An agent skill is a folder with a text file in it. The text file holds the instructions you would otherwise retype every time you ask a model to do a recurring job, and the model loads it automatically when the job comes up. That is the whole idea. The word "agent" makes it sound like infrastructure, and it is not.
Skills are worth an advisor's attention for one narrow reason: most of the useful things a model does in a practice are things you do again next week. If the quality of that output depends on you remembering to paste four paragraphs of instructions, it will drift, and it will drift on the weeks you are busiest.
We tested that claim rather than assuming it. This post covers what a skill actually is according to Anthropic's own documentation, what changed when we ran an identical prompt with and without one, the complete file, and how sharing works, which depends on your plan and matters to anyone in a firm of more than one person. For the wider picture of what advisors actually use AI for, see our read of the 2026 advisor AI surveys.
unknown nodeWhat an agent skill actually is
Anthropic describes skills as "folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks" (Anthropic Help Center, retrieved 2026-08-18). The required parts are small. A skill needs a file called SKILL.md, and that file needs two things at the top: a name and a description.
The description is the part advisors get wrong, because it works as the trigger rather than as documentation. Claude reads every installed skill's name and description, and uses them to decide whether the skill applies to what you just asked. Anthropic's documentation is explicit that the description must state both what the skill does and when to use it, and caps it at 1,024 characters (Anthropic developer documentation, retrieved 2026-10-16). A description that says "prospect research helper" will not fire reliably. One that lists the phrases people actually type, as ours below does ("Use whenever the user asks for a prospect brief, background on a prospect, pre-meeting prep..."), will.
The loading works in three stages, which is why installing several skills does not fill up the model's working memory. Only the name and description of each skill sit in context at all times, at roughly 100 tokens each. The body of the instructions loads only when the skill is triggered. Anything bundled alongside it, a reference document or a script, costs nothing until the model actually opens it. Anthropic calls this progressive disclosure, and the practical effect is that a skill can be far longer than a prompt you would ever paste by hand.
Skills launched on 16 October 2025 and were published as an open standard at agentskills.io on 18 December 2025, with a reference implementation on GitHub (Anthropic, retrieved 2026-10-16). That matters more than a launch date normally would, because a format other tools can read is a format you are less likely to have to abandon.
The test: the same prompt, with and without a skill
The claim to test is that a skill improves the output. We ran it directly.
The prompt was deliberately lazy, the kind an advisor types at speed between meetings:
I'm a financial advisor. Prepare a one-page prospect brief on MargaretReyes-Halloran, who owns the home at 1428 Willow Creek Drive, Highland Park,Texas. Include her estimated net worth, career background, likely financialpriorities, and three conversation openers for our first meeting.
Margaret Reyes-Halloran does not exist. That is the point of the test: if a model is going to invent a person's finances, it will do it most readily when there is no real person to contradict it.
We ran this twice against Claude Opus 5 on 2026-08-18, in an empty directory with no other context, using the Claude Code command line with web access disabled. The only difference between the two runs was whether a prospect-brief skill was installed.
Both runs refused to fabricate. The control, with no skill, replied that any "estimated net worth: $8–12M" or "career background: retired healthcare executive" it produced "would be invention dressed up as research — and you'd walk into a meeting with fabricated facts about a real person in your head". The skill run reached the same place: "even the homeowner claim is unconfirmed here — it came from your prompt, not from a record I checked."
That result is worth sitting with, because it is not the result the marketing around skills would predict. The skill did not prevent a hallucination, because there was no hallucination to prevent.
What did change was the shape of the answer. The control invented its own format: a sourcing table, a fill-in-the-blanks brief template, and three suggested openers written out in full. Useful, and completely different from what it would produce next time. The skill run returned the four headings the skill specifies, in order, every time: VERIFIED FACTS, INFERENCES, UNKNOWNS TO RESOLVE BEFORE THE MEETING, QUESTIONS TO ASK.
Under VERIFIED FACTS it wrote one word, "None", and then explained why. Under UNKNOWNS it listed seven items, each naming the record that resolves it: the Dallas Central Appraisal District record for ownership and exemption status, the Dallas County Clerk for deeds, the Texas Secretary of State for business roles. It also volunteered a caveat the control did not: Texas is a non-disclosure state, so the assessed value is not a sale price.
So the honest finding is that the skill bought consistency, not safety. For a task you run four times a week, and for anything a reviewer has to look at repeatedly, consistency is the thing you were actually short of. But if you adopt skills expecting a guardrail against invented facts, you have bought the wrong tool. Our glossary entry on hallucination covers what a guardrail actually requires.
With pre-meeting briefs, the likelier failure is rarely a dramatic fabrication. More often the brief quietly reorganises itself each time, so the empty section that should have been the loudest thing on the page ends up third from the bottom, and the reader skims past it.
The complete skill
Here is the file. It is short, and its length is not the point of it.
