Skills bank methodology

Solopreneurs lose hours hunting for the right Claude skill, the right MCP server, or the right agent template. Most lists on the internet are either out of date, copied from someone else's GitHub stars, or pumped to sell a course. We curate the bank, rate each entry on real signals, document setup gotchas, and link straight to the source. The goal: turn a 90-minute scavenger hunt into a 5-minute decision.

Every skill page links to the original repository or vendor doc. We don't host the code, we don't middleman the install — we organize, rate, and link.

A skill makes it into the bank when it clears all four of these gates:

• Real, public, tested. Vapor-ware, internal-only repos, and unverifiable claims are out. If we can't install it ourselves, it doesn't ship.
• Has a verifiable source. A GitHub repo, an Anthropic-listed skill, or a registry like mcp.so. Personal blogs alone don't qualify.
• Solves a clear founder workflow. Skills that ship "developer infra for developers" sit lower in the priority queue than skills a non-engineering founder can use directly.
• Active in last 12 months — or has documented sustained use. A repo with no commits in 14 months and no users we can find gets archived.

Every skill carries one score from 0 to 100. It is computed by an open script (scripts/skills-score.mjs, rubric in docs/SKILLS-SCORING.md) from five parts that add up to 100. No vote, no vendor input, no black box:

• 30 — business usefulness. What you would use it for. Marketing, sales, copywriting, SEO, ads, email, e-commerce and finance score the full 30; research, ops and product 24; data, design and automation 16; developer-only tooling 8. Two business uses add a small bonus.
• 30 — traction. GitHub stars on a log scale (100K stars = 24 points) plus forks (up to 6). A skill that lives inside a collection inherits the collection's stars at 90%.
• 20 — recency. Days since the last commit that touched the skill's folder: under two weeks = 20, under a month = 18, under three months = 14, under six = 10, under a year = 5, older = 0.
• 15 — quality. Valid SKILL.md frontmatter (name + description), a README, a license, tests or examples, a known one-line install command, an official vendor account.
• 5 — trending. Stars gained since our previous scrape, or star velocity (stars per day since the repo was created) for repos we index for the first time.

The Trending flag is separate from the score: it lights up when a repo is young and gaining more than five stars a day, or has added 500+ stars (or doubled) since our last pass, and has a commit in the last 90 days. Archived upstreams are capped at 20 and hidden. Scores are recomputed from live GitHub data at every ingest, so a skill that stops shipping slides down on its own.

Each skill gets up to four tags from a fixed vocabulary — marketing, sales, copywriting, SEO, ads, email, e-commerce, research, ops, finance, product, dev, data, design, automation — assigned by keyword evidence in its name, description, folder path, repo topics and the search terms that surfaced it. Tags are what the "Use for" filter reads, and what feeds the business part of the score. "Dev" is dropped when there is strong business evidence, so a marketing skill written in TypeScript reads as marketing, not developer.

Where we scan, at every ingest:

• GitHub repository search — the claude-skills, agent-skills, claude-code and mcp-server topics, "claude skills" / "agent skills" / "cowork skills" free text, and ~40 business terms (marketing, sales, SEO, ads, email, e-commerce, finance…) crossed with those topics, plus "created in the last 60 days and already popular".
• GitHub code search for SKILL.md files mentioning those business terms — the way to find one good skill inside a 200-skill collection.
• Collections — every repo that looks like a skills collection is walked file by file; each SKILL.md becomes its own entry with its own last-commit date and install command.
• github.com/anthropics — the official Anthropic skills repo.
• awesome-mcp-servers and the top "awesome claude skills" lists.
• mcp.so — curated MCP registry.
• Twitter/X tracking on @AnthropicAI and @claudeai announcement threads.
• Skill spotlight requests from the Synapse Circle community.

When sources disagree (e.g., a registry lists a deprecated skill the upstream repo just archived), we trust the upstream repo and flag the registry mismatch in the skill page itself.

How the bank stays current:

• Daily: the scraper watches the official Anthropic repo, awesome-mcp-servers and mcp.so for new entries.
• At every ingest: stars, forks, last commit and score are refreshed for every GitHub-backed entry; the research pass adds new candidates that clear the score floor.
• Draft rule: an entry whose upstream repo is deleted, archived, or whose skill folder no longer exists is marked draft with the reason and stops rendering. Files are never deleted, so a restored repo comes back on the next ingest.
• Noindex rule: skills scoring under 75 keep their page but are noindexed and left out of the sitemap.
• Manual review: Maxime hand-checks the top 50 by score for accuracy of install command and use case.

The negative space matters as much as the positive. We refuse to:

• Pay for placement. No vendor pays to be listed, featured, or boosted in the rankings.
• List skills we can't verify. If we can't install it or read the source, it doesn't ship.
• Fabricate use cases. Use cases are pulled from the README, our own runs, or named operators. We don't invent customer stories.
• Index abandoned or broken skills. See the auto-deprecation rule above. Stale skills stay accessible but leave the search index.
• Inflate the count. Splitting one skill into 4 variants to pump the total is forbidden. One canonical entry per skill.

Our editorial principles are documented in full on the editorial policy page; this section is the skills-specific extension.

Built something worth listing? Found something we missed? Email max@crank11.co with the subject "Skill submission" and include:

• The public source URL (GitHub, official doc, or registry).
• One paragraph on what it does specifically.
• One paragraph on the founder workflow it solves.

We review submissions weekly and respond within 7 days — including when the answer is no, with the reason. A dedicated submission form is on the roadmap; the mailto is the canonical path today.