SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsuse-skill-dl-util

use-skill-dl-util

Writing & research

View repository

Updated 10-16 05:53

About this skill

Skill Research

Search, download, and inspect AI coding skills from playbooks.com using skill-dl.

Trigger boundary

Use this skill when:

  • searching for existing skills on a topic before building a new one
  • downloading one or more skills from playbooks.com for comparison
  • building a corpus of skills for evidence-based skill design
  • piping search results directly into batch download
  • inspecting a downloaded corpus to decide what to inherit or avoid

Do not use this skill for:

  • building or authoring a new skill from a corpus (use build-skills)
  • testing a skill's instructional quality (use enhance-skill-by-derailment)
  • general web research unrelated to skill discovery (use run-research)

Prerequisites

Verify skill-dl is installed before doing anything else.

skill-dl --version

If skill-dl --version succeeds, continue. If it is missing, install with a saved-and-inspected installer, then verify again. On macOS, if bash --version is 3.x, install a newer Bash first with brew install bash.

curl -fsSL https://raw.githubusercontent.com/yigitkonur/cli-skill-downloader/main/install.sh -o /tmp/install-skill-dl.sh
sed -n '1,160p' /tmp/install-skill-dl.sh

# Run one of the next two commands.
# Linux, or macOS where `bash --version` is already 4+:
bash /tmp/install-skill-dl.sh

# macOS if `bash --version` is 3.x (the default /bin/bash is 3.2):
"$(brew --prefix)"/bin/bash /tmp/install-skill-dl.sh

skill-dl --version

Installation requirements: git plus curl or wget. The Bash 4+ requirement applies to the installer script, not to using a preinstalled skill-dl. No API keys are required for normal searches because skill-dl ships with built-in backend defaults. See references/download-patterns.md for output layout, batching, and troubleshooting.

Workflow

1. Define your search intent

Before your first search, read references/search-strategies.md, then write down:

  • The topic or domain you need skills for
  • Whether you need broad coverage (landscape scan) or targeted matches (specific technique)
  • How the downloaded skills will be used (comparison, evidence, reference)
  • Keywords from at least 3 different angles (technology, domain, activity, pattern, or scope)

Keyword diversity matters more than synonyms. This step determines search quality.

2. Search for skills

Run skill-dl search with 3-20 space-separated keywords:

skill-dl search typescript mcp server sdk patterns --top 20

Output is a markdown table ranked by keyword match count. Higher match count means the skill matched more of your keywords.

Search tuning flags:

Flag Purpose
--top N Cap results to top N (default: all)
--min-match N Only show skills matching N+ keywords

Search backends: Serper API (Google), playbooks.com scraping (always active), Scrapedo proxy (fallback). All run automatically.

If the first search returns too many results, add --min-match 2 or narrow keywords. If results are sparse, broaden keywords or run a second search with different phrasing. Read references/search-strategies.md for multi-round search patterns.

3. Evaluate results and build a URL list

Shortlist 3-10 URLs using this order:

  • Match count first — start with skills matching 2+ keywords. If the best results only match 1-2 keywords, treat the topic as niche and curate from the full list.
  • Scope fit second — reject skills whose names clearly signal the wrong depth, domain, or task, even if they rank highly.
  • Diversity third — prefer distinct repos or workflow angles over near-duplicate skills from one source.
  • Owner/repo last — use familiar authors or repos as a tie-breaker, not as the primary signal.

For batch downloads, save selected URLs to a file (one per line, # comments allowed):

# React component skills
https://playbooks.com/skills/owner/repo/skill-one
https://playbooks.com/skills/owner/repo/skill-two

4. Download skills

Read references/download-patterns.md for the full download reference. Quick patterns:

Default layout: skill-dl writes to <output>/<auto-category>/<owner>--<repo>--<skill>/. Example: ./corpus/pro-and-review/mcollina--skills--typescript-magician/. Use --no-auto-category for a flat <output>/<owner>--<repo>--<skill>/ layout, or -c <name> to force a single category folder.