---name: prospect-briefdescription: Build a pre-meeting brief on a prospect or household. Use whenever the user asks for a prospect brief, background on a prospect, pre-meeting prep, research on a household, or "what should I know before I meet this person".---# Prospect briefA brief exists to make the first meeting better, not to look impressive. It isallowed to be mostly empty. An empty brief that says "I do not know" iscorrect; a full brief that guessed is a liability.## The source ruleState a fact only if you can name where it came from. If you cannot, write:```UNKNOWN — verify via: <the specific record, database or question that answers it>```Never state a net worth, income, job title, family structure or age as factwithout a named source. A wealth estimate belongs under INFERENCES, as arange, labelled as an estimate, with its basis (for example, the assessedvalue it starts from). Never fill a gap because the gap looks bad.If you have no source for anything, say so in one sentence and go straight tothe questions section.## Output format```VERIFIED FACTS- <fact> — source: <named source>INFERENCES (each with its basis and confidence)- <inference> — basis: <what it is inferred from> — confidence: low | medium | highUNKNOWNS TO RESOLVE BEFORE THE MEETING- <unknown> — verify via: <specific record or question>QUESTIONS TO ASK1. <open-ended question that does not presume an answer>2.3.```## What belongs in each section- **Verified facts.** Public record, a document the prospect provided, or something the introducer said. Name which.- **Inferences.** Anything derived rather than observed. A property's assessed value is a verified fact; what it implies about liquid assets is an inference, and usually a low-confidence one.- **Unknowns.** The useful half of the brief. Each one names the record or question that would resolve it.- **Questions.** Open-ended. "How did you come to own the property on Willow Creek?" not "I see you bought Willow Creek in 2019 — was that an investment?" The second one confirms your assumption back to you.## Self-check before answering1. Count the statements in VERIFIED FACTS. Every one must name a source. Move any that do not to INFERENCES or UNKNOWNS.2. Search your own draft for numbers. Every number needs a source or it is an inference.3. If the brief reads as a biography, it is wrong. It should read as a list of what you do and do not know.## Do not- Do not present a wealth figure as fact. An estimate is a range with its basis and confidence, or it is left out.- Do not describe a career, education or family the source material does not contain.- Do not write conversation openers that assert facts about the person.- Do not pad. Length is not the deliverable.
Two lines differ from the file we tested. The tested version banned any wealth estimate outright, including a range. That is stricter than it needs to be: an estimate is legitimate when it is labelled as an estimate, given as a range, and shown with its basis, which is the standard we argue for in why a net worth estimate should be a range. So the source rule and the "Do not" list above now forbid stating a wealth figure as fact without a cited source, rather than forbidding the estimate itself. We have not re-run the test with the revised wording.
Three things in that file are doing the work, and they are worth stealing regardless of which tool you use.
The source rule gives the model a legal way to say nothing. Most prompts do not, which is why they get filled in with plausible material. The explicit permission to be empty removes the pressure that produces invented biography. The self-check asks the model to audit its own draft before returning it, which catches the numbers that drift in during writing.
Those three ideas are the same ones that make a prompt work without a skill. A skill is the same technique, stored somewhere you cannot forget it.
Where skills run, and who controls sharing
This is the part that decides whether skills are usable in your practice, and the answer depends on where you run them and which plan the firm is on.
In Claude Code, skills are files on disk. Personal skills live in ~/.claude/skills/, project skills in .claude/skills/, and they are picked up automatically with no upload step (Claude Code documentation, retrieved 2026-08-18). Because they are plain files, they go in version control, they diff, and you can hand one to a compliance officer as a document rather than a screenshot of a settings panel. For a firm that wants a reviewable record of what the AI was told to do, that property is the entire argument.
On claude.ai, the everyday web and desktop app most advisors will actually use, the rules changed on 18 December 2025, when Anthropic added organisation-wide skill management for Team and Enterprise plans (Anthropic, retrieved 2026-10-16). As Anthropic's Help Center describes it today (Using skills in Claude and Provision and manage skills for your organization, both retrieved 2026-10-16):
- Custom skills are uploaded as a zip file and need code execution turned on. The Help Center lists Free, Pro and Max alongside Team and Enterprise.
- On Free, Pro and Max there is no sharing. Each person uploads their own copy.
- On Team and Enterprise, the organisation's owners can provision a skill for every user. It is switched on for everyone by default. Individual users can turn it off for themselves but cannot delete it. Members can also share a skill they wrote with specific colleagues, or submit it to the organisation's library, and shared copies are view-only.
- Skills still do not sync between the app and the API. A skill uploaded to claude.ai has to be uploaded separately if you also use Claude programmatically (Anthropic developer documentation, retrieved 2026-10-16).
