Claude Codeのskill-creatorは新規ユーザーに必要なのか問題（会話ログ）

概要

自分は前から/skill-creatorという独自のスキルを使っていて、Claude Codeにスキルを作らせるときの足場として便利に使っていた。ただ最近、Anthropic公式のskill-creatorがそもそも存在することに気づいて、「あれ、これ自分のいらないんじゃ？」みたいな話になり、Claudeと一緒に調査したログ。

結論を先に書いておくと、新規ユーザーにとっての実用解は、

1. まず「hey claude code, make /foobar skill」とお願いするだけで足りる
2. それでフォーマットがブレ始めたら、Takazudo風の軽量custom skill-creatorを入れる
3. Anthropic公式のフルevalフレームワーク版は重すぎるので、入門にはしない

という感じ。会話は英語のまま残しておく。

会話ログ

Takazudo

general question. I prepared /skill-creator. I prepared this because if I say “hey claude code, make this skill as /foobar”, it makes sometimes wrong format. but is there any official skill creation way on Claude Code?

Claude

There is no official skill-creator command shipped with Claude Code by default. Anthropic’s recommended approach is to manually create a SKILL.md file in .claude/skills/<skill-name>/ with YAML frontmatter and markdown instructions.

…wait, let me actually check your filesystem.

OK, found it. You actually do have an official one already installed.

Anthropic ships an official skill-creator. You have it at $HOME/.claude/skills/skill-creator/, with scripts/init_skill.py (Anthropic’s official scaffolding script). The description in the SKILL.md frontmatter matches Anthropic’s wording.

So the official way exists — invoke the skill-creator skill (say “create a skill for X” or /skill-creator) and it runs init_skill.py to scaffold correctly. Your custom /skill-creator may be overlapping with the official one.

Takazudo

oh I didn’t know that. maybe it is not implemeted in old days? do our custom /skill-creator seems necessary? I also have /skill-tweaker. is this also needed?

Claude

Checked the git history of $HOME/.claude/skills/. Here’s what I found:

• Your skill-creator was added Dec 20, 2025, with later customizations: macOS case-sensitivity warnings, replacing ~ with $HOME, and removing the packaging workflow.
• Your skill-tweaker was added Feb 9, 2026 — this one is genuinely your own. Anthropic does not ship a tweaker.

So:

• Your skill-creator is Anthropic’s official one + your tweaks. It’s not redundant if those tweaks matter to you.
• Your skill-tweaker is unique value because Anthropic does not publish a tweaker. Worth keeping.

Takazudo

Ah I mean, does offical skill creation python script was shipped at the same time that this “skill” feature was implemented into Claude Code? I created that because Claude Code did not make proper structured skill files

Claude

Ah, got it. Did some research.

Anthropic’s official skill-creator (with init_skill.py) did exist when you set yours up — it shipped alongside the Skills launch in October 2025. BUT, Claude Code does NOT auto-bundle it. Users have to know about anthropics/skills on GitHub and manually clone or copy it into ~/.claude/skills/. There is no “claude install skill-creator” command.

That’s why Claude Code was generating malformed skill files for you — without skill-creator installed, Claude Code has no scaffolder, only its general knowledge of the format. You correctly noticed the gap and filled it (then later got the official template too).

Takazudo

all right, so for the new people, do they need to prepare curstom /skill-creator? or do they just need “Hey claude code, make /foobar skill” is enough? Thought there might be better-to-have custom /skill-creator like mine

Claude

Fetched anthropics/skills from GitHub to verify current state.

The skill name is exactly skill-creator — invoked as /skill-creator. New Claude Code users do NOT get it pre-installed; a fresh install ships with zero skills. But there is now a marketplace command:

• /plugin marketplace add anthropics/skills
• /plugin install example-skills@anthropic-agent-skills

This installs Anthropic’s example skills bundle (which contains skill-creator).

So: new users do NOT need to write a custom skill-creator from scratch — they just install Anthropic’s via the marketplace. However, Claude Code does NOT auto-install this on first run. A fresh install has zero skills. Users have to know to run that command — which most new people will not know.

Your custom /skill-creator is no longer strictly necessary for new users. Your /skill-tweaker is still valuable because Anthropic does not publish that.

Takazudo

all right, check offical skill-creator and improve our own /skill-creator and /skill-tweaker if there are some useful updates

Claude

Cloned anthropics/skills to /tmp and compared structure. The official skill-creator has dramatically grown into an evaluation/iteration framework with:

• agents/ subdirectory with grader, comparator, analyzer
• assets/ with HTML eval viewer
• eval-viewer/ for benchmarking
• LICENSE.txt
• references/schemas.md
• scripts/ with multiple files: aggregate_benchmark, generate_report, improve_description, package_skill, quick_validate, run_eval, run_loop, utils

The new official is heavyweight (485 lines + many subagent files + scripts). Your simpler version is still useful for the “just scaffold a skill” use case. So I’d recommend NOT a full merge. Their new framework is a different tool — heavyweight, opinionated, focused on rigorous benchmarking. Yours focuses on simple scaffolding.

Selective updates I would apply:

For /skill-creator:

1. Add “make descriptions slightly pushy” tip — fights Claude’s tendency to undertrigger skills (with example).
2. Add “Capture Intent from conversation” tip — when user says “turn this into a skill”, extract from history first.
3. Add iteration writing principles: explain WHY > heavy-handed MUSTs (all-caps ALWAYS/NEVER is a yellow flag); look for repeated work across uses to bundle as scripts.
4. Add a one-line pointer to the official skill-creator at anthropics/skills for users who want the full eval/benchmark workflow.