Single skill:

skill-dl https://playbooks.com/skills/owner/repo/skill-name -o ./corpus

For 2-4 ad hoc downloads, pass multiple URLs directly in one skill-dl command. Use a URL file when you need comments, reuse, or a larger batch.

Batch from file:

skill-dl urls.txt -o ./corpus --no-auto-category -f

Pipe search directly to download:

skill-dl search react hooks testing | grep -oE 'https://playbooks\.com/skills/[^ |]+' | skill-dl - -o ./corpus -f

Dry run first for large batches:

skill-dl urls.txt --dry-run

5. Inspect the corpus

After download, inspect what you got. Read references/corpus-inspection.md for the full inspection protocol.

Default auto-categorization puts SKILL.md one level deeper than flat mode, so the find ... -maxdepth 3 commands below work for both layouts. If tree is unavailable, use find ./corpus -maxdepth 3 -type d | sort for a directory view.

Quick inspection:

# Overview
tree -L 2 ./corpus

# Count skills and references
find ./corpus -name "SKILL.md" -maxdepth 3 | wc -l
find ./corpus -name "SKILL.md" -maxdepth 3 | while read f; do
  d=$(dirname "$f")
  echo "$(basename "$d"): $(find "$d" -type f | wc -l) files"
done

For each high-signal skill:

  1. Read its SKILL.md fully — understand trigger boundary, workflow, decision rules
  2. Tree its references/ directory — see structure and decomposition
  3. Read the 2-3 most relevant reference files
  4. Check scripts/, rules/, or examples/ if present
  5. Note what to inherit and what to avoid

6. Use the end-to-end script (optional)

Only use this if you have the full skills-by-yigitkonur repo checked out locally. From the repo root, you can use skill-research.sh from the build-skills references for one-command search-to-inspect:

bash skills/build-skills/references/skill-research.sh "keyword1,keyword2,keyword3" ./corpus 6

This runs all three phases (discovery, download, inspection) with parallel execution.

Decision rules

  • If you need fewer than 5 skills, download them individually. Batch only when building a corpus.
  • If results exceed 50 rows, use --min-match 2 before manually scanning. For niche topics where max match count is 1-2, manually curate from the full list.
  • If a skill fails to download ([ERR] Not found in repo), run with -v to see available skills in that repo. The skill may use a non-standard directory layout.
  • If you need skills from the same repo, put all URLs in one file — skill-dl clones each repo only once.
  • If the corpus will feed into build-skills, organize notes per skill before switching to that workflow.
  • If skill-dl is unavailable and cannot be installed, search GitHub manually for repositories containing SKILL.md files.

Do this, not that

Do this Not that
Use 3-20 diverse keywords covering different angles Use 1-2 generic keywords and wonder why results are thin
Run --dry-run before large batch downloads Download 100 skills and then realize half are irrelevant
Inspect each skill's SKILL.md and references before deciding Judge skills by name alone
Pipe search to download for quick corpus building Manually copy-paste 30 URLs one at a time
Group URLs by repo in batch files for efficient cloning Download each URL independently and re-clone repos
Use --no-auto-category when you want flat output Fight auto-categorization after the fact

Reference routing

File Read when
references/search-strategies.md Formulating keywords, running multi-round searches, or tuning result quality
references/download-patterns.md Downloading skills (single, batch, pipe, parallel) or troubleshooting download failures
references/corpus-inspection.md Inspecting downloaded skills, assessing quality, or preparing notes for comparison

Guardrails

  • Do not search with fewer than 3 keywords. skill-dl requires a minimum of 3.
  • Do not exceed 20 keywords per search invocation.
  • Do not skip the dry run for batches over 20 URLs.
  • Do not treat downloaded skills as templates to copy. They are evidence for comparison.
  • Do not download without first verifying skill-dl --version.
  • Do not ignore download errors. Check with -v and verify URLs.