Claude Skills for Sales: Outbound That Gets Replies

Sales is the highest-stakes category we test. One templated-sounding line kills a thread with a prospect forever, and there is no second first touch. So we don’t grade sales skills on plausibility. We grade them on live sends and real (anonymized) accounts, because the question is never “does this look good,” it’s “would a buyer reply.”

That standard matters more now than it did a year ago. Every SDR team has some version of Claude for sales work by this point, which means the inboxes you’re writing into are already full of AI-shaped text. The only outreach that survives is the kind a buyer can’t distinguish from a rep who did the homework. The skills below cleared that bar in our tests. A few didn’t, and we’ll say so.

If you just want the ranked list, it’s at /best/sales. This post is the longer version: the full outbound loop, one worked example from an actual test batch, the math on why personalization quality beats volume, and the places where you still have to be the human.

The outbound loop, covered end to end

Cold outbound is a chain: define who you’re targeting, research the account, write the first touch, follow up, run the call, keep the CRM honest, send the proposal. A weak link anywhere and the rest doesn’t matter. Here’s the tested skill for each step, with scores from our standard rubric (install, triggering, output quality, docs).

Step 0. ICP definition: no skill, and we mean that. We haven’t found a skill worth installing for ICP work, and we’ve looked. Defining who you sell to is a judgment call built on your closed-won data and your founders’ scar tissue. Write it yourself as a plain context document: segments, company size, trigger events, disqualifiers. Every skill below performs measurably better when that document exists, because output quality tracks input specificity almost linearly. This is the one step where “just use AI” is bad advice.

Step 1. Account research: Account Research — 8.8. Feeds everything downstream. Give it an account name and it compiles company context, probable priorities, trigger events and suggested entry angles into a one-page brief. We ran it against five target accounts and had two working SDRs compare the briefs to their own prep. Both rated the skill’s briefs higher. That’s the bar: an SDR would pay for this output, and these two effectively said they would.

Step 2. First touch: Cold Email — 9.2. The flagship of the category. It forces research-backed personalization, meaning verifiable specifics rather than “I loved your recent post,” and it enforces a short single-CTA structure. We tested it the only honest way: a live 15-send batch to real prospects. Three replies. If you’ve run cold outbound, you know what a 20% reply rate on a first-touch batch means, even at small sample size. We get into the math below.

Step 3. Follow-ups: Email Sequences — 8.0, with a caveat. This one lives in our marketing category, and our live test was a 5-email onboarding sequence shipped to a real list, where open rates came back normal-to-good and no subject line read as AI. We haven’t run a dedicated outbound follow-up test yet, so treat the 8.0 as a score for sequencing logic, not for cold follow-ups specifically. What transfers well: it gives each email in a sequence a distinct job, which is exactly the discipline most SDR follow-ups lack. “Just bumping this” is not a job.

Step 4. Before the call: Call Prep — 8.4. Ten minutes before a meeting it produces the prep a good manager would demand: agenda, stakeholder map, predicted objections with responses, and discovery questions ranked by value. We ran it against three real calls. Its objection predictions were right on two of three. The third call went sideways in a way nobody predicted, which is its own kind of reassurance: the skill isn’t clairvoyant, it’s prepared.

Step 5. CRM hygiene: Pipeline Review — 7.2, works with setup. Flags stale deals, stage inflation and next-step hygiene with uncomfortable accuracy. In our test it read a CRM export and called out inflated stages that a human review had let slide for weeks. The “setup” in the verdict is data access: you either export your pipeline to CSV or connect your CRM via MCP. That friction is worth it. It’s the difference between a forecast and a wish.

Step 6. Close the loop: Proposal Builder — 6.8, works with setup. Turns discovery notes and pricing into a structured proposal document. In our test it assembled a branded DOCX from call notes, but only after we installed the DOCX skill and configured a template. Output quality depends on that dependency, hence the modest score. Skip the setup and you get a wall of text nobody signs.