For a solo advisor, none of that is an obstacle. For a five-person RIA, the question is no longer whether everyone can get the same skill but who controls it. On individual Pro accounts it is still a manual process: one person writes the file, everyone else installs it, and nothing tells you when someone's copy has drifted. On a Team plan it becomes an owner's job: the owner uploads the skill once and every advisor gets it. That is better, and it also makes the skill a governance question. Whoever holds the owner role decides what every advisor's AI is told to do, so a compliance reviewer will reasonably ask who that is and how a change to the skill gets reviewed before it reaches everyone. A file in version control answers that question by itself. A settings panel needs a written process.
What a skill is not
It is not a compliance control. A skill is instructions, and instructions are advisory. Nothing enforces them. Our own test is the evidence: the model followed the format, but it followed it because it chose to, not because anything stopped it doing otherwise. Anything that must not happen needs a human review step, not a paragraph in a markdown file.
It is not a knowledge base. A skill can bundle reference files, but the model still has no access to your CRM, your custodian, or a county record unless something connects it. In our test it could name the right source and could not look anything up there. Finding out what a record actually says is a separate job, and it is the one our owner research does.
It is not free of risk. Anthropic's own security guidance is blunt: installing a skill is equivalent to installing software, so audit skills before use and be wary of any that fetch external URLs. A skill from a stranger is a set of instructions you have not read, running against your files. Treat a downloaded skill the way you would treat a browser extension that asks for everything.
Three skills worth building for a practice
Three where a skill earns its place, and one where it does not.
The pre-meeting brief, above. It qualifies because it is recurring, it has a format worth enforcing, and its failure mode is quiet. To work through the method before automating it, see the prospect research brief method and our guide to researching a prospect before the first meeting.
The plain-language translation. Turning a plan output into something a client can read is recurring, and the thing you want enforced is a reading level and a ban on jargon substitutions that quietly change meaning. A skill holds that better than your memory does at 6pm.
The claim audit. Before anything client-facing goes out, a pass that marks every checkable claim and every claim that cannot be checked. This is the highest-value one, because it is the step people skip. You can try the approach without building anything using our AI claim auditor, and if the copy is marketing, the SEC marketing rule checker covers the adjacent ground.
Where a skill does not help: anything requiring current facts. A skill cannot make a model know this quarter's numbers. That needs retrieval, which is a different mechanism entirely.
How to build one this week
- Pick a task you did three times last month and will do again.
- Write down what you tell the model each time, including the corrections you make after the first draft. That text is most of your skill already.
- Add the three elements from ours: a source rule, explicit permission to return an empty answer, and a self-check.
- Save it as
SKILL.mdwith anameand adescriptionthat says when to use it, in a folder named after the skill. - Run your laziest version of the request and see whether it fires. If it does not, fix the description first.
- Run it twice more on different inputs. You are checking that the shape holds, because the shape is what you are buying.
Frequently asked questions
Do I need to be able to code to use an agent skill?
No. A skill is a markdown file, which is a text file with headings. Everything in the example above was typed in a plain text editor. Skills can optionally bundle scripts, and that part does require code, but the format does not require it and most useful advisor skills will not have any.
Are agent skills the same as a custom GPT?
They overlap in purpose and differ in form. A custom GPT is a configured assistant that lives inside ChatGPT. A skill is a file. The practical difference for a firm is that a file can be version controlled, diffed, and handed to a reviewer as a document, and it is portable to any tool that reads the format, which since December 2025 has been an open standard rather than one vendor's feature.
Will a skill stop the model from making things up?
Not reliably, and our test is the reason we say so rather than assume it. Both the run with the skill and the run without it declined to invent a stranger's net worth. A skill changes the format and the emphasis of an answer. Treat any claim that it eliminates fabrication as unproven, and keep a review step for anything client-facing.
Can my whole firm share one skill?
On claude.ai, yes, if the firm is on a Team or Enterprise plan. The organisation's owners can provision a skill for every user, and members can share view-only copies with colleagues. On Free, Pro or Max accounts there is no sharing, so a shared standard means each person installing the same file and someone owning the job of keeping those copies identical. Either way, decide who is allowed to change the firm's skill and how that change gets reviewed. In Claude Code, where skills are files in a project directory, normal version control handles both.
How often should I revisit a skill I have written?
Whenever the underlying job changes, and once after the first month of real use. The instructions you write on day one reflect what you thought the task needed. The corrections you find yourself typing after the model responds are the actual specification, and they belong in the file.
What to do with this
If you take one thing from our test, take the unglamorous version of the result. A skill did not make the model more truthful. It made it produce the same, reviewable, deliberately-mostly-empty document every time, which is what you want from anything that touches a client relationship.
Start with one task. Write the file. Run it three times. If the shape holds, you have removed a recurring decision from your week, and the tools underneath you can change later without taking your instructions with them.
For where this fits alongside everything else advisors use AI for, start with how financial advisors use AI, and see what we publish about our own limits for how we handle the same questions about our product.