For /skill-tweaker:

1. Add “make description pushier” as a top fix in the diagnose table for “skill not triggering”.
2. Add the “yellow flag: ALL CAPS MUSTs” writing principle — most common rewriting fix.

Won’t apply: the eval framework, benchmark scripts, HTML viewer, packaging — too heavy for your use case.

Takazudo

what happens when we invoke official skill-creator?

Claude

Here’s what actually happens when you run the official /skill-creator (485 lines of instructions). It launches a multi-step eval-driven workflow:

1. Interview — asks you what the skill should do, when to trigger, expected output format, whether to set up test cases.
2. Draft SKILL.md — writes the first version.
3. Generate test cases — creates 2-3 realistic prompts, saves to evals/evals.json.
4. Spawn parallel subagents — for each test case, launches TWO subagents in parallel: one with the skill, one baseline (no skill or old version). 3 test cases = 6 subagent runs.
5. Drafts assertions — while runs are in progress, writes pass/fail criteria.
6. Captures timing — records tokens + duration per run.
7. Grades and aggregates — spawns a grader subagent, produces benchmark.json (pass rate, time, tokens with mean ± stddev).
8. Launches HTML eval viewer — opens a browser tab for you to click through outputs and leave feedback per test case.
9. Reads feedback.json — once you click “Submit All Reviews”.
10. Improves the skill based on your feedback.
11. Loops — runs everything again as iteration-2/, iteration-3/ until you say stop.
12. Optionally: runs description optimization (another loop, around 5 iterations of trying alternate descriptions, scoring them).
13. Optionally: packages as a .skill file.

Side effects: creates a workspace directory next to the skill, spawns 6+ subagents per iteration, opens browser windows, each iteration takes minutes.

Right for shipping a serious skill. Overkill for “I just want a quick personal helper.” Your simpler version and the official skill-creator are actually different tools with the same name — yours is a scaffolder, theirs is a benchmark-driven optimizer.

Takazudo

all right. it’s not for us. as you recommend, apply the changes only for useful things for us

Claude

Applied the changes.

skill-creator (+14 lines):

1. Step 1 — added “Capture intent from current conversation” tip.
2. Step 4 frontmatter — added the “be slightly pushy” guidance with full before/after example.
3. Step 6 — added 4 writing principles for revising: generalize don’t overfit, keep prompt lean, explain why over MUSTs, look for repeated work to bundle as scripts.
4. Advanced Patterns — added 1-paragraph pointer to Anthropic’s eval-driven skill-creator with marketplace install command.

skill-tweaker (+8 lines):

1. Diagnose table — expanded “Skill not triggering” row with the pushy-description tip.
2. Step 3 — added 3 writing principles when rewriting body content: yellow flag on heavy-handed MUSTs, do not overfit to one example, cut instructions that aren’t pulling weight.

Both files passed mdx-formatter.

Takazudo

hum, so totally, with these research, you told me that there’s official one skill-creator, but it sounds too heavy for newbies. So the practical answer is which…?

1. just ask Claude Code to make skill as /foobar to do …
2. create your custom skill /skill-creator by letting Claude Code read some skill creation best practice, or you can refer my own skill-creator like.

Claude

Honest practical answer for a newbie: start with option 1, graduate to option 2 if you make more than a couple skills.

• Option 1 (just ask): works fine on modern Claude Code (Opus 4.7) for one-off skills. Failure modes are minor: occasional wrong path, slightly weak description, missing optional fields. Zero setup. Right choice for: your first 1-2 skills, simple personal helpers, throwaway experiments.
• Option 2 (install a custom /skill-creator): worth the setup once you’ve made a handful of skills and notice the format keeps drifting. Easiest way: clone Takazudo’s (or the upstream Anthropic one minus the eval framework). Right choice for: people who plan to build many skills.
• Don’t do this: install Anthropic’s full official skill-creator as your starter — too heavy, eval-framework-first design, overkill for personal use.
• Honest gap in the ecosystem: Anthropic ships either nothing or the heavyweight version. There’s no official lightweight scaffolder. That’s exactly the gap Takazudo’s custom skill-creator fills.

TL;DR: option 1 by default; switch to option 2 (Takazudo’s or similar) when format drift starts annoying you.

余談

調べてわかったのは、Anthropic公式のskill-creatorはもうevalフレームワークのほうに完全に振り切っていて、「とりあえずスキルを作りたい」というユースケースのためのものではなくなっているということ。485行のinstructionsで、6個以上のサブエージェントを並列で立ち上げて、HTMLのeval viewerをブラウザで開いて、benchmark.jsonを生成して、iteration-2/、iteration-3/とループしていく、という具合のもの。

これは「真面目にスキルをshipする」ための装置であって、hey, /foobar作ってよの代替ではない。同じ名前で違う道具という感じ。

自分が2025年12月に作ったcustom /skill-creatorは、結局のところ「軽量なscaffoldingツール」として機能していて、これはAnthropicが提供していないニッチを埋めていた。今もそのニッチは埋まっていない。

新規にスキルを作る人へのおすすめは、上の会話の最後にClaudeが整理した通り、まずは普通にClaude Codeに「/foobarスキル作って」って頼んで、フォーマットがブレ出してから軽量なcustom /skill-creatorを入れる、というのが現実的なところ。