One more in the pipeline: Objection Handler is in our test queue. Two account executives are running it against their real objection logs. No score until they’re done, because a verdict without a test is just a vibe.

A worked example: from account name to first touch

Abstractions don’t sell anyone on a workflow, so here’s one prospect from our test batch, anonymized the way we anonymize everything.

The input. A single line: the account name and the contact’s role. Call the company Northwind and the contact Dana, a revenue operations lead. That’s all the Cold Email skill got, plus the account brief from step 1.

The research output. The Account Research brief came back one page: what the company does, its probable priorities for the quarter, and, most usefully, trigger events. The one that mattered here was hiring. The company’s careers page had added three RevOps roles inside a month. The brief flagged that as an entry angle, with the reasoning attached: a RevOps hiring spike usually means the current CRM process is buckling under growth. This is the kind of inference a good SDR makes on a good day and skips on a busy one.

The email the skill produced. Four sentences:

Hi Dana — Northwind’s careers page added three RevOps roles last month; that’s usually when CRM hygiene starts hurting. We cut manual entry 34% for Loop. Worth 15 minutes Thursday?

Notice what’s in there. A verifiable, dated observation about the prospect’s own company. A stated inference the reader can agree or argue with. One proof point with a number. One specific ask. And notice what’s absent: no “I hope this finds you well,” no paragraph about the sender, no compliment, no “quick question” subject-line bait. Across the 15-send batch, this pattern held. Every email referenced something checkable, and none of them opened with filler. Three of the fifteen got replies.

Could a human have written that email? Obviously. A human who does 15 minutes of careers-page archaeology per prospect. The point of the skill isn’t that the output is superhuman. It’s that the output is good-human, at a per-email cost of a few minutes of review instead of twenty minutes of work.

The reply-rate math

Here’s why personalization quality is the whole game, with numbers you can check against your own funnel.

Take a standard SDR sending 200 cold emails a week. On templated spray, a 1% reply rate is a normal, unembarrassing result in 2026. That’s 2 replies a week. If a quarter of replies convert to booked meetings, that SDR books 2 meetings a month from cold email. Whole outbound programs run on arithmetic this thin.

Now suppose better research and a better first touch double the effective reply rate to 2%, which is a conservative read given our small-batch results ran far above that. Same 200 sends: 4 replies a week, 4 meetings a month. Double the quality signal again, to a 4% reply rate, and you’re at 8 replies a week and roughly 8 meetings a month, from the same list, the same sender, the same volume. Nothing else in the stack moved. Only the words did.

The skeptical objection writes itself: our 3-of-15 result is a sample of fifteen. Correct, and we said so in the test notes. We are not promising you 20%. What we’re claiming is narrower and more defensible: the binding constraint on cold email reply rates is personalization quality, personalization quality was previously constrained by SDR time, and these skills remove most of that time constraint. Hand-researching 200 prospects at 15 minutes each is 50 hours, which is why nobody does it. Reviewing 200 skill-drafted emails at 3 minutes each is 10 hours, which fits inside a workweek with room left for calls.

The version of this that goes wrong is reading the math as a volume argument. It isn’t. If the time savings tempt you to send 600 instead of 200, you’ll trip the deliverability tripwires we describe below and your effective rate goes to zero. The math only works in one direction: same volume, better emails.

What the scores don’t say

The pattern across every sales skill we tested: output quality tracks input specificity almost linearly. The Cold Email skill with a rich account brief produces something a human SDR would send. With just a company name, it produces polite noise, grammatically flawless and instantly deletable. This is why the loop order above matters. Research feeds writing, writing feeds calls, calls feed the CRM. Teams that install only the email skill and skip the research step are buying a very articulate template engine, and templates are the thing we’re all trying to escape.

Honest limits

We sell skill reviews, not miracles, so here’s where this stack stops working.

