephele

Nephele's home screen is a conversation. Unlike an ordinary chatbot, Nephele doesn't just reply with text—it actually operates your computer: opening a browser to search for images, organizing your Eagle library, batch-packaging your work, running Python scripts to process files.

This page isn't about how it works under the hood—it's about how you actually use it once the app is open.

Working with the interface

One input box, three layers of routing

You never have to switch "modes" by hand. Just type what you want to do into the home input box, and Nephele routes it automatically:

1. Rule engine (local · zero latency) — Explicit commands like "open Photoshop" or "open the Downloads folder" run directly, without going through AI, without a network call, and without spending any credits.
2. Guide channel gate (free · lightweight model) — While you type, a lightweight model takes a first look: light questions like how to use the app, navigation, or one-line facts get answered for free on the spot, without spending stamina. By design it leans toward letting things through, and on any failure it always passes you through—it will never block your input.
3. Cloud MAX (full Agent) — Requests that genuinely need work done enter the full Agent, which calls tools to complete the task for you and is billed in cloud credits.

In the top-right of the toolbar there's also a "?" guide assistant you can ask "how does this feature work" at any moment—it's driven by the same lightweight model, it's free, and when it can't answer or you want to go deeper it carries your exact words and "escalates" to Cloud MAX. More on that below.

Dragging images onto the input box

Drag an image file straight onto the input box and a 60×60 thumbnail appears. Click the × in its top-right corner to remove it. PNG, JPG, JPEG, WEBP, BMP, and GIF are supported.

The dropped image is sent to the Agent as an attachment alongside your text. For example, drag in an image and ask "write me a caption suited for posting on Pixiv," and Nephele will generate the content based on that image.

Stopping generation

While Nephele is producing output, a stop button (a square icon) appears at the right of the input box. Click it and the current stream is interrupted, showing "[Stopped]".

The Skill system: why Nephele can "understand" your intent

You may have noticed that the same instruction sent to different AIs gives wildly different results. Nephele's secret isn't a smarter model—it's Skills.

What a Skill is

A Skill is a plain-text file (SKILL.md) that describes in natural language "when the user says X, here's what you should do." It isn't a plugin, it needs no installation, and it involves no code—it's essentially a "work manual" for the AI.

A Skill looks like this:

---
name: find-references
description: 找参考图
emoji: "🎨"
tools: [find_references, find_similar_by_image]
keywords: [参考图, 找图, 灵感, 搜图, 风格参考, Pixiv, 以图搜图]
---
 
## Decision tree
 
- 关键词 → 调用 find_references（Pixiv+Eagle 并行）
- 图片 → 调用 find_similar_by_image（Eagle 视觉相似度）
- Pixiv 标签必须用日文：初音未来 → "初音ミク"，赛博朋克 → "サイバーパンク"

When you say "find me some mecha references," Nephele will:

1. Match a Skill: scan the keywords of every Skill, hit "reference" and "mecha" → load find-references
2. Inject the prompt: drop that Skill's full strategy into the AI's system prompt
3. Execute: the AI not only knows which tools to call, it also knows to use Japanese for Pixiv tags and not to search with Chinese

Built-in Skills

There are four built-in Skills right now:

trace-artwork-origin has a thoughtful design: if you drop in an image and ask "who drew this," Nephele looks at the image first—if it recognizes the character or art style, it tells you for free; only when it really can't tell, or when you want a specific URL, does it call a paid reverse image search engine. Not to save money, but to avoid spending it for nothing.

Where Skills come from

Nephele loads Skills from three sources, in order of priority from high to low:

1. Cloud push (remote) — delivered through remote config, with no action needed from you. A same-named Skill overrides the two layers below. Good for emergency strategy fixes or pushing new capabilities.
2. User custom — C:\Users\<username>\.nephele_workshop\skills\ (or the corresponding user directory on your platform). A same-named Skill overrides the built-in one.
3. Built-in — the skills/ folder that ships with the app.

If you create a new folder under the layer-2 directory and put a SKILL.md inside it, Nephele will load it automatically on the next launch.

