Transformations vs generators: a pattern from 20 Claude Skills
I shipped 20 Claude Skills as a paid pack two weeks ago. Reviewing them now from the seat of "which would I actually rerun", one structural cut keeps coming up. The skills that hold up are transformations of existing material. The ones that generate from a blank input read as plausible noise.
The breakdown
Sorting the 20 by the shape of their input:
| Category | Examples | |
|---|---|---|
| 8 | Transformation | changelog-from-commits, support-ticket-to-faq, meeting-actions, standup-from-git, testimonial-extractor, weekly-metrics-narrative, twitter-thread-from-blog, docs-to-faq |
| 5 | Template + context | customer-reply, cold-email-personalize, sales-followup, onboarding-email-sequence, investor-update |
| 3 | Audit / critique | landing-page-audit, pricing-page-feedback, bug-triage |
| 4 | Pure generator | icp-generator, competitor-pricing-table, press-release-launch, lead-qualifier-bant |
The 8 transformations are the only category that has a stable acceptance criterion. You can hold the output next to the input and ask "did the skill faithfully reshape what I gave it". You either yes or no.
The template-and-context skills are next. There's a fixed scaffold (a follow-up email, an onboarding sequence) and the skill is fitting context into that scaffold. Verifiable: did it use the inputs, did it stay inside the scaffold.
The audit skills are slightly weaker because the criterion is taste-based, but the output points at concrete things on a real page. You can disagree with item 3, agree with item 1, and still ship a usable artifact.
The pure generators are the weakest. They take a one-line prompt and produce a multi-paragraph artifact that I have nothing to verify against.
What "rerun" means
The honest test for a skill isn't "did it produce something good once". It's "would I run this on Tuesday and again on Friday".
For changelog-from-commits the answer is yes every release. The input changes (the commits), the desired shape stays fixed, the output is checkable against the commits. For icp-generator the answer is no. I run it once, get a plausible profile, and either accept it (and I could have written it myself) or reject it (in which case the skill didn't help). It doesn't repeat.
The Ilja rule
The sharpest framing I've seen on when a workflow earns being made into a skill is from Ilja Becker:
If I'd say "do it like last time" 3+ times, build the skill. — Ilja Becker, on Bluesky
Reread that with the transformation/generator cut in mind. "Like last time" presumes there is a last time. Presumes raw material that needs the same reshaping again. The skill is reshaping it, not making it appear from nothing.
Generators don't have a "last time". The blank input is fresh every run. That's part of why they don't compound.
The two I'd kill from my own pack
press-release-launch. The user gives me a one-paragraph product description. I produce 350 words of marketing voice. I have no way to verify it represents the product correctly, and the output has the same fingerprint every time you run it. Better as a Word template than as a skill.icp-generator. The user describes their product in a sentence. I invent an Ideal Customer Profile. The output is plausible. It is not grounded. It is the same shape every time. A real ICP is built from interviews and analytics, not from one prompt.
Both are fill-the-blank-page work. The model can do them. Every output reads the same. They were the easiest to write and they are the least useful to ship.
Where the cut might break
Two cases I'm not sure about:
- Brainstorm-style generators where the user wants 30 ideas to throw most away. The input is small and the output is unverifiable, which my cut would call a generator. But the goal isn't "produce one correct output", it's "produce 30 prompts that get the user thinking". Maybe a different mode.
- Specification skills where the user describes what they want and the skill produces a spec or schema. There is a verification path (the spec compiles, runs, fits the existing system). My cut would call them generators because the input is descriptive, but I think they belong with audits.
What I'm doing about it
This week I'm rewriting the 4 pure generators and probably replacing 2 of them with skills that have a real input. icp-generator wants to become icp-from-customer-list — give it your last 20 paying customers, get a sharper profile. press-release-launch can become press-release-from-changelog: feed it the actual release notes, get a release-grounded press release.
If you bought the pack, those updates land in the same ZIP at no cost. If you haven't, two of the twenty are free here with full source rendered in place.
Worked example: Camille Roux's tech-debt-audit skill
To make the framework concrete, here's the audit shape applied to a real public skill: Camille Roux's tech-debt-audit — one command on a repo, produces a TECH_DEBT_AUDIT.md.
Input: a codebase (substantial real material — files, structure, dependencies, test setup).
Output: TECH_DEBT_AUDIT.md — a named artifact in a fixed format.
Acceptance criterion: can the human verify the audit's claims against the actual repo state? (yes for each item — file path, dependency version, missing test, etc.)
Verdict: SKILL ✅
Reruns: yes — every release, every quarter, every time the codebase changes. The input is different each run; the output shape stays fixed; the diff between runs IS the value.
Why it isn't a generator: there is a "last time" — last quarter's audit. The new audit isn't blank-page work; it's reshaping the new repo state into the known format. Comparable. Trackable. Compounds.
Stronger version (what an audit verdict would suggest): two additions worth shipping — trend across runs (which dimensions regressed since last quarter), and actionable in this sprint markers. Both build on the transformation shape rather than fight it.
That's roughly the shape of what you'd get for €19 — except about your workflow, and delivered in 12 hours.
Want me to do this for your workflow?
€19. Same-day audit. One workflow.
If you read this and you're staring at one of your own workflows wondering which side of the cut it's on, that's the question this service answers.
You give me one paragraph describing the workflow plus one real input/output pair. I send back a 200-400 word verdict: skill / not-skill / borderline, and if skill — the exact input → output shape it would take and the trigger language to make Claude pick it up. Same day, within 12 hours.
No skill files built. No fluff. Just the shape decision so you only spend the 30-90 min of skill-building when it actually compounds.
Refund if I can't deliver an honest verdict within 24 hours.
Open question I'd like an answer to: where does the transformation/generator cut break? Is there a class of skill that's purely generative and still earns its rerun?
Send a reply on Bluesky or email hello@from100to200.com. I read everything.