Happy Path Works
Standards

How these are built.

A skill is craftsman-class only where it reliably makes the model do the thing that does not survive being done from memory — the step the model, left alone, would forget, skip, fudge, or confabulate. The operational form of the same test: remove the feature and ask whether the output changes. If it does not, the feature is not craftsmanship. It is packaging.
The quality floor

Pass/fail. A skill that fails any gate is a prompt, not a product.

1
Names the gap
Can you state in one sentence the specific thing the model gets wrong or does inconsistently without this skill?
2
Deterministic entry
Does the skill have a hard trigger — a prefix or an unambiguous condition — that fires it without relying on keyword guessing?
3
Enforced hard-fails
Is every stated hard-fail backed by a check that returns a real signal — exit code, count, diff — before the skill emits, not by prose alone?
4
Verify before voice
Is any output with a countable or specified constraint written around a verified result rather than asserted and hoped?
5
Loud failure, named exit
Does each scope boundary say "if X, refuse and do Y" rather than "try to stay in scope"?
6
No prior-chat references
Does the procedure stand without a conversation it does not contain — no "as we discussed", no "like last time" — given that a fresh session never had it?
7
Deciding example
Does at least one worked example show the skill catching a case that would otherwise go wrong, rather than the happy path succeeding?

Gate 2 may be supplemented by a keyword list but is not satisfied by one. Gate 3 passes only if every claimed hard-fail is mechanically checked — one unenforced hard-fail fails the gate.

Above the floor

What the price is for.

Mechanical enforcement. The market ships prose that says "always". A skill worth paying for ships a check that returns non-zero. A rule enforced by memory is a rule that is not enforced.
Graceful degradation. The skill knows its own failure mode and says so — "I can't do this cleanly, here is the handoff" — instead of confabulating a worse answer. Rare, because finding the failure mode requires testing rather than intention.
Re-runnable test. "Tested end to end" counts only if the test can be run again on demand. The eval is the asset. The memory of a passing run is not.
Anti-patterns

What looks like craftsmanship and isn't.

Fat keyword trigger lists. Padding that reads as thoroughness and fires unpredictably.
Prose hard-fails. A hard-fail that lives in a sentence is a suggestion.
Length guidance as a quality marker. Token economy is real and worth practicing; it is not evidence. A file trimmed to a budget and a file where every line is load-bearing look identical to a line counter, and only one of them changes what the model does. It may appear as guidance; it does not count as craftsmanship.
Comprehensive skill-file bloat. Non-load-bearing lines are context the model must hold and can misweight. Documentation is half the build — that governs your setup guide, not the size of the skill file.
Happy-path worked examples. An example of the obvious case succeeding is decoration.

One test covers all five. Does removing it change what the model does? If not, it is there to look like craftsmanship, which is the opposite of the thing being sold.

What never ships
Bare prompts. Untested skills. Undocumented behavior. Examples that only cover the happy path.
How things are tested
Every skill runs end to end before it enters a pack.
Update policy
Versioned, changelogged, buyers keep updates across a major version. Details at /policies/.
Support
One person answers hello@happypath.works. It is not staffed daily, so allow a few days for a reply. Bug reports get a fix or a documented workaround. Questions about whether a pack fits the way you work get an honest answer, including "no, and here is why". There is no ticket system, no tiered support, and no chatbot.
Why there are no reviews on this site

There are no testimonials here, no ratings, no customer count, and no founder to vouch for. This is a pseudonymous one-person studio and it intends to stay one.

What is offered instead is inspectable. The floor above is the same floor every skill sold here has to clear. The free audit tool measures your files against it and its source is public. The setup guides and worked examples are readable without a purchase. The changelog is dated. If that is not enough evidence to spend the price of a pack, then it is not enough — buy nothing until it is.

What this is not

It does not ask a model what it thinks of your skill. It runs a script, applies the same seven tests every time, and returns an exit code you can put in a build. Two runs against the same file produce the same verdict, which is not a property a model's opinion has.

It does not produce a score. Scores compress a specific finding into a number that feels like progress, and a number is the wrong output for "gate 3 fails because line 40 declares a hard-fail that nothing checks".

It does not fix your file. It tells you which gate failed and what is missing. The edit is yours — an unattended rewrite of a skill file is the failure mode this catalog exists to argue against.