Claude, Client Letter.
Liam explains that Claude's client-letter skill does not teach it to counsel clients, it is a single SKILL.md file of ordered steps that guarantees the same letter structure every time, nothing more.
Liam, in for Bear, takes on a natural but wrong assumption: that Claude "learned" how to write client letters from a skill called client-letter. It did not. It was handed a file of instructions, and client-letter is just the name of that one file.
Why this sounds like a capability upgrade
When Claude has a client-letter skill available, it sounds like a writing upgrade, as if it studied real client letters and came back knowing how to counsel someone in its own words. But nothing about the model itself changes when a skill is added. Delete the skills folder and Claude does not forget how to write client letters, because there was nothing to learn in the first place. It just stops following that one routine.
What a skill actually is
A skill is one file. The client-letter skill is a folder holding a single SKILL.md file, containing instructions for turning a case update into a clear client letter, written in plain language that Claude reads before it starts. Claude reads that file top to bottom and works through it step by step, in the order it is written, with no branching unless the file itself says to branch.
Specification, not power
This makes a skill a specification, not a new capability. The payoff is a consistent structure applied to every letter: what happened, what it means, what happens next. The limit is that anything outside those written steps is genuinely off the map for that routine. The client-letter skill never taught Claude how to counsel a client. It guarantees only that every time it runs, Claude reads the same file and produces the same letter structure.
Neither a good letter nor a bad one proves much
The video makes a point of running this in both directions. Watching Claude produce a clear client letter does not prove it understood the underlying matter, a file can be followed precisely on a case Claude never really grasped. Equally, watching it produce a clumsy letter does not prove the skill itself is broken, it may simply be a case the file's structure does not fit well. Neither outcome is proof of understanding or its absence.
Try it yourself
The video closes with a direct prompt: pick one document you produce the same way every time, ask Claude to write a SKILL.md for it in plain language with ordered steps, then have it read that file back to you and walk through exactly what it will do before doing it. Running this against a document you actually produce repeatedly shows the mechanism directly, rather than taking the claim on faith.
Key takeaways
- A Claude skill is a specification the model follows, not a capability it learns or internalizes.
- Deleting a skill's folder does not erase any ability, because nothing was actually learned, Claude just stops following that one routine.
- The client-letter skill's payoff is a consistent letter structure, what happened, what it means, what happens next, every time it runs.
- The limit of a skill is anything its written steps never cover; Claude does not improvise beyond the file.
- A well-written or poorly-written output is not evidence, on its own, of whether Claude understood the underlying matter.
Who this is for
This is for legal professionals and anyone using Claude skills who wants an accurate mental model of what a skill actually does, rather than assuming it represents new learned ability.
Chapters
- 0:00The naive framing: "Claude learned client letters"
- 0:11Sounds like a writing upgrade
- 0:20Broken, with a case — delete the folder, nothing is forgotten
- 0:29The anchor: one file — client-letter's single SKILL.md
- 0:43Read it, then follow it in order
- 0:51Spec, not power — the payoff and the limit
- 1:03The anchor returns — same file, every run
- 1:14Both directions — neither one is proof
- 1:29Carry-out
- 1:38Your turn
- 1:51Outro
Full transcript(auto-generated, with timestamps)
The naive framing: "Claude learned client letters"
[0:00]You guess Claude learned how to write client letters from a skill called client letter. It didn't. It was handed a file of instructions. Client letter just names one such file. Let's see what's actually inside it. Here Claude
Sounds like a writing upgrade
[0:11]Has a client letter skill, and it sounds like a writing upgrade, like it studied real client letters and came back knowing how to counsel someone in its own words. But nothing in the model
Broken, with a case — delete the folder, nothing is forgotten
[0:21]Changes. Delete the skills folder and Claude doesn't forget how to write. There was nothing to learn. It just stops following that one routine.
The anchor: one file — client-letter's single SKILL.md
[0:29]Here's what a skill actually is, one file. Client letter, for instance, is a folder holding a single skill. MD instructions for turning a case update into a clear client letter, written in plain language Claude reads before it starts. Claude reads that file top to
Read it, then follow it in order
[0:44]Bottom and works through it step by step in the order it's written, no branching unless the file itself says branch. That
Spec, not power — the payoff and the limit
[0:52]Makes a skill a specification, not a new power. The payoff, the same structure every letter, what happened, what it means, what happens next. The limit, anything outside those written steps, and Claude is off the map. So client
The anchor returns — same file, every run
[1:03]Letter never taught Claude how to counsel a client. It just guarantees that every time client letter runs, Claude reads that same file and produces the same letter structure. That's the whole trick. Watching Claude produce a
Both directions — neither one is proof
[1:15]Clear client letter doesn't prove it understood the matter. A file can be followed to the letter on a case it never really grasped. And watching it produce a clumsy one doesn't prove the skill is broken. It may just be a case the file structure doesn't fit. A skill
Carry-out
[1:29]Doesn't teach Claude how to counsel a client. It's a file of steps Claude reads before it starts. So every letter comes out with the same structure every time. Your turn. Paste this into Claude.
Your turn
[1:40]Pick one document I produce the same way every time. Write me a skill MD for it, plain language, ordered steps, then read it back to me and walk me through exactly what you'll do before you do it. Claude client letter, Liam in for bear.
