“Add rate limiting
to search, like login.”
Clear to you. The agent still has to find the files, guess the pattern and work out how to check its own work.
THE INTENT IS THEREV8 HONE ADVANCED PROMPT / 0.5.20
It knows
your project.
Type a rough request in Claude Code or Codex and press one key. v8 rewrites it in place as a short, outcome-first brief grounded in your repository: real file paths, your actual test command, and questions only when they matter.
Start with v8MADE FOR YOUR FLOW
Claude Code/Codex
What install, setup and the first Ctrl+G look like
npm install -g v8-advanced-prompt
added 1 package
v8-prompt setup
Claude Code and Codex will use v8-prompt on Ctrl+G.
Get started
- claude or codexopen a NEW terminal and start your agent
- Ctrl+Gturn what you typed into an advanced prompt
- v8-prompt loginconnect v8 Pro (optional)
add pagination to the orders list endpoint, it returns everything right now
## Goal
Add pagination to the orders list endpoint so it returns a limited set of orders per page instead of all orders.
18 more lines: Context, your request as written, Requirements. Show full promptHide full prompt
## Context
- Before changing anything, read the code involved and the repo's docs and conventions (@docs/decisions/0004-list-pagination.md, @README.md); what they require is part of done. Leave the documentation as it is unless my request or those conventions ask for a change.
- The listOrders handler is defined in @src/routes/orders.js and currently returns all orders.
- The application is created in @src/app.js via createApp and started in @src/server.js.
- Tests for orders are in @test/orders.test.mjs.
- Documentation about pagination exists at @docs/decisions/0004-list-pagination.md and may describe expected behaviour; check it by path.
- Follow the existing pattern in @docs/decisions/0001-error-format.md.
## Material: My request, as written (where this brief reads differently, follow my request)
"""
add pagination to the orders list endpoint, it returns everything right now
"""
## Requirements
- Apply the pagination parameters to the orders data source, returning only the requested slice.
- Provide sensible defaults when parameters are omitted.
## Done when
- [ ] The endpoint returns the correct subset of orders based on page and limit parameters.
- [ ] Default pagination works when no parameters are supplied.
- [ ] `npm test` passes
- [ ] Report the commands you ran and their results.
Advanced prompt ready. Review it, then press Enter.
Press Ctrl+G again right away to get your original back.
npm install -g v8-advanced-promptThe welcome screen is the CLI's own, drawn as text. The advanced prompt is a recorded run (compiler c2026.10.28, 2026-10-15, repository amb-orders-pagination), replayed as it came back with 18 middle lines folded (Show full prompt opens them, word for word); the wait is shortened. Yours uses your repository's paths and commands.
ONE COMMAND. YOUR OWN TERMINAL.Start free. No account needed. Node.js 20+. Install guide ↗
npm install -g v8-advanced-prompt01 / GROUNDED IN YOUR REPOSITORYSCROLL TO EXPLORE
02 / THE DIFFERENCE
Your repo, not a template.
Before the first line.
A generic prompt leaves the agent to search, guess and invent a test command. v8 reads a few facts about the repository you are in (names and commands, not your code) and writes them into the brief.
“Add rate limiting
WITH V8 HONE ADVANCED PROMPT
01
Your files, by path.
The route, the limiter to copy and its tests, named from your repository so the agent starts in the right place.
02
Done when your checks pass.
Observable outcomes, ending with the test command your project actually uses: pnpm test, pytest, cargo test.
03
Questions only when they matter.
At most three, and only for unknowns that change the result. Simple requests stay two to four lines.
Example prompt structure. Actual output depends on your request and repository, and requires your review.
MEASURED, NOT CLAIMED
The agent starts in the right place.
Fewer turns, same solve rate.
We run real coding agents on tasks graded by hidden tests, written by people who never saw v8. On those, Claude Code with Sonnet took 9 to 42% fewer turns with v8's brief. The share of tasks solved did not clearly change (−6 to +2 points): on clean repositories the agent solves most tasks from the raw request already.
- Turns, Sonnet
- 9 to 42% fewer
- Solve rate, Sonnet
- −6 to +2 points
- Cost per run, Sonnet
- 3 to 24% higher
- Time per run, Sonnet
- 83% shorter to 40% longer
The brief is longer than your request, so each run costs a little more. Before c2026.10.27, agents also edited documentation more often. Codex is supported but not yet measured. In 12 other languages (144 translated tasks), Claude Code with Haiku solved 92% with v8's brief and 89% from the raw request, a difference within noise, and the brief tells the agent to answer in your language. When a request leaves out a value only you know, such as the new name in a rename or the limit for a new rule, v8 asks that one question: on 12 held-out tasks of this kind, Claude Code with Sonnet solved 50% with the answer and 17% without it. Every study, win and loss ↗
03 / YOUR WORKFLOW, WITH CONTEXT
One shortcut.
