Reply Manage

Reply Manage watches an X account's mentions and drafts a reply for each new one, using the same brand-voice grounding as regular post generation, plus a knowledge base and a set of guardrails that decide whether a draft is safe to act on at all.

Turning it on

Each connected X account gets its own reply config. From the Reply Manage page, press Enable replies on an account's row to create one. Two things are configurable from there:

• Active: whether mentions are being scraped and drafted for this account at all.
• Require approval: if on, drafts sit in the queue below for you to approve, dismiss, or skip. If off, a reply is sent as soon as it's drafted and passes the guardrails below, with no human step in between.

There's no default here that's "safer" than the other in Pegacraft's eyes: both are real, first-class modes, and you can flip the toggle at any time without losing the config.

Knowledge base

A collapsible Knowledge base panel on the Reply Manage page holds documents scoped to your project: facts, FAQs, and policies the model can cite when drafting a reply. Add a title and a body, separating distinct facts with a blank line; each becomes its own retrievable chunk. Retrieval is phrase and keyword similarity against what a mention actually says, not full semantic search, so a document reads best when it's written the way a person would actually ask about it, not as marketing copy.

A document can be disabled without deleting it: its chunks stop being retrieved, but the text stays in place to re-enable later.

Guardrails

Two settings on each account's reply config decide what a mention has to clear before a reply goes anywhere:

• Confidence gate: after drafting, the model scores its own reply from 0 to 1. Below the configured threshold (0.8 by default), the mention is escalated instead of sent or queued, even when the account is in auto-send mode. A confident wrong answer is worse than an honest "I'm not sure."
• Red-line topics: before anything is drafted, the mention is checked against a list of topics (financial advice, legal advice, and press inquiries by default, editable per account). A match escalates immediately with no draft attached at all: Pegacraft never even attempts to answer a red-line question.

Both checks fail closed. If the classifier's own output can't be parsed, that's treated as a hit, not a pass. An ambiguous read on a sensitive path escalates rather than guesses.

The queue and escalations

Drafts waiting on approval, and mentions that hit a guardrail, both show up in the same keyboard-driven, one-at-a-time queue, escalations first.

An escalation carries a reason: a named red-line topic, "low confidence," or "classification failed" if the guardrail check itself broke. A red-line escalation has no AI draft to fall back on, so the card shows a plain text box instead: write the reply yourself, and it goes out bypassing the guardrail entirely, since you're the one vouching for it now. A low-confidence escalation keeps its draft visible, so "Send anyway" is a real option once you've read it.

Reply Manage's escalations also surface in the Inbox's Escalated tab, so you don't need to remember to check two separate places.

How often mentions are checked

Mentions are scraped on a recurring schedule, roughly every 10 minutes, jittered per account so every connected account isn't hitting X at the exact same second. A given mention is only ever drafted once: if the same mention shows up on a later scrape, Pegacraft recognizes it and doesn't draft (or charge for) a second reply.