Claude Code · 5 groups · 40 skills

Of the skills you installedmost never fire

Not because they are bad. Claude reads only the name and description of each skill at session start, so a skill whose description does not match your sentence is never opened. That is a grouping and wording problem, and it is the one this fixes.

name it and it loads
the fix that takes one word
4 more
causes, in the playbook

Start by measuring, not by buying

Two scripts read the session logs Claude Code already writes on your own disk and tell you where the tokens went. Free, no account, nothing is uploaded. They print what is leaking and which of these packs, if any, actually fixes it.

Why this first

Of the leaks it commonly finds, most are fixed by the free pack and a discipline file it writes for you. It says so instead of pointing everything at a purchase.

Where the money went

Spend, how much of it was re-reading the conversation rather than working, and which tools fill your context.

any computer
npx --yes claude-tokenscan

What you carry before you type

Every installed skill loads its name and description at the start of every session. This weighs that block and finds the folders that can never load.

any computer
npx --yes claude-tokenscan preamble
Never used a terminal? Step by step nothing is uploaded · no sign-up · about two minutes

Five groups, by the problem they solve

Each group is the set of skills you reach for in one situation. Group 01 is free. Groups 02 to 05 come together in the $9 bundle, with the playbook.

01

Say less, build less 11 skills

For anyone who keeps hitting usage limits, or gets 300 lines back for a 3-line job.

Same work, fewer tokens, and no code you did not ask for.

Plain Claude
Long answers, the same explanation every session, an interface and a config file for a one-off function.
With this group
Short answers, smallest diff, and it reads the code before it edits.
  • Fewer tokens per answer
  • Stops over-building before the code exists
  • Investigates the cause before touching files
the 11 skills
cavemancaveman-commitcaveman-reviewponytailponytail-auditponytail-reviewinvestigate-firstlean-buildsafe-refactorsurgical-patchverify-and-stop
Free
pay what you want
Start free

npx claude-skillpacks <key> say-less

02

Plan it, then run it 7 skills

For long jobs where the agent drifts, forgets the plan, or declares victory halfway.

A plan on paper, then steps that actually get executed in order.

Plain Claude
Starts coding from the first idea, loses the thread in long sessions, and says "done" from memory.
With this group
Interrogates the idea first, writes a step plan you can check, then works one step at a time.
  • Big jobs split into verifiable steps
  • Independent tasks handed to subagents
  • Your own rules become a skill you reuse
the 7 skills
brainstormingwriting-plansspec-driven-developmentexecuting-planssubagent-driven-developmentdispatching-parallel-agentswriting-skills
$9
in the bundle
Get the bundle

npx claude-skillpacks <key> full-bundle

03

Make it work, prove it 5 skills

For anyone who has heard "fixed!" and found the same bug an hour later.

Tests before code, root cause before the patch, and a real run before "done".

Plain Claude
Guesses at the fix, patches where the symptom showed up, and calls it finished without running it.
With this group
Writes a failing test first, traces the cause, and runs it before saying done.
  • The same bug stops coming back
  • Every caller of the broken function gets checked
  • "Done" means it was executed, not assumed
the 5 skills
test-driven-developmentincremental-implementationverification-before-completionsystematic-debuggingdiagnosing-bugs
$9
in the bundle
Get the bundle

npx claude-skillpacks <key> full-bundle

04

Review, merge, harden 9 skills

For the hour before a PR, when nobody else is going to read the diff.

A second pair of eyes on the diff, and a clean way out of the branch.

Plain Claude
Praises the code, flags formatting, and leaves the real bug at item three. Merge conflicts get resolved by coin flip.
With this group
Problems first, ranked by severity, and conflicts resolved after reading both intents.
  • Correctness before style
  • Security and performance passes before release
  • Branches that start and finish cleanly
the 9 skills
requesting-code-reviewcode-reviewreceiving-code-reviewcode-simplificationresolving-merge-conflictssecurity-and-hardeningperformance-optimizationusing-git-worktreesfinishing-a-development-branch
$9
in the bundle
Get the bundle

npx claude-skillpacks <key> full-bundle

05

Ship it and show it 8 skills

For when the code works but the page looks templated and the launch post will not write itself.

Frontend that is not purple-gradient default, diagrams, launch copy, plus the three skills we wrote.

Plain Claude
Purple gradient, three rounded cards, emoji icons. Then a launch post full of "revolutionary".
With this group
Reads the brief and picks a direction, draws the architecture, and writes copy that sounds like a person.
  • Landing pages that do not look templated
  • Architecture and sequence diagrams as self-contained HTML
  • Windows shell traps, real-UI recording and pre-publish checks, written by us, not on GitHub
the 8 skills
design-taste-frontendarchifycopywritinglaunchsocialwindows-claude-coderecording-real-uiverify-before-publish
$9
in the bundle
Get the bundle

npx claude-skillpacks <key> full-bundle

All 40 skills, all five groups

Five groups in one zip, plus the playbook. Nothing to choose between.

