Reply Cadence Engine

Knowing how to reply well was never the problem. Doing it every day is. The Reply Cadence Engine turns the reply workflow into a deliberate loop: point a Hunt at the conversations you want to be part of, send it hunting on X, review the drafts it stands up, and send the ones worth sending. Every draft is anchored to something you actually wrote, and everything the engine can't prove gets held back.

Sending stays yours. The X API cannot post replies, and BlackOps would not want it to. Every draft is staged into X's native composer by the Chrome extension for you to read, edit, and fire.

How it works

1. You define Hunts. A Hunt is a saved definition of a conversation you want to be part of: search keywords, a plain-English brief of who you're looking for, and what you want to convey when you find them.
2. You run a hunt. Open the Hunt Targets panel in the Chrome extension on x.com, pick a Hunt, and hit Run this hunt. One click turns the Hunt's keywords into a live search on X itself and opens it in your browser. No API quota is spent — the search runs where you're already logged in.
3. The extension scores what X returns. As results render, each tweet is scored against the Hunt's brief: is this really the conversation you're looking for, is the author a person rather than a content feed, and would a reply from you actually answer them? Matches get a Hunt badge in place. A tweet no Hunt fits still gets a relevance score if it overlaps with what you recently posted. Otherwise it gets a skip badge with the reason, and it is dimmed so the matches stand out.
4. Angles first, drafts on request. Each target gets two or three one-line reply angles, each with the fact it stands on, and you write the reply. Ask for a full draft and the full drafting pipeline runs: grounded in the Hunt's intent and your bound brains, anchored to a verbatim-verified fact from your notes or your published posts, critiqued, refined, and conformed to your brand voice. If nothing you've written speaks to the tweet, the engine says so instead of writing filler.
5. The bonus catch. Between hunts, the timeline you're already reading gets relevance scores against your own recent tweets, posts, and replies, so the occasional tweet worth answering stands out. Open one and your active Hunts are checked against it too. A matching Hunt takes it over, and otherwise the angles build on what you already said. Each tweet shows one score from one source, never two. Deliberate first, serendipity on top.
6. The queue keeps score. The Reply Queue page in your admin is the ledger: staged angles with the fact each one stands on, targets found but held back with reasons, the rules the engine has learned from you, and progress against your daily goal.

Found, but held back

Not every good target gets angles. When a tweet is worth answering but nothing concrete in your brains bears on it, the drafting gate holds it back instead of writing a generic reply. Held-back targets appear in their own section on the queue page with the tweet, its score, and why it was held.

Each one has a Draft anyway button. Click it and the full drafting pipeline runs for that single target, and the result moves into your working queue for review. Nothing is ever drafted for a held-back target until you ask — the button is the only path. You can also dismiss any of them.

Replies that carry your post

When a Hunt has bound posts, drafts can stand on facts quoted verbatim from those posts — and a draft whose fact came from a bound post may carry that post's link. The queue card shows exactly which post a draft is anchored on, so you always know what a reply is standing on and what it links to before you send. Links stay rare by design: the budget caps them at roughly one in twenty drafts per Hunt, and a link never appears on a reply that isn't anchored on that same post.

It learns your judgment, not just your voice

Edit a draft before sending and the engine keeps your version next to its version, feeding the corpus that trains your voice profile. Skip a draft and tap why — the reason becomes a standing rule applied to every future draft. Every learned rule is visible on the queue page under Learned Reply Rules, and one click deletes any rule you stop agreeing with.

The streak

Progress compounds when the loop runs daily, so the queue tracks it:

• Progress shows sent count against your daily goal, on the web queue and in the extension.
• Done for today renders the moment you hit the goal.
• Streak increments each completed day and resets on a miss. Your best streak is kept.

Getting started

Install the Chrome extension, create at least one active Hunt with search keywords (at Hunt Missions, or ask your AI assistant to create_hunt), open x.com, and run it. No Hunts, no targets.

Reading the queue from chat

Any connected MCP client can read the queue with get_reply_queue: the staged items in work order, drafts included, plus progress and streak state. Useful for a morning "what's in my queue today?" without opening the app. Filter it to source: "agent", a specific source_agent, mode, or status to see just what an agent queued. The tool itself is read-only. Sending always happens on x.com.

Agents can queue replies for you to send

An AI agent (your own assistant, a bot you run, anything with MCP access) can put a reply suggestion directly in front of you with queue_reply, instead of only reading the queue. Each item carries:

• mode — full (a finished reply, ready to paste) or angles (2-3 one-line angles; you write the reply yourself, same as the engine's own angles-first targets above)
• why — the agent's one or two sentence reason the reply is worth sending
• source_agent — which agent queued it (e.g. "claude", "grok-bot"), shown as a badge everywhere the item appears

Queuing the same target twice does nothing — the tool returns the existing item instead of creating a duplicate, and an already-answered tweet is reported back rather than re-queued. Items expire automatically (48 hours by default) if nobody acts on them.

Agent-queued items show up in their own Agent Queue section on this page, separate from the engine's own staged targets, and in a slim queue panel in the Chrome extension on x.com with the same per-mode actions. Exactly like every other reply on this page: nothing is ever sent by the API or the extension. You always click Reply. Marking an item posted — whether you paste it verbatim or edit it first — is captured the same way a regular send is, and counts toward your streak.

An agent that already has an item queued can adjust it with patch_reply_queue_item (edit the text or angles, change priority, or withdraw it) as long as it's still waiting on you — it can only ever touch its own queued items, never the engine's.

Related

• Ranked Reply Targeting — the Hunt definitions the engine runs from
• Fact-Anchored Replies — why drafts lead with a concrete fact or don't exist at all
• War-Story Notes — writing the facts your replies can quote