Coming up: a UGC Skill marketplace

We're planning a Skill marketplace. You'll be able to share the SKILL.md files you write with other illustrators, and download ones others have written—like "Pixiv operations strategy," "Comiket exhibitor checklist," or "a guide to analyzing a specific art style."

The beauty of a Skill is that it doesn't depend on coding ability. Can't program? No problem—as long as you can describe your own workflow, you can write a useful Skill.

Cloud mode (Cloud MAX): what it can actually do

This is Nephele's full form. What's listed below is already implemented in the code and can actually be called.

Searching for references

You can just say:

• "Search Pixiv for some cyberpunk illustrations"
• "Find Japanese-style character design references on Pinterest"
• "Search Eagle for images tagged 'mecha'"

Nephele opens the built-in browser, runs the search, and returns the results as image cards in the conversation. You can check off the ones you want and save them to your computer or to Eagle with one click. If you're not happy with the results, just keep going: "show me a different batch" or "only vertical compositions."

Boundaries: Pixiv search goes through a cloud node; Pinterest, ArtStation, and Huaban go through your own Edge browser via the Nephele Wisp browser extension (the extension must be online—there's no Playwright fallback); Konachan, Eagle, and Wikimedia Commons are also available sources. By default it only searches your local Eagle library; network sources have to be specified explicitly. If the target site is unreachable or pops a CAPTCHA, the search may fail.

Reverse image search (local Eagle)

Upload an image and say "find images in Eagle similar to this one," and Nephele uses the visual embeddings produced as a by-product of tagging to compute cosine similarity, layers in tag-overlap to rank, and returns the most similar assets in Eagle.

Prerequisites:

1. These images must first have been through library indexing (visual embeddings are written during tagging); the first indexing pass takes a few minutes
2. Only already-indexed images can be matched by reverse image search

Eagle asset management

• "Import every illustration-type image in the Downloads folder into Eagle"
• "Move the images tagged 'mechanical' in the 'references' folder to the 'mecha references' folder"
• "Batch-add the 'to evaluate' tag to this set of images"
• "Delete the duplicate images in Eagle"

Boundaries: imports are automatically tagged by the local tagging model (if the model is loaded). Batch operations involving large numbers of files can take a while, and Nephele will show progress.

Working on artwork

• "Review this image" — calls Creative Review, which gives a letter grade (S to E) by artwork type + three-part narrative feedback (what you nailed / what you could explore / what to try next time)
• "Check whether this image is AI-generated" — calls AI Credential Detection, which scans EXIF/IPTC/generation-parameter fingerprints
• "Write me a caption suited for posting on Pixiv" — calls the multi-platform caption generation in Aggregate Upload
• "Add a watermark to this image and package it" — calls Publish Packaging, which outputs a processed image
• "Issue a digital timestamp for this image" — calls Digital Timestamp

Boundaries: these tools need you to upload an image or specify a file path. Timestamping requires a network connection (a timestamp server). Packaging writes files to a specified directory.

Running Python scripts

• "Write a script that renames every PNG in the current folder by date"
• "Count how many images are in this folder"

Nephele generates the code and displays it on a dark code card. You must click Confirm before it runs. You can edit the code before it executes. This is a safety design—Nephele will never run any code without your knowledge.

Boundaries: the script is executed in a subprocess by a separate Python interpreter in isolated mode (-I, ignoring environment variables and user site-packages); it can access the file system and the Python libraries that ship with the app (PIL, numpy, onnxruntime, etc.). For irreversible operations like deletion, Nephele first lists everything and asks for a second confirmation before executing.

System shortcuts

• "Open Photoshop"
• "Open C:\Users\Artist\Downloads"
• "What's the weather today"

Simple commands like these don't go through AI; they run directly through the rule engine with near-zero latency.

Boundaries: opening an app relies on system path resolution—if the software is installed in an unusual location, it may not be found. A weather query opens the browser to a Bing search.

