How we use Claude to run sales operations: an agent that drafts, a person who sends

Claude runs the back office of my sales company. It reads the inbox, decides what each thing is, writes the draft, logs one row, and waits. I read the row and click send. That is the whole relationship.

It replaced about 30 chat sessions that contradicted each other. Now one session runs everything, with an hourly tick and a morning brief. The agents are files in a folder, one per job, each with a lane and a model.

This page is how it is wired, what each part costs, and the rules that hold it together. The prompts themselves are free at the bottom.

How it is wired, in six steps

1. 01

   Start with two rules, and end every prompt the same way

   Drafts only, never send. Every output line must be true. Those two lines sit at the top of all ten prompts, and every prompt ends with the same sentence: you still need real people to make the calls.

   Money, contracts, pricing, pausing a client, complaints, anything you cannot undo: the agent logs one line and waits for me. Everything else it does, logs, and reports.

   screenshot: claude-rules

   Real screenshot goes herePrompt 1, the loop, open in a terminal with the two rules at the top

2. 02

   Give every job its own prompt and its own clock

   The loop runs every 10 minutes: read what arrived, decide the intent, act, log one row. Owed reply runs every minute and finds any lead where their newest message is newer than our newest reply. The show-rate desk runs every 5 minutes. The post-call note runs after every calling block. The weekly report on Friday, the revenue check Monday at 8, the learn pass Sunday night.

   One job, one prompt, one schedule. When a job breaks, I know which file to open.

   screenshot: claude-schedule

   Real screenshot goes hereThe schedule table: ten jobs, their cadence, and the model each one runs on

3. 03

   Route the model to the job

   Detection is cheap. Decision is not. Plain Python or the smallest model finds candidates all day and never writes copy. A second model decides whether a draft should exist at all, and it is advisory only. A mid-size model writes the draft. The biggest model judges pricing and deal shape, and nothing else.

   The rule I wrote on the wall: the big model writing a status field is the waste. The small model deciding whether to email someone who said no is the risk.

   screenshot: claude-agents

   Real screenshot goes hereThe agents folder: eight files, each with a name, a lane and a model line

4. 04

   Put every rule in a gate, not a doc

   Every draft passes a voice check before it lands in Close: no dead dates, no naked links, no banned phrases, the right length, an honest gap since the last touch, a near-future ask. A second check knows that a follow-up, a nurture, a check-in and a request are four different shapes.

   The test for whether a rule belongs in code is one question: could I forget it? If yes, it goes in a gate. In early August every rule that lived in a doc got broken inside two days. Every rule I had moved into a gate held.

   screenshot: claude-gates

   Real screenshot goes herevoice_check output on a draft: two flags, one fix, then pass

5. 05

   Qualify before you pay for data

   Claude plus Apify checks each candidate against the buyer profile before a single credit is spent on a reveal. Company and person, cross-referenced against what is public. The measured result on one client pass: 788 researched, 237 passed, 30 percent. The list built without the gate ran about 80 percent wrong.

   The gate is not a nice-to-have. It is where most of the money is saved.

   screenshot: claude-qualify

   Real screenshot goes hereThe qualify verdicts: QUALIFIED, BORDERLINE, DQ with the reason column

6. 06

   Write in the owner's voice, and measure the voice first

   I ran 700 of my own texts through a script. Average 6.8 words. 97 percent start lowercase. Zero "let me know". So texts are short and lowercase, and emails use normal casing, because I asked for that in writing.

   Every draft is 55 to 65 words in three beats: acknowledge the context, one thing of value, one soft next step. One link, opened first, never the homepage. The subject line is two first names. Names come from the contact record, never from the transcript.

   screenshot: claude-voice

   Real screenshot goes hereA draft next to the measured voice stats it was checked against

Run this today

1. Pick one job. The loop is the right first one.
2. Paste prompt 1 next to your CRM. Swap the CRM name and your name.
3. Add the two rules at the top and the closing line at the bottom. Do not remove them.
4. Run it by hand three times. Read every row. Fix what is wrong in the prompt, not the output.
5. Put it on a clock. Ten minutes for the loop, one minute for owed replies.
6. Write one gate for the rule you broke most this month.
7. On Sunday, read what you edited before sending. Write the rule down. That is prompt 10.

ceo                opus     sets the day, routes work; a wrong call cascades
revenue-director   opus     pipeline, pricing, proposals; money is hard to unwind
ops                sonnet   hygiene sweeps, job health; checks against explicit rules
internal-sdr       sonnet   Coda's own pipeline: sourcing, follow-ups, reminders
client-sdr         sonnet   delivery inside a client's lane; the playbook is written
client-success     sonnet   client updates, reports, onboarding; a known format
data-ops           haiku    gates: mobile, still-in-role, dedupe; mechanical, checkable
copywriter         opus     every word a prospect reads; a bad send costs a relationship

Shared law: never send, always draft. No task lists. No em dashes. Plain language.
Read memory first. Report what was done, what was drafted, and what needs me.