AI & Ghost Suggestions
Flow can suggest content as you type, explain errors, and nudge you toward better writing — all configurable, all optional.
What are ghost suggestions?
Ghost suggestions are inline completions that appear as dimmed, italic text after your cursor. They're called "ghosts" because they're transparent — you see them, but they're not actually in your document until you accept them.
For example, if you're filling in a \title{} field and start typing "An Intro", Flow might suggest "duction to Machine Learning" in faded text. You choose whether to take it or ignore it.
Accepting and dismissing
| Action | Shortcut |
|---|---|
| Accept suggestion | Ctrl+Right Arrow |
| Dismiss suggestion | Escape |
| Manually request a suggestion | Ctrl+Right Arrow (when no ghost visible) |
You can also just keep typing — the ghost disappears as soon as you press any character key.
How to enable ghost suggestions
Ghost suggestions are off by default. To turn them on:
- Open Settings (gear icon in the sidebar).
- Go to the AI Apprentice section.
- Toggle on Ghost Suggestions.
- Optionally, toggle on Freeform Suggestions to also get suggestions for prose text outside template fields.
You'll need to have at least one LLM provider configured (see below).
Fillable vs. freeform modes
Ghost suggestions work in two contexts:
- Fillable mode — Suggestions for structured template fields (the default). The AI uses the field type, hint, and surrounding context to suggest appropriate content. For instance, it knows that
\email{}should contain an email address. - Freeform mode — Suggestions for general prose text. This is more open-ended — the AI reads your surrounding paragraphs and continues your thought. Must be enabled separately in settings.
Fillable mode uses shorter, more constrained completions (up to 50 tokens). Freeform mode allows longer continuations (up to 150 tokens) with more creative temperature.
Configuring your LLM provider
Flow supports several cloud providers and local models. You pick one in Settings > AI Apprentice.
Anthropic (Claude)
Flow's default cloud provider. Available models:
- Claude Sonnet 4 — Latest, recommended for most users
- Claude 3.5 Sonnet — Previous generation, still excellent
- Claude 3.5 Haiku — Fastest, good for quick completions
Requires an Anthropic API key. Get one at console.anthropic.com.
OpenAI
Supports GPT models. Enter your OpenAI API key in settings.
Google Gemini
Supports Gemini models including Gemini 2.0 Flash. Enter your Google AI API key in settings.
Z.AI (Zhipu)
Supports GLM models. Enter your Z.AI API key in settings.
For all cloud providers, Flow validates your API key with a test request when you save it. If the test fails, you'll see an error explaining what went wrong.
Local LLM setup
If you prefer to keep everything on your machine, Flow can run local models using llama.cpp. No API key needed, no data leaves your computer.
- Go to Settings > LLM Providers.
- Select Local as your provider.
- Click Browse & Download to see available models.
- Choose a model based on your hardware. Options range from compact (~1 GB) to powerful (~20 GB).
- Once downloaded, click Load to start the model.
Flow auto-detects your GPU and uses the best available acceleration:
- Metal — Apple Silicon Macs
- CUDA — NVIDIA GPUs
- Vulkan — AMD and other GPUs
- CPU — Fallback, slower but always works
Available local models
| Model | Category | Size range |
|---|---|---|
| Llama 3.2 1B Instruct | Compact | 0.77–1.32 GB |
| Llama 3.2 3B Instruct | Compact | 2.02–3.42 GB |
| Phi 3.5 Mini Instruct | Balanced | 2.39–4.06 GB |
| Qwen 2.5 7B Instruct | Balanced | 4.68–8.10 GB |
| Llama 3.1 8B Instruct | Balanced | 4.92–8.54 GB |
| Mistral 7B Instruct v0.3 | Balanced | 4.37–7.70 GB |
| Qwen 2.5 14B Instruct | Powerful | 8.99–15.7 GB |
| Qwen 2.5 32B Instruct | Powerful | 19.9–23.3 GB |
Each model comes in multiple quantization levels (Q4_K_M, Q5_K_M, Q8_0) — smaller quantizations use less memory but sacrifice some quality.
Auto mode
Set your provider to Auto to get the best of both worlds. Flow will use your local model when it's loaded, and fall back to your cloud provider when it isn't. This way you get fast, private completions when possible and reliable cloud completions as a backup.
Writing nudges
Separate from ghost suggestions, Flow can also show writing nudges — subtle suggestions that appear when it detects common writing issues. These are especially useful for resumes and cover letters.
The eight nudge types:
- Missing metric — Achievement statement without quantified results ("Improved performance" → add specific numbers)
- Inconsistent dates — Multiple date formats in the same document
- Thin section — A section that's significantly shorter than the rest
- Weak opening — Generic phrases like "I am writing to..." or "To whom it may concern"
- Long bullet — Bullet points that are too wordy (over ~160 characters)
- Missing verb — Resume bullets that don't start with an action verb
- Passive voice — Passive constructions where active voice would be stronger
- Missing required — Template fields that were skipped
Some nudges can be auto-fixed with one click; others are informational. Enable them in Settings > AI Apprentice > Writing Nudges.
AI error explanations
When compilation produces errors, Flow can ask your configured AI to explain the error in plain English. This shows up as an "Ask AI" button next to each error message.
You can also turn on Auto Ask AI in Settings > AI Apprentice to have this happen automatically for every error — no clicking required.