affiliate-skills cannot agree on its own skill count, and its demo table has no rows

Fifty, fifty-two, eight stages, nine rows

The project cannot agree on its own size. The repository description says fifty skills. The readme's tagline says fifty-two AI-powered skills across eight stages. The capability table underneath has nine rows: program search, trending scout, research brief, angle ranker, traffic analyzer, content and blog, landing and distribution, analytics and automation, and compliance and meta. Nine rows against eight stages, reconciled only by the last row being described as operating across all stages while the diagram puts it at a position after automation. One of the rows further claims to be twelve skills on its own, which is how a single table row can carry more than the stage count it sits in. Nothing on the page says how many files are under the skills directory, so the count a reader can verify is the one nobody is given.

The demo's program table has headers, a rule, and no rows

Two transcripts are presented as a demo, and the first one is specific. The agent claims to have scanned 47 videos across two platforms, reports the top format as comparison at 45 percent of top content with average engagement of 35.2, names a hook with an engagement figure, identifies a gap nobody has covered, and sets a benchmark of a median 18 thousand views where the top ten percent needs 85 thousand. The second transcript then asks for the best program and answers with a table:

       Name         Commission        Cookie    Stars   Traffic Score
       ─────────────

That is the whole table. Column headings and a separator rule, and not a single row, which is the exact point at which the page needed either a row or a note saying the rows were omitted. Nothing on the page says whether the figures in the first transcript come from an actual scan.

The no-install path points your agent at someone else's API with a tracking tag

The section headed try it now, no install needed, is four lines of text to paste into any assistant. The first line asks it to search a named affiliate directory for AI video tools. The second hands over the request:

Use this API: GET https://openaffiliate.dev/api/programs?q=AI+video&sort=relevance&limit=5&utm_source=affiliate-skills

A source parameter is appended to every call, which is how the directory attributes traffic from this repository. The third line asks for a table with name, reward value, cookie days and stars, and the fourth asks the assistant to recommend the best one and explain why. So the ranking the toolkit exists to produce comes from a directory operated by someone else, reached over the network, with the call attributed back to this project. The Program Search row in the capability table says the same thing in a shorter form: live program data covering commissions, cookies and comparisons.

Four install routes, and one repository URL stops mid-word

Installation is offered four ways. One command adds the skills through a package for Claude Code and Pi, and the same command is repeated for Cursor and Windsurf. A second route clones the repository straight into a Claude skills directory inside the home folder, changes into it and runs a setup script named with no extension:

git clone https://github.com/Affitor/affiliate-skills.git ~/.claude/skills/affiliate-skills
cd ~/.claude/skills/affiliate-skills && ./setup

A third goes through a different tool's installer for OpenClaw. And the last route at the bottom of the page, for Cursor and Windsurf, begins a clone command whose repository URL stops one letter short of the real name. So the manual route writes into a directory belonging to one agent even when the reader is using another, the setup script is invoked without a word about what it does, and the last block on the page cannot be pasted.

Bun, no dependencies, and a CLI that ships as a compiled binary

The package manifest is thirty lines and has nothing in the dependency fields. The devDependencies object is present and empty, and there is no dependencies object at all, so an npm install on this package installs nothing. Every script needs Bun: the build compiles the CLI entry point into a single binary under the tools directory, the dev script runs it uncompiled, and three test scripts run the unit tests, the registry invariants and the documentation contract checks. A fifth script is declared for the build itself. So the toolchain is one runtime, the dependency surface is zero, and the shipped artefact is a compiled executable rather than a script that runs on an interpreter, which matters for the claim that this works with any agent that reads text.

The skills are markdown with schemas and a next-step field

The repository layout is the most informative part. Skills live at a path of stage, then skill name, then a markdown file, so the flywheel order is in the path rather than in prose. A shared references directory holds the doctrine, the compliance notes and the flywheel documentation that every skill draws on. A registry file is the machine-readable catalog, an evals directory holds evaluation cases and their results, and a tools directory holds the CLI source. What makes chaining work is a field: the page says every skill has typed input and output schemas, a chain metadata object, and a suggested next value, and that every skill knows both what comes after it and what feeds it. That is the mechanism behind the closed loop, and it is checked by two test suites.

Three plan documents, a version file, and a compliance skill

The root of the repository holds its working papers: three plan documents for a content launch, a social intelligence upgrade and a second-generation skills expansion, alongside a changelog, a version file, a spec directory, a drafts directory and a template directory. The version file and the manifest both carry a version, and the repository publishes no tags, so the two have to be kept in step by hand. One row of the capability table is worth singling out: compliance and meta, described as a safety net covering an FTC audit, funnel planning and self-improvement, with the compliance references living in the shared directory. The same loop that writes the posts also contains the step that audits them, which is a sensible arrangement to have in the same package rather than a separate one.

Editorial conclusion

The engineering in this repository is the part worth having. The skills are markdown files with typed input and output schemas, a chain metadata field naming the next skill and what feeds it, a machine-readable catalog, evaluation cases, and two contract test suites that check registry invariants and documentation promises. The build is one Bun compile step producing a single binary, with no runtime dependencies at all. What the front page does not settle is the size of the thing. The repository description says fifty skills, the readme says fifty-two across eight stages, and the capability table has nine rows, one of which bundles twelve content skills. The demo transcripts are labelled as demos and the one program table stops after its column headers. And the program's core recommendation comes from an external directory, reached over HTTP with a tracking parameter, so the ranking you are buying is someone else's. Before adopting it, read the registry rather than the tagline to see what actually ships, and decide whether you want your agent's recommendations to come from a directory you do not control.

Frequently asked questions

Official sources

1. Affitor/affiliate-skills on GitHub
2. Issues
3. License: MIT
4. Project website
5. README