The problem
A seller’s reply obligation scales with the size of the account book. The revenue does not. Past a few hundred accounts, most inbound is simultaneously low-stakes and non-optional — a rate-limit question, a link to the right doc, a scoping clarification, a nudge on a renewal date. None of it is hard. All of it has to be answered by someone who knows what was actually said on the last call.
That work has no natural slot, because calls take the day. It lands in the hours that are left, which are the evening ones. The visible symptom is a seller answering mail at 10 p.m.; the invisible one is that reply latency becomes the thing customers judge you on.
The two obvious fixes both fail, and they fail for the same reason.
Add a process step — snippets, macros, a triage SLA. A template covers the greeting and the sign-off, which were never the expensive part. It stops where the work starts: the account-specific middle that requires this customer’s plan, this thread’s history, and this week’s version of the docs.
Buy a tool that sends for you. This does remove the composition cost, but relocates the risk to the one place you cannot inspect it — after the message reached the customer. Sellers who understand their own liability work that out fast and quietly stop using it. Adoption dies; the evening queue comes back.
The need is a way to pay the composition cost without taking the send risk.
How it works
Anthropic’s own account, published on its Claude blog in June 2026, describes what one seller built. The article states that Jared Sires joined Anthropic in 2024 as a startup account executive and had never written a line of code; that his book grew to 600 or 700 accounts; that he was running 10 to 15 customer calls a day; and that he was answering customer email until 9 or 10 p.m. every night.
His first build was not the drafting tool. Using Apps Script and Claude, he pulled product usage data from internal systems and had Claude rank his accounts each morning by growth rate.
The practice this page is about is the next one. CLAFTS — short for Claude Drafts — is described as an application that lives inside Gmail and uses the Claude API to draft replies to customer email. Per the article, it pulls context from a shared Google Drive folder and other third-party tools, references Anthropic’s public documentation through web search, and matches his writing style.
The output does not send. It accumulates where unsent mail already lives:
By the time he opens his drafts folder at the end of the day, the responses are waiting for review.
The article says Claude’s out-of-the-box writing ran longer and hedged more, and that Sires reworked the system prompt until drafts matched his own style. On the volume of that tuning, in his words:
I've probably gone through hundreds of iterations with CLAFTS in the system prompt to generate different pieces of writing for me
He then added a feature the article calls CLAFTS Tones, which it describes as using pattern matching to mimic his voice across different relationships — customer, peer and family threads read differently, and the drafts adjust. He tested it by writing himself a sequence of increasingly angry emails on his personal account:
Claude started to mimic that, and at some point I started to have refusals because Claude didn't want to generate angry emails to customers.
On scale and spread, the article reports: the app runs to roughly 4,300 lines, almost all written by Claude Code; a colleague, John Albert, helped co-build it; and within 24 hours of him sharing it in Slack, others in the sales organisation had started using it.
Separately, the article describes a Sales plugin he ships through Claude Cowork, carrying 20-plus skills wired to Salesforce, Intercom, Gong, Google Calendar, Gmail, Google Drive and BigQuery, and anchored by two: /customer-context, which it says assembles a 360-degree account view in about 90 seconds, and /pipeline-management, which surfaces at-risk deals, forecasting guidance and progression recommendations. The source does not say the plugin is CLAFTS repackaged. Within months of launch, the article says roughly 80 percent of Anthropic’s sales org was using it.
On time saved, the article reports the same estimate at two granularities: two to three hours a day, and 10–15 hours per week. Both are his own figure, and neither is audited. What he says changed:
Before CLAFTS, I felt like I was doing more administrative work than actually spending time with customers. After CLAFTS, I was actually able to do more of what I wanted to do, which is sales.
One more reported detail, because it bears on who can copy this:
Claude Code, having the terminology 'code' at the end of it, made me feel a little bit intimidated just to even start
The source does not specify how much of a typical draft he edits, what share he discards, whether any quality bar or accuracy check exists, or whether anyone other than the sender reviews. Those absences matter, and we return to them below.
Why it works — our read
Everything from here is RevGuild analysis, not the practitioner’s claim.
