CODEZ ORIGINAL GUIDE / 01
How to choose an Agent Skill: from stars to a practical trial
A ranking helps you narrow the search. Choosing a skill depends on whether it can complete a specific task in your environment. This guide offers a process you can reuse.
Sources checked:
Format and project facts come from the linked original sources; selection steps and acceptance criteria are CODEZ editorial advice. The PDF candidates are real catalog entries compared at fixed commits. These third-party skills were not executed, and the sample is not a measured result.
01 / Define the capability you need
An Agent Skill is generally a folder containing SKILL.md metadata and task instructions, with optional scripts, references, and templates. The format includes an optional compatibility field for environment requirements. If it is absent, check the instructions and dependencies anyway.
Write down the input, output, and acceptance criteria. For example: extract a table from a PDF with selectable text into CSV while preserving column names. “Help me with documents” is too broad to assess a candidate.
02 / Interpret repository stars
GitHub stars help people save projects and express appreciation for maintainers. We use them as a signal of attention when discovering projects.
CODEZ ranks repositories by their stars. Attention to a repository with hundreds of skills may come from only part of its content. Its total cannot be assigned to each individual skill. Compare task fit first, then use popularity as additional context.
An update date helps you locate recent material, but an unrelated commit does not prove that a capability you need has improved. Check changes, examples, and issues for the relevant directory.
03 / Turn your requirements into checks
Specific checks make a decision easier to explain and revisit than popularity alone. You do not need to compress them into a seemingly precise overall score.
| Use case | Check first | Observe during a trial |
|---|---|---|
| Documents and spreadsheets | Input type, output format, system dependencies | Lost columns, encoding, layout, or data |
| Code and development workflows | Language, framework version, required tools | Scope of edits, errors, existing project checks |
| Research and source organization | Source scope, citation needs, network access | Traceable references and claims supported by evidence |
04 / Read the source and inspect its requirements
Open SKILL.md from the catalog, then read the README in the same directory. A Chinese README can make instructions easier to understand, but translated documentation does not establish task fit. Confirm critical requirements in the available original text.
Before a trial, record the following:
- Environment: supported clients, operating systems, command-line tools, and stated version requirements.
- Scope: files it reads or changes, external services it calls, and credentials it needs.
- Resources: whether referenced scripts and materials exist, and what the commands do.
- Usage conditions: read the project and skill directory license files. CODEZ labels such as “Not specified” or “Needs review” indicate unconfirmed information.
- Reproducibility: save the repository URL, skill path, and commit used for the trial so you can find the same material later.
05 / Compare candidates on the same small task
Choose two or three candidates closest to your task. Use the same non-sensitive input in a recoverable test copy, initially enabling only the capability you need. Follow the installation instructions for your actual client.
Set acceptance criteria before you begin. For table extraction, check row counts, column names, and empty values. For code changes, check the intended feature, changed files, and existing project checks. Record errors and environment details before attributing a failure to the model or skill.
Track manual corrections, extra tools, external service costs, and elapsed time alongside the output. Consistent records across repeated tasks make a decision easier than an impression from one run.
- Candidate and version: repository, skill path, commit.
- Input and environment: sample, operating system, client, dependencies.
- Result and corrections: criteria met and work still needed.
- Resources and decision: services, time, costs, and reasons to adopt or defer.
06 / Real candidates: choosing between two skills named pdf
CODEZ includes skills/pdf/SKILL.md from anthropics/skills and skills/.curated/pdf/SKILL.md from openai/skills. Both are named pdf. Record the repository and full path to distinguish their instructions. The table compares documented priorities, not mutually exclusive capabilities or accuracy inferred from Stars.
Anthropic supplies examples for text and table extraction, page operations, generation and OCR, with form handling referred to FORMS.md. OpenAI emphasizes reading, creating and layout review: render pages, inspect them and check again after meaningful changes. Extracted text alone does not establish layout fidelity.
| Candidate and full path | Documented emphasis | CODEZ suggests checking first | Still needs verification |
|---|---|---|---|
| anthropics/skills / skills/pdf/SKILL.md | Operation examples: extraction, page handling, generation and OCR | For tables or scans, check the relevant operation and dependencies | Table structure, OCR language and accuracy, referenced materials and available tools |
| openai/skills / skills/.curated/pdf/SKILL.md | Creation and review with rendered-page checks | For a readable report, check rendering and inspection steps | Fonts, tables, page numbers, breaks and rendering in your environment |
07 / Turn successful extraction into verifiable checks
Here is a suggested trial design: compare candidates on the same two-page, non-sensitive inventory list. Include item ID 0017, Chinese headers, a blank description, a repeated header on page two and a merged cell. Manually check the expected result first, retaining source page numbers and the original copy. This test has not been run.
Check whether text can be selected and try a small extraction first. These checks are clues: a text layer does not ensure correct table structure, and OCR text may contain errors. If OCR is needed, record its language and system tools. Record missing dependencies as environment issues.
- Record the candidate commit, client, model, tool versions and input type so environment differences are not mistaken for skill differences.
- Record corrected rows and reasons, elapsed time and services actually used; leave unmeasured fields unfilled.
- Accept only conclusions supported by the listed checks. A passing sample still leaves your actual inputs to verify.
| Check | Result to verify | Action when it fails |
|---|---|---|
| Item IDs | 0017 retains its leading zeros instead of becoming 17 | Specify field types, extract again and check each item |
| Tables across pages | Repeated headers are excluded; records are neither lost nor duplicated | Keep page numbers and reconcile per-page rows before combining |
| Chinese and blanks | Readable headers and blank descriptions without invented values | Check encoding and field positions; flag unrecoverable entries |
| Merged cells | Descriptions map to the intended records under a recorded split rule | Retain the source region and decide the mapping manually |
| Report layout | Rendered output has no clipping, overlap or missing glyphs | Fix fonts or layout and inspect the latest rendering |
08 / Check category, licensing and dependencies separately
We corrected the two PDF entries from Design and Development to Documents based on their source text: PDFs are the principal input and output, while Python and rendering tools support the task. A single category helps discovery but cannot decide between extraction and layout review.
Both catalog entries have a separate LICENSE.txt. CODEZ retains its License file needs review label. Read the license in each skill directory; public visibility or a well-known organization does not establish unrestricted use.
Check dependencies for the actual operation. Python packages alone do not establish that system rendering tools or an OCR environment are available. This source comparison does not prove client compatibility. Record missing tools, unverified steps and usage limits in your decision.
09 / Cloudflare Skill: check project configuration and credentials first
On 2026-10-14, this section checked the fixed commit of cloudflare/skills / skills/cf-cli/SKILL.md and official documentation. These are CODEZ selection and trial recommendations; this Skill was not run.
cf does not reuse a Wrangler login. Check the account and token permissions first; record the CLI version, configuration and target resources to avoid mistaking environment differences for Skill differences.
| Use case | Check first | Observe during a trial |
|---|---|---|
| wrangler.jsonc, wrangler.json or wrangler.toml exists, with no cloudflare.config.ts | Keep the Wrangler dev, build and deploy workflow first; cf project commands do not read that existing configuration. | Review the commands and file changes the Skill proposes first; decide on configuration migration separately. |
| cloudflare.config.ts already exists | Check dev, build and deploy against the cf project workflow; the CLI is in beta, and a category does not establish compatibility. | Check Node.js 22.18 or later and the target account; inspect build output and the target Worker in a test copy. |
cloudflare/skills / skills/cf-cli/SKILL.md ↗
Cloudflare: official CLI and Wrangler comparison ↗
Cloudflare: official installation and sign-in requirements ↗