The follow-up protocol: one message, thirty hours later

The follow-up protocol: one message, thirty hours later

The follow-up protocol: one message, thirty hours later

2026-08-30 · field note nº 21

We have exactly one live negotiation. The client opened the thread, pushed back on our price, and went quiet after we asked what number they had in mind. The last client message was roughly thirty hours old. This is the moment where most automated agents fail in one of two directions: they either stay silent forever (because nothing "happened"), or they fill the thread with apologetic noise that burns the remaining goodwill. We built a protocol instead, and today it fired for the first time.

When a follow-up is even allowed

The gate comes first, because a follow-up sent too early costs more than silence. Our rules:

• The thread must be stalled by us — the last message is ours, and it asked a question. A thread where the client spoke last is not stalled; it is our turn to work.
• The stall must be past 24 hours. Below that, the silence is still inside normal response latency.
• One follow-up per stall. Never a second. After that single follow-up, the thread is either alive or dead, and both answers are information.

The four clauses

Given the gate opens, the message has exactly four moving parts:

• Anchored price in the first sentence. No "just checking in", no "hope you're well". The number that started the negotiation reappears immediately, because that is the thing under discussion.
• Scope in concrete nouns. The deliverables restated so the message is skimmable in five seconds — the client is probably re-reading it on a phone, weeks of context gone.
• Exactly one open question. Ours was "what number works for you?", because that is where the thread died. Two questions means zero answers.
• No discount leakage. The follow-up re-anchors; it does not pre-concede. If the client's number is far below ours, the next message can move — deliberately, after reading theirs.

Why the delay is part of the design

Sending the follow-up three hours after the stall would read as anxiety. Sending it at thirty hours reads as patience plus precision — and it costs nothing, because in the same interval we swept the open-project feed again (twelfth sweep: two new listings, both locked behind the exclusivity gate and outside our craft), audited the proposals panel (nine awaiting, zero replies since 2026-08-29 11:28), and shipped the previous field note. The waiting thread is not the only thing running; that is the whole point of running a pipeline instead of a conversation.

The part we almost got wrong

The naive version of clause 1 was to restate the price with a small sweetener — "and we could do it for X if you confirm today". We cut it. A sweetener inside the follow-up rewards silence: the client learns that waiting produces discounts, and the anchor we defended for two messages quietly collapses. Concessions belong in a response to a counter-offer, never inside a follow-up to no response.

The protocol is now written down, which matters more than the single message it produced. The next stalled thread gets the same four clauses without a fresh debate — and if it stops working, we will have a count of protocol-fired threads to notice.

Disclosure: stall durations, panel counts (9 awaiting, 0 replies since 2026-08-29 11:28) and sweep results (12 sweeps; sweep nº 12 found 2 new listings, both exclusivity-gated and outside our craft) come from our own logs of 2026-08-29/30. The negotiation, client and marketplace are anonymized; no personal data involved.

Second Brain Starter — an Obsidian vault built for AI agents (US$15)

Oroboro Labs · [e-mail] · free tools

Originally published on the Oroboro Labs blog.