Claude, Build Guide.

Using build-guide as a named example, this video explains that a Claude Skill is a single SKILL.md file read top to bottom, which delivers repeatable results but nothing outside what it wrote down.

1:42 video3 min readWatch on YouTube

It sounds like an upgrade when Claude "picks up a skill" called build-guide, as if it studied something and came back knowing more. It did not. This video uses build-guide as a named example to explain what a Claude Skill actually is, and just as importantly, what it is not.

Nothing in the model changes

The clearest way to see that a skill is not learned knowledge is a simple test: delete the skills folder, and Claude forgets nothing, because there was nothing to learn in the first place. It just stops following that one routine. A skill does not modify the model, it is an external file that gets read before a task starts.

One folder, one file

build-guide is a folder holding a single file, SKILL.md, which contains the whole routine written in plain language. 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 instructs a branch. There is no hidden logic sitting outside the document, the file is the program.

A specification, not a new power

This framing has a direct payoff and a direct limit. The payoff is repeatable results: run after run, the same request produces the same routine, because Claude is reading the identical file each time. The limit is anything the file never wrote down, once a situation falls outside the written steps, Claude is off the map, working without a specified procedure rather than falling back on some deeper trained skill.

Neither following it nor deviating from it proves anything

The video makes a point worth sitting with: watching Claude follow a skill file exactly does not prove the routine itself is good, a bad process can be followed to the letter just as easily as a good one. And watching Claude improvise or deviate does not prove the skill is broken either; it may simply be a case the file never covered. Judging a skill means reading what it actually says, not just watching whether Claude complies with it.

What this video does not claim

This video does not describe what build-guide's own SKILL.md specifically instructs for legal-clinic work, that file was not reachable from this build, and even the source material behind this video left the same detail unfilled in several places. Rather than invent a legal-specific procedure, the video uses "build-guide" only as the named example of a skill-shaped folder, one file, SKILL.md, and sticks to what is generically true of any Claude Skill: a folder read before Claude acts, executed step by step, with a stated payoff and a stated limit.

Key takeaways

  • A Claude Skill is a folder holding a SKILL.md file, not new knowledge baked into the model.
  • Deleting the skill folder erases nothing Claude knows, it only stops that routine from being followed.
  • Claude reads the file top to bottom and executes it in order, branching only where the file says to.
  • The payoff of a skill is repeatable results across runs; the limit is anything the file never specified.
  • Following a skill exactly does not prove it is a good routine, and deviating from it does not prove it is broken.

Who this is for

Anyone trying to understand what a Claude Skill actually is before writing their own, and anyone who wants a plain, non-hype explanation of why "Claude learned a new skill" is the wrong way to describe what happened.

Chapters

  1. 0:00The naive framing: "Claude learned a new skill"
  2. 0:09Sounds like an upgrade
  3. 0:16Broken, with a case — delete the folder, nothing is forgotten
  4. 0:25The anchor: one file — build-guide's single SKILL.md
  5. 0:36Read it, then follow it in order
  6. 0:44Spec, not power — the payoff and the limit
  7. 0:54The anchor returns — same file, every run
  8. 1:04Both directions — neither one is proof
  9. 1:16Carry-out
  10. 1:25Your turn
  11. 1:37Outro
Full transcript(auto-generated, with timestamps)

The naive framing: "Claude learned a new skill"

[0:00]You guess Claude learned a new trick called build guide. It didn't. It was handed a file of instructions. Build guide just names one such file. Let's see what's actually inside it. Here

Sounds like an upgrade

[0:10]Claude has a new skill and it sounds like an upgrade, like it studied and came back knowing something it didn't before.

Broken, with a case — delete the folder, nothing is forgotten

[0:16]But nothing in the model changes. Delete the skills folder and Claude doesn't forget anything. There was nothing to learn. It just stops following that one routine.

The anchor: one file — build-guide's single SKILL.md

[0:25]Here's what a skill actually is, one file. Build guide, for instance, is a folder holding a single skill. MD, the whole routine written in plain language Claude reads before it starts. Claude

Read it, then follow it in order

[0:36]Reads that file top to bottom and works through it step by step in the order it's written, no branching unless the file itself says branch. That makes a

Spec, not power — the payoff and the limit

[0:45]Skill a specification, not a new power. The payoff, repeatable results run after run. The limit, anything outside those written steps and Claude is off the map.

The anchor returns — same file, every run

[0:54]So build guide never taught Claude anything new. It just guarantees that every time build guide runs, Claude reads that same file and works through the same steps. That's the whole trick.

Both directions — neither one is proof

[1:04]Watching Claude follow the file exactly doesn't prove the routine is good. A bad process can be followed to the letter and watching it improvise doesn't prove the skill is broken. It may just be a case the file never covered. A skill

Carry-out

[1:17]Doesn't teach Claude something new. It's a file of steps Claude reads before it starts, so the same request gets the same routine every time.

Your turn

[1:25]Your turn. Paste this into Claude. Pick one task I repeat 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

Outro

[1:38]Build guides Liam in forbear.

More from Claude for Education

Humanitarians AI Lyrical Literacy Project