Deliverability is not a writing problem. No skill fixes a burned domain, missing SPF/DKIM/DMARC records, or a sending pattern that looks like a bot woke up at 9:00 and fired 400 emails in one hour. Worse, the major filters now profile AI-pattern text at scale: uniform sentence lengths, the same twelve transition phrases, identical structure across a batch. The Cold Email skill’s output varied enough to pass in our test, but if you crank volume, you re-create the pattern the filters are trained on. Warm your domain, throttle your sends, and treat every skill-drafted email as a draft.

Data freshness is your job. The Account Research skill reasons well over the information it can reach, and it flagged trigger events accurately in our five-account test. But it cannot know about the reorg announced in a private all-hands, the champion who quit last Tuesday, or the funding round that closed but hasn’t been reported. Every fact in an outbound email needs a ten-second human sanity check, because one hallucinated or stale “fact” about a prospect’s company costs more than all the time the skill saved you that week.

Some emails a human must write. Reply handling, anything involving pricing negotiation, breakup emails to a deal you actually care about, and any thread where the prospect has shown real emotion, positive or negative. The moment a conversation becomes a relationship, drafting assistance should drop to zero or near it. Buyers forgive a mediocre first touch. They do not forgive discovering that the “thoughtful reply” they engaged with was unreviewed machine output.

Two practices that burn teams. Sending without human review, and volume worship. We keep repeating these because every horror story we’ve collected from readers reduces to one of the two.

A note on GDPR and EU prospecting

If you prospect into the EU, cold B2B email is generally defensible under legitimate interest (GDPR Article 6(1)(f)), but that defense has conditions: the outreach must be relevant to the person’s professional role, you need a documented legitimate-interest assessment, every email needs a working opt-out that you honor immediately, and you must be able to tell a prospect where you got their data if they ask. AI enters the picture through data minimization: feeding a prospect’s scraped personal history into a model to generate “hyper-personalization” is exactly the kind of processing that regulators have started asking pointed questions about. Our advice, which is operational rather than legal (talk to a lawyer for that): personalize from company-level, publicly stated facts, like the careers-page signal in our worked example, rather than from an individual’s personal data trail. It’s also, conveniently, what performed best in our tests. Note that several EU member states layer stricter e-privacy rules on top of GDPR, so check the countries you actually target.

FAQ

Can Claude send the emails for me? Not with these skills, and by design. They draft; you review and send through your existing sequencer or mail client. Some teams wire up sending via MCP integrations, but we recommend keeping a human between draft and send. Every result in this post assumes that review step exists.

Which skill should I install first? Account Research, then Cold Email, in that order, because the second is only as good as what the first feeds it. Installation takes about two minutes per skill; the walkthrough is at how to install Claude skills. If you want all five pre-configured to chain together, that’s the Sales Pro Pack.

Does this work with my CRM? Pipeline Review needs your pipeline data, either as a CSV export (works with anything that can export) or a direct CRM connection via MCP. Salesforce and HubSpot MCP connections are the common paths. The other skills in the loop don’t touch your CRM at all.

Won’t AI-written emails hurt my deliverability? Volume and infrastructure hurt deliverability; well-reviewed AI drafts at human volume don’t, as far as any evidence we’ve seen shows. The risk appears when you scale sends because drafting got cheap, and identical AI patterns start repeating across hundreds of messages. Keep volume flat, vary your sends, and run drafts through Humanizer if you want an extra pass against AI tells. It scored 9.6 in our blind-editor test.

Is it legal to use AI for prospecting EU leads? Generally yes for B2B under legitimate interest, with the conditions in the GDPR section above: professional relevance, documented assessment, honored opt-outs, and restraint about feeding personal data into the model. B2C cold email is a different and much harsher regime. When in doubt, ask a privacy lawyer, not a skills directory.

Where to go from here

Every skill above is free to install; the Sales & Outreach category has install commands and the full test notes, and the ranked list lives at /best/sales. If our test batches convince you of anything, let it be this: the teams winning with Claude for sales aren’t the ones sending more. They’re the ones who made 200 good emails cost what 200 bad ones used to.