Drafts replies and posts worth sending, in your own voice.
A Chrome extension that drafts three replies to the post you are reading, or three original posts from an idea.

---

What it does

Click Riposte on any post to draft replies, or hit the floating launcher to write something of your own. Either way you get three drafts that take genuinely different approaches. Each one is labelled with the angle it took, counted against your character limit, and one click away from being in the reply box.

The point is not to produce more replies. It is to produce replies that are worth reading:

• It writes in your voice. Paste your own posts once and it copies your rhythm, punctuation, and level of formality. This is the single thing that stops the output reading like everyone else's AI.
• It refuses to pad. The prompt bans the whole register of AI slop: no "This.", no "Great point", no restating the post back at its author, no engagement-bait question tacked on the end. When it has nothing specific to add it writes something shorter rather than something longer.
• It writes your own posts too. Give the launcher a rough idea and it drafts three takes on it, with a separate set of angles and its own list of banned X-post tics.
• It reads the whole thread. Replying to a comment pulls in the post that started the conversation and the comment being answered, so the draft makes sense to someone reading top to bottom.
• It can check its facts. With web search on, it looks up a real number or source before replying, and is told explicitly that a confident wrong figure posted from your account is the worst possible outcome.

Nothing is ever posted automatically. Riposte drafts, you decide.

Requirements

You need your own Anthropic API key. Riposte has no backend and no subscription. It calls the Claude API directly from your browser using your key, and you pay Anthropic for what you use. Expect roughly 3 to 6 cents per set of three drafts on Opus 5, less on Sonnet or Haiku.

Chrome, or any Chromium browser that supports Manifest V3.

Install

From a release

1. Download riposte-v0.1.0.zip from the Releases page and unzip it.
2. Open chrome://extensions and turn on Developer mode (top right).
3. Click Load unpacked and select the unzipped folder.

From source

git clone https://github.com/generalizingai/riposte.git
cd riposte
npm install
npm run build

Then load the dist/ folder via Load unpacked, as above.

Extensions loaded this way do not auto-update. To upgrade, pull and rebuild, then hit reload on the extension card in chrome://extensions.

Setup

The options page opens on first install. Paste your API key, press Test to confirm it works, then paste 15 to 30 of your own posts into Your voice, one per line. Topics do not matter, only how you write.

Setting	Notes
Model	Opus 5 for the best writing, Sonnet 5 to halve the cost, Haiku 4.5 for speed.
Effort	How hard Claude thinks. low is snappiest, medium is the default balance.
Default angle	The approach used when you have not picked one in the panel.
Character limit	280, or the higher Premium limits. Drafts over it are flagged in red.
Web search	Grounds replies in a real fact. Slower and costs more.
Your voice	Your own writing samples. The highest-leverage setting here.

The seven angles

Switch between these from the panel at any time, and redraft without losing what you already have.

Angle	What it does
Conversational	Writes the way you would say it out loud. The one that reads least like writing.
Insightful	Adds the mechanism or second-order effect the post is missing.
Founder	Speaks from having actually run the thing: what it cost, what broke, what you would redo.
Supportive	Agrees, but earns it with a concrete example instead of applause.
Contrarian	Pushes back on the weakest part of the claim, generously.
Witty	One dry, understated observation.
Question	The one question whose answer would change how you read the post.

You can also type a free-text steer ("shorter", "push back harder", "mention the study") and redraft.

Writing your own posts

The floating launcher opens compose mode. Type a rough idea, press Write, and get three takes on it. Posts get their own angles, because "supportive" means nothing when there is no post above you.

Angle	What it does
Conversational	Says the thing the way you would say it to one person.
Observation	One sharp noticing about how something actually works.
Story	One specific thing that happened. No moral bolted on the end.
Contrarian	The part the consensus has wrong, and why.
Lesson	Something learned the expensive way, told flat rather than as advice.
Question	A real question, narrow enough to answer in one line.

Posts also get their own anti-slop list, aimed at the register X rewards and readers resent: no "Here's the thing", no "Unpopular opinion", no hook-colon-list, no thread emoji, no closing "Thoughts?", no one-sentence-per-line profundity.

The review queue

Finding posts worth answering costs more time than writing the answers. The queue takes that part.

While a feed is open, Riposte collects the posts that scroll past. Nothing is sent anywhere and no API call happens during collection. When you press Find posts, it makes one batched call to triage the whole pool, picks the few you could genuinely add something to, drafts replies for those, and queues them with a line explaining why each was picked.

You then read the batch and either Bin it or Open it, which navigates to the post and loads the draft into X's reply box. You press Reply.

Triage is deliberately reluctant. It is told to skip posts where the only honest reply is agreement, posts outside what you actually know, engagement bait, and posts already buried in replies. An empty result is a valid answer, not a failure. Fill in What you can speak to in the options or it has no basis for judging relevance.

On autonomy

Riposte will not post on your behalf, and this is a design decision rather than an unfinished feature. Every reply is one a human read and chose to send. There is no autonomous mode and no batch-post button.

Three reasons, in the order that will matter to you. Automated replying is specifically named in X's platform manipulation policy and reply-spam is what their detection is tuned hardest against. A tool whose whole premise is "replies worth sending" cannot ship a mode that sends unread replies. And everyone receiving an autonomous reply would reasonably believe a person read their post.

The queue removes the part that costs an hour. The part that costs a minute is the part worth keeping.

Privacy

Riposte has no servers. Nothing is collected, and nothing reaches the author of this extension.

• Your API key and settings are stored in chrome.storage.local, on your machine only.
• The text of a post is sent to api.anthropic.com when you ask for drafts, and nowhere else.
• No analytics, no telemetry, no error reporting.

Your key is stored unencrypted, which is standard for extension storage but worth knowing: anyone with access to your computer account can read it. Use a key scoped to this tool so you can revoke it on its own.

Full policy: docs/PRIVACY.md.

Development

npm run dev        # rebuild on change (reload the extension after each build)
npm run build      # one-off build to dist/
npm run package    # build and zip dist/ for a release

Validate the API request shape without involving Chrome at all:

ANTHROPIC_API_KEY=sk-ant-... node scripts/smoke-test.mjs

It runs the real prompt against a sample post and prints the drafts. Add --search to exercise the web search path too.

Layout

File	Responsibility
src/prompt.js	The actual product. Angles, the anti-slop rules, voice matching, output schema.
src/extract.js	Every X DOM selector, in one table. The only file a redesign should break.
src/background.js	Service worker. Talks to the Claude API, maps errors to readable messages.
src/content.js	Injects the trigger and launcher, owns the flow.
src/panel.js	The floating panel.
src/select.js	Themed dropdown, since native <select> menus cannot be styled.
src/settings.js	Defaults, shared by the worker and the options page.

If the replies come out wrong, tune src/prompt.js. Everything else is plumbing.

Known limitations

• X redesigns will break it. Everything DOM-dependent is isolated in src/extract.js, so fixes are usually a one-file change, but they will be needed.
• Insert uses document.execCommand, which is deprecated but remains the only reliable way to get text into X's React editor. There is a fallback, and a clipboard fallback behind that.
• No auto-update on unpacked installs.
• X only, for now. The DOM layer is deliberately separate so other sites can be added.

License

MIT. See LICENSE.

Not affiliated with, endorsed by, or connected to X Corp. or Anthropic.