Skip to main content
Tested with receipts

5 free Claude Code skills we tested on our real Shorts

We use all five of these skills to make our Shorts. Before recommending them, we gave each one a real job from our own pipeline and ran it twice in headless Claude Code, with the pass rule written down before the first run.

Four passed both runs, one passed one of two, and the hook scorer called our best Reel weak. All 10 runs together cost $1.29 in billed usage.


At a glance

The results

Job: Score the first lines of our last 11 Instagram Reels

Ran cleanly. Called 10 of 11 weak, including our 1,200-play Reel. 0.07 correlation with real plays.

2 of 2 · $0.13

Job: Write an Instagram caption for a real Short

AI tells went from 4 (no skill) to 1, with 0 banned patterns in both runs

2 of 2 · $0.36

Job: Build a word-caption video component

2 of 2 rendered on the first try, captions on the right word

1 of 2 · $0.39

Job: Build an animated scorecard from remocn components

1 of 2 rendered; the other imported a font our Remotion version does not ship

2 of 2 · $0.26

Job: Turn our top 3 Shorts scripts into a voice skill

Valid skill both times, in 47 to 54 seconds, for 12 to 14 cents

Cost = billed usage for both runs of each skill. Total for all 10 runs: $1.29.


Skill 1

instagram-agent-skill (hook scorer)

By Jakeschincariol · License: MIT · Passed 2 of 2 runs · $0.15

What it did for us

A set of Instagram skills. The part we use is hookscore.py, which scores a Reel's first line on length, specifics, stakes, front-loading and who it talks to. We gave it the opening lines of our last 11 Reels. It rated 10 of them weak, including the one that got 1,200 plays (it scored 30.5). The rest got 126 to 225 plays. Its ranking had a 0.07 rank correlation (Spearman) with real plays, which is close to none.

Honest limit

It is a formula, not a model trained on views; the repo's own notes put its within-creator accuracy (AUC) at 0.56. Eleven Reels is a tiny, noisy sample. We still run it on new scripts because it reliably flags greetings, slow openers and lines with nothing concrete in them.


Skill 2

marketingskills

By Corey Haines · License: MIT · Passed 2 of 2 runs · $0.13

What it did for us

A large library of marketing skills. The copywriting skill ships an AI tells reference: a list of patterns that make copy read as machine-written. Without the skill, Claude's caption had 4 tells (2 emoji, a colon reveal and 1 flagged word). With the skill, both runs had 1 (a single flagged word) and no banned patterns.

Honest limit

One run came in at 161 words against the 150-word limit we asked for. Our checker is a fixed pattern list and misses some tells, such as "X, not Y" endings.


Skill 3

Remotion skills

By Remotion · License: No license file in the repo at time of writing · Passed 2 of 2 runs · $0.36

What it did for us

Remotion's own agent skills for writing Remotion video code. Both runs built a 1080x1920 caption component that passed all 5 of our checks, rendered on the first try and highlighted the right word at the right time (frame 75, the word "real").

Honest limit

The skill tracks the latest Remotion. We run 4.0.259, so check the APIs it suggests against your installed version.


Skill 4

remocn

By the remocn team · License: MIT · Passed 1 of 2 runs · $0.39

What it did for us

Free copy-paste Remotion components (text animations, counters, transitions). Run 1 built the scorecard from two remocn components, BlurOutUp and RollingNumber, and rendered. Run 2 imported the Geist font from @remotion/google-fonts, which our Remotion 4.0.259 does not include, so type-checking and rendering failed. The fix is one line.

Honest limit

Even the passing run needed a human look: the font fell back to a serif and the headline had blurred out by the last frame. remocn assumes the Geist font and a recent Remotion. We also handed it the registry as local files; the real install (shadcn add) pulls in extra helpers.


Skill 5

skill-creator

By Anthropic · License: Apache-2.0 · Passed 2 of 2 runs · $0.26

What it did for us

Anthropic's skill for writing new skills. From the transcripts of our three best Reels it wrote a theaihow-voice skill both times, with valid frontmatter and 4 of the 6 voice features we checked for, in 47 to 54 seconds and 12 to 14 cents per run.

Honest limit

Our checker missed the opening-line feature in both runs, though run 2's skill does describe it; we kept the score as registered. We skipped skill-creator's eval loop to keep the test cheap. For a skill you depend on, run the evals.


Free download

Download our test files

Everything behind the numbers on this page: the pass rules we wrote before the first run, the results, the harness and graders, the inputs, and a billed receipt for every run (cost, turns, duration, tokens), plus what each agent produced. The README inside explains each folder and how to rerun the test.

It does not include the skills themselves. Get those from the repos linked above; the zip lists the exact commit we tested.

Download test files (zip, 380 KB)

Get the next tested setup

Get our tested Claude Code setups and test files by email, one email when a new test ships.

Double opt-in. Every email carries a one-click unsubscribe link.


Method

How we tested


Questions

FAQ

Are these Claude Code skills free?

Yes. All five are free on GitHub. You pay only for the Claude usage they consume; our 10 test runs came to $1.29 in total.

How do I install a Claude Code skill?

Copy the skill's folder (the one with SKILL.md) into ~/.claude/skills/ to use it in every project, or into .claude/skills/ inside one project. Some, like remocn, also ship as a Claude Code plugin.

Does the hook scorer predict views?

No. On our 11 Reels its scores had a 0.07 rank correlation with real plays, and it called our best Reel (1,200 plays) weak. Use it to catch weak openers, not to forecast performance.

Which skill should I install first?

If you write any copy, marketingskills' AI tells list gave the clearest before-and-after in our test (4 tells down to 1). If you make videos in code, start with the official Remotion skills, which rendered first try in both runs.

Why did remocn fail one run?

The agent imported the Geist font from @remotion/google-fonts, which our Remotion version (4.0.259) does not include. Adding the font or picking one you have installed fixes it.

Can I see the raw test files?

Yes. Download the zip on this page: pass rules, results, harness, graders, inputs and a receipt for every run, with a README on how to rerun it.

How did you test them?

Headless Claude Code (claude -p --output-format json) on Sonnet, acceptEdits, safe mode, in an empty folder outside our repo. Each skill ran twice, and the caption and hook jobs also had a no-skill baseline with the same prompt. Pass rules were written and committed before the first run, and every receipt (cost, tokens, outputs) is kept.

Can Eight Labs help my team use Claude Code?

Yes. Eight Labs consults for enterprise, SMEs and startups on Claude Code adoption, agents and AI systems. Reach us through the contact form on theaihow.com.

All five skills belong to their authors; links and licenses above. We are not affiliated with any of them. Building AI at your company? Eight Labs consults for enterprise, SMEs and startups. Want your own setup checked? See the Claude Code Setup Review.