prebuilt ai skills vs the ones you write yourself.
a vendor just shipped 37 ready-made sales skills into claude. whether that helps you depends entirely on which of the three layers you are actually missing.
a prebuilt skill is a vendor's workflow, shipped with their product and shaped by their data model. a skill you write is your workflow, shaped by how your business actually runs. take the prebuilt one when the job is generic and the data already lives in that vendor's system. write your own when the judgment is yours.
that split is the whole decision, and last week gave everyone a clean example of it.
what salesforce actually shipped
on august 26, 2026 salesforce and anthropic announced claudeforce, an expanded partnership whose first release is "salesforce in claude": a plugin carrying 37 prebuilt sales skills. the skills let a seller read live crm context, update pipeline, and take governed action without leaving claude.
the count is 37 prebuilt sales skills at launch, with more scheduled for late 2026 (salesforce, august 26, 2026). salesforce publicly names skills for daily briefing, pipeline review, forecast narrative, meeting preparation, deal-health review, win/loss analysis, activity logging, and crm hygiene. availability is select pilot customers now, with open beta expected in september 2026 (salesforce, august 26, 2026).
anthropic ceo dario amodei, quoted in the joint announcement of august 26, 2026, framed it as bringing frontier intelligence into systems where much commercial activity happens
.
read that literally and it is the correct read. the value is not that claude got smarter. it is that claude got access, plus a set of workflows written against a known data model. those are two different things, and telling them apart is the practical skill.
the three layers, side by side
every working ai setup has three layers: access, instructions, and judgment. a vendor can ship you the first two. the third one is always yours. here is what each layer is and who has to build it.
| the layer | what it is | who builds it | when it wins |
|---|---|---|---|
| prebuilt vendor skill | a workflow the vendor wrote against their own product, installed as a plugin | the vendor | the job is identical at every company running that software, and the data is already in it |
| connector or mcp server | the pipe: what data the model can read and which tools it can call | the vendor, or you, or an open server | the model gives good generic answers and just cannot see your numbers |
| skill you write | a plain-text folder holding your process, your rules, your standard for good | you, once | the judgment is specific to your business and no vendor can guess it |
if you want the fuller taxonomy of these pieces, i mapped prompts, skills, plugins and mcps against each other in prompts vs skills vs plugins vs mcps, and what a skill file actually is in claude skills, explained in plain english.
which layer you are actually missing
most stalled ai setups are missing exactly one layer, and people usually guess wrong about which. three questions sort it in about a minute.
- can it see your real data? if the answer is no, you have an access problem, not a prompting problem. that is a connector or an mcp server, and there are plenty of free ones. i mapped nineteen of them by deal stage in mcp servers for sales teams.
- does it do the job the generic way? if it can see everything and still produces the average version, you have an instruction problem. that is a skill, and it is the layer a vendor can hand you if their skill happens to match your job.
- does it get the call wrong in a way only you would catch? then no prebuilt anything will fix it. your qualification bar, your pricing logic, what a good lead means to you: that is judgment, and it only exists in your head until you write it down.
the ordering matters. people buy at layer two when they are broken at layer one, then conclude ai does not work for their business.
the same week, anthropic standardised the other end
on august 27, 2026 anthropic previewed the model hardware standard, a shared specification for ai agents to safely operate physical lab and manufacturing equipment. it is not for your business. it is a research preview with labs and manufacturers, and the honest thing to say is that nothing in it lands on your desk this year.
it is still worth thirty seconds, because of what it says about where the effort is going. anthropic reports that carnegie mellon completed a full setup in 8 hours, against the several weeks such integrations normally take (anthropic, august 27, 2026). that is a lab number, not a small-business number. but the pattern behind it is the same one salesforce is chasing: nobody is racing to make the model cleverer. they are racing to standardise the layer between the model and the thing it needs to touch.
which is the same message as the table above, arriving from the other direction. access and instructions are being commoditised, fast. judgment is not.
what a prebuilt skill cannot know
a prebuilt skill knows the vendor's data model. it does not know your business rules, and that gap is where most ai adoption dies. a "deal-health review" skill can read every field on the opportunity. it cannot know that deals from one particular referral source close at twice the rate and deserve a different follow-up, because that fact lives nowhere in the crm. it lives in you.
this is the shape of the adoption gap generally. roughly 85% of people have access to ai and about 25% use it on real work (IBM 2026 Global CEO Study), and the missing 60 points are not a model quality problem. i wrote the long version in the ai adoption gap. more access, faster, does not close it. writing down the judgment does.
so the useful move when a vendor ships 37 skills is not to install all 37. it is to run them, notice the two or three places the output is confidently generic, and write those down as your own skill. then test it against the no-skill baseline before you rely on it, which is a ten-minute job i walked through in how to test a claude skill before you trust it.
the takeaway
prebuilt skills are good news. they take the generic work off your plate and they are getting better and more numerous, from salesforce and from everyone else. what they do not do is remove the one job that was always yours: deciding what good looks like in your business and writing it down where the machine can read it. take the shelf version for the generic job. write your own for the judgment. that order does not change no matter how many skills ship next.
common questions
what is claudeforce and what does salesforce in claude actually do?
claudeforce is the expanded salesforce and anthropic partnership announced on august 26, 2026. it launches with salesforce in claude, a plugin carrying 37 prebuilt sales skills that let sellers read live crm context, update pipeline and take governed action from inside claude. salesforce lists skills for daily briefing, pipeline review, forecast narrative, meeting preparation, deal-health review, win/loss analysis, activity logging and crm hygiene. it is with select pilot customers now, with open beta expected in september 2026 and more skills in late 2026.
should i use prebuilt ai skills or write my own for my business?
use a prebuilt skill when the job is generic across every company that runs the same software, and the data it needs already lives in that vendor's system. write your own when the judgment is specific to you: your qualification bar, your pricing logic, your voice, your definition of a good lead. a prebuilt skill knows the vendor's data model. it does not know your business rules, and nobody can ship those for you.
do i need salesforce to use claude skills for sales?
no. a claude skill is a plain-text folder of instructions plus any files it needs. salesforce in claude is one vendor packaging its own workflows as skills against its own crm. if you do not run that crm, you get the same shape by writing the skill yourself and pointing it at wherever your data lives, a spreadsheet, a notes folder, or an mcp server that connects your own tools.
what is the difference between a claude skill, a connector, and an mcp server?
a skill is instructions: how the work gets done. a connector or mcp server is access: what data and which tools the model can reach. a skill without access is a well-written prompt that cannot see your numbers. access without a skill is a model that can read everything and still does the job the generic way. most stalled setups have one and not the other.
want the written-down version, already built?
the ai builder toolkit is a set of claude skills for the jobs in this post, written for a business of one or a few, ready to install and run in claude today. no course, no call.
see the toolkitbuilding this out with other people is faster than doing it alone. the ai builders lounge is where the weekly builds happen.
or just follow along. new field notes most weeks on x, instagram, and tiktok.