Still your workspace.
No extra editor. No new conversation to manage.
Write it your way.
Type what you want in Claude Code or Codex. A rough thought is enough. Follow-ups like “now do the same for settings” work too when Claude Code shows its last reply in the editor.
It sends names, not your code.
Press Ctrl G and v8 searches your repository on your device, then sends a few facts: stack, test command, matching file paths and the function names in them, recent commit subjects. The prompt service writes the brief back into the same input box.
Answer, or skip.
If an unknown would change the result, you get up to three numbered questions in the terminal, in Claude Code or Codex; 0 skips. Answers are remembered for that repository, and answering is free.
You make the call.
Review before Enter. Press the shortcut again to restore your original words.
04 / ROOM TO BUILD
A small tool.
A place in your flow.
Start free. Go Pro when it becomes
part of your everyday work.
Free
NO ACCOUNT NEEDED$0 / free plan
- 10 advanced prompts a week · 3 a day
- Claude Code and Codex
- Answering its questions is free
- Undo with the same shortcut
Pro
FOR YOUR DAILY FLOW$39 / year · or $4.99 / month
- Up to 350 advanced prompts a week: about 50 a day, with no daily cap
- Every terminal and machine you connect, on one allowance
- Reserved capacity for Pro; service and provider limits still apply
- Answering v8's questions is free
- Cancel anytime from your account; 14-day refund on first payment
A few things to know.
How does it know my project?
When you press the key, the CLI collects a small context pack from the repository you are in: its name, branch and whether there are uncommitted changes; the stack from your manifests; the test, build, lint, typecheck and run commands; which instruction files exist and which of them govern the files involved (CLAUDE.md, AGENTS.md, .cursor/rules); up to 8 file paths most relevant to your request, their test files, up to 2 existing files to follow as a pattern and up to 2 documentation files; up to 4 commands your CI runs (redacted); the last 5 commit subjects; and answers you confirmed earlier for that repository. To rank files, v8 searches file contents on your machine; it sends up to 8 function, method and class names from each of up to 4 of those files and their first test file (the ones that match first, then each file's top-level definitions), never the code itself. It is capped at 4 KB.
What leaves my computer?
The text in your input box and that context pack, only when you press the key, sent to our service and its AI provider, which does not use it for model training. Your computer's time zone goes along too; we keep only the country it implies. No file contents: not your code, not your .env. The exceptions: up to 4 command lines quoted from CLAUDE.md or AGENTS.md when your manifests list none, up to 4 redacted CI commands, and, for a follow-up in Claude Code, up to 4,000 characters of the agent's last reply. Prompt text and repository context are never stored on our servers. Run v8-prompt context show to see exactly what would be sent, or v8-prompt context off to send the prompt alone.
When does it ask questions?
When you press the key in Claude Code or Codex, only when an unknown would change the deliverable, stack, scope or data model: at most 3, with numbered options, and 0 skips. Answers are remembered on your computer for that repository (v8-prompt forget clears them). The follow-up with your answers does not count as a second prompt.
Is it different in Codex?
The brief is tuned for Codex to act rather than ask. When you press the key, v8 can still ask you up to 3 questions in the terminal if an unknown would change the result; other unknowns become explicit assumption lines in the brief, which you can correct before pressing Enter.
Is there a Claude Code plugin?
Yes. The package includes a plugin with one command, /v8:prompt <request>, which shows you the advanced prompt and, once you choose to use it, hands it to Claude as the task. It needs the CLI installed. See the install guide.
Can I use it without an account?
Yes. The free plan includes 10 advanced prompts per week per terminal, with a daily limit of 3. On a shared network, such as an office, only the first few new terminals each day get the free plan. If v8 returns your text unchanged, that prompt doesn't count, up to 20 a day per terminal.
What if I prefer my original prompt?
Press the shortcut again to restore it. Your coding agent receives the prompt only after you press Enter.
Which key do I press, on Windows, macOS and Linux?
Ctrl + G works in every terminal on every system; on a Mac it is Control, not Command. Setup also binds Ctrl + Shift + . in Windows Terminal and in the VS Code, Cursor and Windsurf terminals (Control on a Mac there too). Press the same key again to get your original text back.
YOUR NEXT IDEA IS ENOUGH TO START.
Give it direction.