What it can't do (not even in the cloud)

• Can't directly upload work to platforms — Publisher only generates captions and tags; it doesn't actually upload
• Can't modify your artwork — Nephele has no drawing features and won't touch the contents of your PSD/CLIP files
• Can't bypass site logins — if a site like Pinterest requires login, Nephele can't handle it automatically
• Can't guarantee 100%-accurate AI detection — AI Credential Detection relies on metadata, which can be tampered with

The guide channel: ask "how do I use this" for free

Not every question is worth invoking the full Agent. The guide channel is driven by a lightweight model (Axioma Zephyr), dedicated to handling light "how do I use the app" questions for free, without spending stamina (subject to a daily cap). It shows up in two places:

The toolbar "?" assistant

Click the "?" in the top-right of the toolbar to open a guide overlay. It has two ways to use it:

• Option-card flow (chips) — a set of preset guide cards, purely local, no AI call; whichever card you tap takes you down that path. Good for deterministic guidance like "I want to do X, where do I click first."
• Free input — just type your question and get a streamed answer. It knows which screen you're currently on, so the answer is "in context." When it can't answer, or you say "tell me more / do it for me," it carries your exact words and escalates to Cloud MAX (asking you to confirm before escalating).

The gate in front of the main input box

When you type into the home input box, the guide channel also takes a first look as a "gate." App usage, navigation, one-line facts, and questions that need one round of clarification get answered for free on the spot; everything else is passed through to Cloud MAX.

Consumption and limits

Stamina and Nepheline (cloud credits)

Nephele has two independent resource pools—don't mix them up:

Stamina

• A daily free allowance during the test period, refreshed automatically every day
• Used to try out cloud features, no payment required
• Stops when you've used it up for the day, recovers the next day

Nepheline (cloud credits)

• A paid resource pool, included with your annual license + top-up-able
• Nepheline purchased via top-up never expires; the annual grant is valid during your license term, frozen when it lapses, and restored when you renew
• The test period (including the current Beta) keeps the daily free allowance; the free policy will change after the full release, per the announcement at that time

What is and isn't consumed:

• Shortcut commands handled by the rule engine: zero consumption
• Light questions answered on the spot by the guide channel: free, no stamina deducted (subject to a daily cap)
• Requests that enter Cloud MAX: billed in cloud credits; multi-turn tool tasks cost more

Tool-call limit

Cloud MAX runs at most 20 tool calls per message. Beyond that it wraps up automatically. This is a hard limit, to prevent accidental consumption and infinite loops.

Context window

Cloud MAX is driven by Axioma Breeze, with a 1M-token context that usually won't overflow. Single-output is capped at 8192 tokens—a guardrail against runaway consumption, not a ceiling on the model's capability.

Large-result compression

If a tool returns a very large raw result (say, a reverse search or a batch of references), the server first uses a lightweight model to compress it into the factual summary the main model needs to make decisions, then feeds that to Cloud MAX. Compression keeps only the key information (failure reasons, source trustworthiness, the next blocking point, etc.) to prevent long text from blowing out the context window. If compression fails, the raw result is kept and the conversation isn't interrupted.

FAQ

Q: Why does Nephele sometimes run the same tool repeatedly?

If a tool returns an error, Nephele will try to correct it and run again. But if it runs the exact same tool call twice in a row, the loop terminates automatically to prevent it from getting stuck.

Q: Do simple questions deduct from my credits?

No. Explicit shortcut commands go through the local rule engine at zero consumption; light "how does this feature work" questions are answered for free by the guide channel, with no stamina deducted. Only requests that genuinely need the Agent to do work enter Cloud MAX billing.

Q: How much can I do without the cloud AI?

The vast majority of things. Nephele embeds several local AI models (auto-tagging, reverse image search, cross-lingual semantic search); tagging, indexing, image search, semantic retrieval, watermarking, timestamping, rights enforcement, and file processing all run on your own computer and work offline. AI Credential Detection relies on locally parsing image metadata/credentials (not cloud inference), so it also works offline. What needs a network connection is mainly Agent capabilities like natural-language conversation with Nephele (having it plan tasks, call tools, search the web).

Q: Can I have multiple conversations open at once?

Yes. In the session-history panel that slides out on the left, click "New session" to start a new conversation. History isn't shared between sessions.

Q: Does Nephele save my chat history?

Conversation history is saved in a local database and isn't uploaded to a server. But cloud-mode AI inference does pass through Axioma's servers—that's a necessary condition for using cloud features.