The input this changes is narrow and precise: the seller no longer opens a blank message. Composing from nothing and judging something already written are different cognitive tasks with different costs, and only one of them uses what a seller is actually paid for. Deciding is this the right commitment, is this promise safe to make, does this match what we said on the call is the job. Typing “Hi Sam, thanks for flagging this” is not. Pre-drafting deletes the second and leaves the first.
That only holds if the draft sounds like the person sending it. Voice-matching is not a nicety here, it is the load-bearing constraint: a draft in the wrong register has to be rewritten, and rewriting costs what writing cost, so the whole economic case collapses to zero. This is why the honest cost line is the prompt-tuning, not the build. Any version of this that skips it produces drafts that are cheaper to delete than to fix.
And it survives contact with a real week because of where the output lands. The drafts folder is a review surface that already exists: no new dashboard, no new tab, no new habit to maintain, and — the part that matters — a safe failure mode. Forget to review, and nothing happens; the mail sits there. Under auto-send, forgetting is unrecoverable, because the error is already in the customer’s inbox.
The transferable principle: put the AI upstream of an approval surface your team already uses, never downstream of one you have to invent. The counterfactual is the same drafting quality behind a new internal review queue — which nobody opens by week three, and which then gets “temporarily” bypassed.
Where it breaks
Ours, not the source’s.
- Review fatigue is intrinsic, and improvement accelerates it. A queue of consistently plausible drafts trains the reviewer toward approval as the default. The better the model gets, the faster that happens. The practice then degrades into auto-send without anyone ever deciding to auto-send. Earliest visible symptom: your edits get shorter week over week while your send volume rises.
- The reported gains are self-reported and unaudited. No edit rate, no discard rate and no quality bar appears in the source. Treat the direction as evidence and the magnitude as not yet measured.
- A personal voice model does not transfer the way a shared skill does. Hundreds of prompt iterations bought one person’s register. The next rep starts near zero. Note the tension inside the same article: the thing that reached roughly 80 percent adoption was the shared plugin of skills, which is a different artefact from a tuned personal voice.
- Approval is still authorship, legally and commercially. A rubber-stamped delivery date, discount or security assurance is a commitment your company made. The source does not raise this.
Don’t do this if the bulk of your inbound is commitments rather than information — pricing exceptions, security questionnaires, contractual language. There the review cost approaches the writing cost, the saving disappears, and the only thing you have added is a fluent, confident first draft of a promise nobody approved.
What has to be true at your company
Volume. Enough repetitive inbound that composition is a genuine block — the cited case sits at several hundred accounts and a double-digit daily call load. At three replies a day, tuning and review overhead cost more than they return.
A fetchable source of truth. The drafting step is only as good as what it can read: current docs, a shared folder, account records with recent state. If the authoritative answer lives in a senior rep’s head, you get fluent, wrong drafts — more expensive to review than a blank page.
A usable voice corpus. Your own sent mail, in volume, readable by the tool. Without it there is nothing to match against.
Permission to build in the seller’s own tooling. Something has to be allowed to read and write mail on a work account. That is an IT and security sign-off — the step most likely to kill this at a regulated company.
By company shape: at eight people, skip the app — the practice is a prompt and a habit, and the founder’s sent folder is the corpus. At 800 the corpus problem inverts: docs exist but are stale in three places at once, so the first project is deciding which source wins. PLG inbound is high-volume and mostly informational — the best case. In enterprise the commitments-to-information ratio is far worse, and the honest scope is internal follow-ups and recaps, not customer replies.
Try it this week
One person, about an hour, no approval needed, nothing installed.
Take your last 20 sent replies and paste them into a project or system prompt as a voice sample, along with your current pricing and docs links. Take the five oldest unanswered inbound threads. Generate a draft for each. Send nothing without reading it.
Then measure the edit, not the time. For each draft, log one of three outcomes: sent as-is, light edit, rewritten from scratch. Do that daily for a week, tuning the prompt when a rewrite happens.
If it is working, the share landing at “light edit or better” climbs across the week and passes a third by day five.
Conclude it isn’t when “rewritten from scratch” is still over half after two rounds of prompt tuning. That result is informative rather than disappointing: it means your bottleneck was never composition. You did not know the answer yet, and no amount of drafting fixes that.