One hotkey, anywhere on Windows
Anvyll lives in the tray. Press Ctrl+Space in Cursor, a browser, Word, Notion or a terminal and the overlay opens over whatever you were doing.
For Windows · Ctrl+Space from anywhere
Claude wants XML tags. GPT-5 wants numbered steps. Anvyll rewrites what you already typed using the target model's own methodology, then puts it back in the field you took it from.
How it works
Anvyll sits in the tray doing nothing until you press the hotkey. It reads the field you're focused on, rewrites what's there, and writes it back.
Put the cursor in any text field and press the hotkey. Anvyll grabs your selection, or the entire field if nothing is selected.
Pick the model you are aiming at and how hard to push the structure. The matching methodology pack does the rest.
Press Enter and the refined prompt replaces what you captured. If a secure field blocks injection, Copy is always there.
Why it isn't just "make this better"
Anvyll ships a prompt pack per target model, each carrying that vendor's documented guidance, an exemplar and its rubric weights. Pick the model you're aiming at and the structure changes with it.
Your input write something about our new pricing page for the blog
<role>
You are a B2B content strategist who has shipped
pricing-page launch posts for developer tools.
</role>
<task>
Draft a 600-word launch post announcing the new
pricing page. Lead with what changed for existing
users, not the feature list.
</task>
<constraints>
- Sentence case headings, no exclamation marks
- Name the actual price and what it replaces
- One concrete before/after example
</constraints>
<output_format>
Markdown. H2 sections. No preamble.
</output_format> Anthropic's guidance leans on explicit XML delimiters and a role set before the task, so the rewrite emits tagged blocks rather than headers.
# Role
B2B content strategist. You have shipped pricing-page
launch posts for developer tools.
# Task
Draft a 600-word launch post announcing the new
pricing page.
# Steps
1. Open with what changed for existing users.
2. State the new price and what it replaces.
3. Give one concrete before/after example.
4. Close with a single next action.
# Constraints
- Sentence case headings
- No exclamation marks, no hype adjectives
# Output
Markdown, H2 sections, no preamble. OpenAI's guidance favours a flat instruction hierarchy with numbered reasoning steps and the output contract stated last.
**Persona**
A B2B content strategist specialising in developer-tool
pricing launches.
**Context**
The product is a Windows desktop utility. Pricing moved
from usage credits to a flat monthly plan. The audience
is existing users who need to know what changed.
**Task**
Write a 600-word launch post for the company blog.
**Constraints**
- Lead with the change, not the feature list
- Name the actual price
- One before/after example
- Sentence case, no exclamation marks
**Format** Markdown with H2 sections. Google's guidance front-loads persona and context before the instruction, and asks for the constraint block to sit adjacent to the task.
## Instruction
Write a 600-word blog post announcing a new pricing
page for a Windows developer utility.
## Requirements
- Audience: existing users
- Lead with what changed, not the feature list
- State the new monthly price and what it replaces
- Include one before/after example
- Sentence case headings; no exclamation marks
- Output: markdown, H2 sections, no preamble
## Voice
Plain and specific. Assume the reader already knows
the product. DeepSeek responds best to dense, low-ceremony instructions — the pack strips decorative structure and keeps the requirements as a tight list.
**Brief**
Announce our new pricing page in a 600-word blog post.
**Who's reading**
Existing users of a Windows desktop utility who want to
know whether this costs them more.
**What matters**
- Say what changed, up front, without softening it
- Name the price and what it replaces
- One before/after example
- Take a position on why flat beats credits
**Avoid**
Hype adjectives, exclamation marks, title case.
**Format** Markdown, H2 sections. Grok's pack keeps the brief conversational and explicitly licenses a point of view, which the model handles better than rigid scaffolding.
## Deliverable
A 600-word markdown post at content/blog/pricing.md
## Task
Announce the move from usage credits to a flat monthly
plan.
## Content requirements
1. Open with the change as it affects existing users
2. State the new price and what it replaces
3. One before/after example
4. Close with a single next action
## Style
- Sentence case headings
- No exclamation marks or hype adjectives
- Frontmatter: title, description, date Composer is a coding-surface model, so the pack front-loads the concrete deliverable and file target before any prose direction.
The target model selects whose methodology gets applied. Rewrites run server-side through the Anvyll proxy, so you never wire up six providers.
What you get
The whole point is that refining a prompt shouldn't cost you a context switch, a browser tab, or your clipboard.
Anvyll lives in the tray. Press Ctrl+Space in Cursor, a browser, Word, Notion or a terminal and the overlay opens over whatever you were doing.
Selected text is captured if present, otherwise the whole field. Apply writes the result straight back where it came from.
Anvyll snapshots what you had copied and puts it back afterwards. Nothing you were holding gets clobbered.
Every rewrite reports a 1–100 rubric score and a line diff tagged by what changed — added role, added constraints, tightened scope.
Each model's pack carries its vendor guidance, an exemplar and rubric weights. Guidance updates ship as data, not as a new build.
Settings, library, history and project memory live in a per-user encrypted vault. Sign in on a new machine and your state follows.
No per-provider API key needed. Bring your own OpenAI key if you prefer, or fall back to the local optimizer with no account at all.
Where your text goes
Anvyll reads whatever field you're focused on, so it matters that you know exactly what leaves the machine.
Pricing
One plan, one price, billed in the app through Stripe. No seat maths, no usage dashboard to babysit.
€0
A one-time grant of 1,000 credits — roughly 50 refines. No card.
Works without an account
€13.99 / month
VAT inclusive. Cancel at period end from the Stripe portal, in the app.
Trial starts in the app, after install
Questions
No. Signing in routes rewrites through the hosted Anvyll proxy, which calls the model server-side — no key ever lives on your machine. If you would rather use your own OpenAI key, Settings accepts one and stores it through Windows Credential Manager.
Yes, for now. Anvyll depends on Win32 capture and injection to read and write the field you are focused on. Windows 10 or later.
Expected until Azure Trusted Signing finishes business-identity verification. Choose More info, then Run anyway. It means the certificate chain is not established with Microsoft yet, not that the installer is compromised.
Nothing breaks. The local fallback optimizer keeps producing structured, model-specific prompts with no account and no key. Unlimited removes the metering entirely.
Reassign it in Settings — the hotkey is not fixed. If the overlay does not open, check that Anvyll is running in the tray and that no other app has claimed the combination.
No. Prompts, sessions and projects stay in an encrypted vault on your device. Our servers hold account, billing and token-metering data. Rewrite text passes through the proxy to reach the model and is not retained.
Install it, press Ctrl+Space in whatever you're already working in, and see what comes back.
Windows 10 or later · 1,000 credits included · no card to start