  • One download, one install, one licence file per author
  • PLAYBOOK.md: why an installed skill never fires, a symptom-to-skill map, five workflows, prompts to paste
  • Overlapping skills were removed. One tool per job, not three that do the same thing
  • Three skills written for this bundle, so it is not a folder you could have cloned
$9
one-time
updates included
group 01 stays free
Buy everything

npx claude-skillpacks <key> full-bundle

One command, no installer

The free group and the bundle install the same way. Your licence key is in the Gumroad receipt email.

npx claude-skillpacks <your-license-key> full-bundle
# skills go to ~/.claude/skills
# the guide, notices and licences go to ~/.claude/skillpacks/full-bundle
# restart Claude Code to pick them up
say-lessGroup 01, free
full-bundleAll five groups, $9

It checks the key against Gumroad, unpacks the skills, and stops there. Nothing stays running, and it writes to those two folders only. Prefer to do it yourself? The zip is a plain download and the folders copy by hand.

What you get

A skill is a SKILL.md file: a procedure your agent reads when your request matches it. These are grouped, deduplicated, and documented.

40skills across five groups
14guides in the bundle: the order to use the skills in, and prompts to paste
3skills written for this bundle
1playbook, ten sections

What you actually type

Skills are not slash commands. You name one in a normal sentence, and it loads. These come straight out of the guides.

investigate-first. Read only, no edits.

Stops the agent from editing while it is still guessing.

systematic-debugging on this error. Do not fix anything yet.

Cause first. Without the second sentence it starts patching.

surgical-patch. Minimal diff. Tell me the file count first.

Turns a six-file rewrite into a one-line fix.

Check every claim in this copy against the facts file.

From verify-before-publish, one of the three we wrote.

Inside the zip

claude-skillpacks-bundle/
├── PLAYBOOK.md           install fixes · why skills never fire · 5 workflows · prompts
├── field-notes/          windows-claude-code · recording-real-ui · verify-before-publish
│   ├── GUIDE.md          when to reach for it, in what order
│   └── NOTICE.txt        original work, not redistributed files
├── debug-squad/
│   ├── systematic-debugging/SKILL.md
│   ├── diagnosing-bugs/SKILL.md
│   └── GUIDE.md · NOTICE.txt · LICENSES/
└── … 12 more packs

What we left out, and why

Seven skills that would have padded the count did not make it in. Here they are by name.

Dropped for licensing

karpathy-guidelinesno licence file in the source repo
hallmarkwe could not verify who owns it
high-end-visual-designwe could not verify who owns it
baoyu-designwe could not verify who owns it

Popular is not the bar. If we cannot ship the licence, we do not ship the skill.

Dropped for overlap

code-review-and-qualitythe same job as code-review
debugging-and-error-recoverycovered by the other two debugging skills
planning-and-task-breakdowncovered by writing-plans

Three skills that all claim the same work make the agent pick badly. One per job.

We ran the test

Same task, same model, twice. One run got the pack's skill text in front of the prompt; the other got nothing. A third model graded both, blind, against an answer key written before either answer existed, six criteria at five points each.

TaskPackWith packPlain
Diagnose a bug where the file looks fine and the audio is silent debug-squad 27/30 25/30
"Can we ship?" on a status report with no command output in it build-verify 30/30 24/30
Six Windows shell failures that have nothing to do with the code field-notes 26/30 25/30
Total 83/90 74/90

Where the pack lost

  • On the Windows task both answers got all six diagnoses right. The pack did not make it more correct. It made it half the length, and won by one point on compression.
  • On the debugging task the plain answer was the better-written document and won signal-to-noise 5 to 3. The pack answer padded.
  • Only the "can we ship" task was one-sided. There the plain answer won no criterion at all.

What it cost

  • The skill text is input tokens every time it loads: roughly 1,000 to 4,700 per run in this test.
  • It paid that back once, halving the Windows answer, and did not on the debugging task, which came back 34% longer.
  • Three tasks, one run each, one grader. A one-point margin at this size means nothing. Only the six-point margin would likely survive repeating.

Every prompt, both answers and all three scorecards are on one page. We did not edit them.

Before you buy

Four things that are true, and three things this is not.

SourceMost of these skills are free on GitHub. Every pack ships the author's licence and repo link.
Token costA loaded skill is extra context, so some answers cost more tokens. Group 01 pushes the other way.
InstallOne command, or a folder copy if you prefer. Nothing runs in the background and nothing is added to your PATH.
Tested onIt is tested against Claude Code. Other skill-reading tools should work; we do not promise it.
Not a big benchmarkOurs is three tasks, one run each. It is published in full, losses included, and it is too small to settle anything.
Not a subscriptionOne payment, updates included.
Not a smarter modelNo skill makes the model smarter. It changes what it reaches for, not what it knows.

Questions

It is installed but it never fires. Why?

Your request never matched its description. Claude Code reads only the name and description of every skill at session start. Name the skill in your message and it loads. Section 3 of the playbook covers the rest, including the folder sitting one level too deep.

How do I install it?

Run npx claude-skillpacks <your-license-key> <group-id> with the key from your receipt, then start a new session. By hand works too: unzip and copy the skill folders into ~/.claude/skills/, where each must end up as <skill>/SKILL.md. One level deeper and nothing loads. That is the most common mistake.

Does it work outside Claude Code?

A skill is a SKILL.md file with a YAML header, so anything that reads skills can use it. Claude Code is what we test against. Some packs also carry an agents/openai.yaml from their authors.

Updates and refunds?

Updates are free and arrive through Gumroad. If it is not useful, reply to your receipt and ask for a refund.

Can I resell it?

No. Third-party skills stay under their own licences, which ship in the packs. The field-notes skills and the playbook are ours. Use them in your work, do not repackage them for sale.

Stuck on something?

Reply to your Gumroad receipt with a screenshot of what you see. That reaches us, and a screenshot usually answers it in one message.