uygnoey Prospecting a whole book overnight
Derived from a first-person account by John Albert, a business development rep at Anthropic, who works "upwards of a hundred accounts at any given time" and covers them through a skill that runs as a scheduled task overnight.
The output is not a list of accounts. It is, per account, a brief, a score, and an outbound play — waiting when the rep opens their workspace in the morning.
Instructions
1. Define what "the current state of an account" means
The sweep observes the current state of each account. In the post that means questions like:
- Who are we in touch with?
- How do they use the product today?
- What signals are relevant right now?
Write these down before wiring anything up. They determine which sources you need and what the brief has to contain.
2. Connect the sources that answer those questions
The post names the connections behind the sweep: the CRM, sales tools such as Apollo and Common Room, call recordings from Gong, and the team's data warehouse. Claude connects to these, performs deep research, and validates the result.
Mapping of source to question in references/context-sources.md.
3. Validate research against curated guidance
Deep research alone produces plausible accounts of a business. The post adds a validation step: findings are checked against outbound guidance and ICP criteria that the team has curated.
- Curate the guidance as a document the skill reads, the way the inbound side reads a knowledge base.
- Score the account against the ICP criteria explicitly, so the score is reproducible rather than a vibe.
- Let the guidance decide the play, not the research alone.
This is what the post means by adding context "to help Claude work more like a BDR" at the company — the team's own standards are part of the input.
4. Produce a brief, a score, and a play per account
The morning artifact has three parts. Keep them separate:
- Brief — the current state, with the evidence behind it.
- Score — the account against ICP criteria.
- Play — the specific outbound move this state and score imply.
Format in templates/account-brief.md.
5. Keep a memory file and a ledger
The skill keeps a small memory file and ledger, which prevents repetitive or duplicative work. Without it, an overnight sweep re-discovers the same signals and re-proposes outreach that already went out.
- Memory: what the sweep concluded about an account, and when.
- Ledger: what was actually acted on, so the next run does not repeat it.
6. Feed rep feedback back into the skill
The workflow "becomes increasingly useful over time as each BDR can provide feedback on Claude's results, which then feeds back into the skill." Treat a rejected play or a corrected brief as input, not just as a fix for today.
7. Do not build a skill for every ad-hoc request
The post is explicit that one-off requests are often just a prompt: a spend analysis dashboard for a top account, a sweep for accounts already using the product with no matching opportunity, a scored list of webinar invitees from an AE's book. Reach for a skill when the request repeats.
Worked examples in examples/ad-hoc-requests.md.
Examples
The morning open. The rep opens their workspace to a brief, a score, and an outbound play for each account in the book. The research they would have spent hours compiling is already done; the judgment about which plays to run is not.
Undiscovered usage. A prompt considers an AE's full book and finds account-level usage signals where no sales opportunity exists yet — often a good signal to start a conversation about optimizing that usage.
Event outreach without a skill. An AE flags an upcoming webinar and asks which accounts in his book would be interested. Claude checks usage data and CRM history across the book, scores each account against the ICP, and flags the best fits with contacts worth inviting. No skill was built for this.
Avoiding a duplicate touch. The sweep surfaces a signal it already surfaced last week and the ledger shows outreach went out on it. It does not propose the same play twice.