Turn replies into meetings
Use Inbox context and CRM records to draft follow-up work.
Use this ongoing play when an outbound motion has produced inbound messages that need an answer. It helps a founder or operator work from the existing conversation, classify what the person meant, and draft a useful next step without restarting the original pitch. Read Inbox for the conversation model and CRM records for the durable facts that belong with a person, company, or deal.
The play uses the inbox-replies skill. That skill teaches Architect to treat a reply as part of the same thread, to inspect the full exchange before writing, and to read the thread again after sending. A suitable outcome is a meeting request or another concrete next step, but an unsubscribe, decline, referral, or out-of-office reply has a different correct outcome.
Needs and produces
Start with a bounded set of unanswered threads, meaning threads whose last message is inbound, for one channel and a defined time window. You also need the conversation's channel and sender account, the person or worksheet row when it exists, and the Context pages needed to support a factual answer. The skill names Product Information as the source for business facts and Tone of Voice as the source for register.
The output is a per-thread decision, not a bulk message. It includes the person, channel, classification, the quoted line that supports that classification, the sequence stop state, a reply draft, and the Context pages used for its claims. When a message is sent, the final result also confirms that it appears in the original conversation and gives the conversation link when available.
This play does not create a new outreach sequence for a person who replied. A new conversation with a person who has never been contacted is a sequence, even when it contains only one message. Existing threads stay in Inbox, while lasting relationship facts may be filed in CRM after the conversation has been understood.
Read the conversation before drafting
In the Inbox screen, open each selected thread and read every message in both directions. Architect uses the inbox-replies skill with Inbox reads to locate the active conversation and the sequence state. It then reads the person's worksheet row and only the Context pages the answer depends on. This keeps a question about a feature from being answered with a general sales summary or a guessed date.
Classify the latest inbound message from the words and conversation context, not from a keyword. The skill distinguishes interested, question, objection, not now, wrong person, referral, unsubscribe, and out of office. If the thread supports more than one interpretation, keep that uncertainty visible instead of choosing a category merely to move the thread forward.
Before drafting, check whether the lead's sequence run is stopped. A reply ends that person's run on every channel. If the run is still active, stop it before preparing the response. This check prevents a reply from being followed by another step from the earlier cadence while a person is trying to have a direct conversation.
Draft for the actual reply
Architect applies the inbox-replies skill to the classification, then produces a short reply in the existing thread. For interest, the skill calls for one short paragraph, a concrete next step, and a specific way to book. For a factual question, it answers from Context, cites the page used, and makes one ask. An objection is acknowledged in the person's words, followed by one sourced proof point and one low-effort ask.
A "not now" answer should agree and ask permission to follow up at a stated time. A wrong-person or referral reply should thank the person and ask for, or confirm, the right name. An unsubscribe or decline gets no pitch, only confirmation of suppression. An out-of-office response is evidence about availability, not interest. These choices make the draft match the conversation rather than applying the same meeting request to every thread.
The screen check is specific: before an external action, verify the exact channel, sender account, recipient, conversation, and final copy. Use the visible thread as the source of truth for salutation and context. Do not promise a discount, feature, or date that the relevant Context page does not support. When Context does not contain the needed fact, ask the user instead of filling the gap with a plausible claim.
Work a bounded slice and preserve the record
- 1
Select unanswered threads
Ask Architect to use theinbox-repliesskill for one channel and time window. In Inbox, confirm that each thread's last message is inbound before treating it as work for this play. - 2
Read evidence and stop state
Architect reads the complete thread, the related worksheet row, relevant Context pages, and sequence state. Check the quoted classification evidence and that the reply has ended the lead's run before moving to a draft. - 3
Draft by classification
Architect produces the thread-specific response using the inbox-replies practice. In Inbox, compare the copy with the actual message and confirm the recipient, channel, and sender account shown for the conversation. - 4
Capture lasting facts carefully
Use CRM only for durable record truth such as a new objection pattern or a competitor named in the thread. Keep the conversation itself in Inbox and link the record work to the evidence rather than treating a draft as a confirmed fact. - 5
Read back the sent thread
After an external reply, Architect reads the conversation again. Check that the new message appears in the same thread, then report the confirmed state and link instead of assuming a send response proves delivery.
Credits and external effects
Inbox reads, classification, and draft preparation are different from the external effect of replying. Search for the Inbox capability needed for the current task, read its delivered schema when necessary, and inspect any estimate before a paid live action. A priced live call returns a quote, then the same call runs under the supplied maximum-credit ceiling. Above the server's row or provider-spend limit the server answers approval_required and the user must say yes first.
Keep the slice small enough that every draft receives a conversation-level review. A bulk archive or tag change must state its count first. The finished conversation readback is what establishes the outcome. It also prevents a second send when a client returned an unclear result or the thread took time to refresh.
When it goes wrong
If a thread's classification is ambiguous, do not force it into interested or objection. Return the quoted line, explain the uncertainty, and ask the user for the intended interpretation. If a question needs a product, pricing, feature, or timing claim that Context does not support, pause the draft and ask for the fact. That is safer than making an unsupported promise in a conversation that may become a record.
If the sequence is not stopped after an inbound reply, correct the stop state before drafting or sending. Never create a new enrollment to answer the reply, and never reach the same person through another sender or channel after that stop. If a prior send result is unclear, a conversation status alone does not prove delivery. Read the thread and look for the message before retrying, so the person does not receive it twice.
If the thread belongs to the wrong sender account, recipient, or channel, do not move it to a different conversation just to continue a draft. Read the correct thread and repeat the bounded-slice check. For a failing workflow or sequence state that cannot be explained from the thread, continue with Find out why something failed and report the state already proven.
Leading metric and next play
The leading metric is replies converted into booked meetings. Interpret it alongside the classifications and stop checks, because an unsubscribe or a well-handled "not now" reply is not a failed draft. The immediate proof of this play is a bounded set of threads with their classification evidence, correct stop state, grounded draft, and readback where a reply was sent.
Use Weekly review to compare reply outcomes with runs, sequence analytics, worksheet qualification, and the credit ledger. When the issue is a workbook input rather than a conversation, use Clean up a worksheet before asking a live motion to work around bad source evidence.