Skillselion Deep Dives
How we write a deep dive
A deep dive reads one skill or MCP server the way a reviewer reads a contract: the actual file, top to bottom, quoted section by section. Every quote is checked against the origin on the day we publish and links to the exact commit and lines it came from; a skill with no public repository (some vendors serve the canonical file from their own domain) is cited by that file and the date we verified it. This page is the standing method. Individual dives add only what belongs to that skill.
What a deep dive is, and is not
Most writing about agent skills paraphrases a README. A deep dive quotes the file that actually runs, because the instructions inside it are the product. When a skill tells your agent to install packages without a second confirmation, or to trust some publishers before others, that sentence is in the file and nowhere in the marketing.
- It is a source reading. The skeleton of every dive is the source file itself, walked in its own order, so a reader can follow along with the file open.
- It is not a review score. We publish no ratings and no rankings inside a dive. Adoption numbers live on the listing, sourced and dated.
- It is not a launch post. Fresh releases belong in Skill of the Day. A dive covers work that many people already run.
How we pick the subject
Adoption first: we work down the most-installed skills and MCP servers in the catalog, because the files a million agents already load deserve to be read carefully. Search demand can promote a subject out of order when readers are clearly already looking for it. Every pick is deliberate and hand-curated, never automated.
The freshness gate
Before a word is written, we verify that the copy we are about to quote matches the live original. The catalog copy is refreshed from the registry, then compared as full text against the origin repository's default branch. When the two disagree, the repository wins and the catalog gets fixed.
This is not a formality. In August 2026 the gate caught a skill whose published file had been completely rewritten upstream while catalogs still served the old version, and a second skill that had quietly dropped a command from its documented interface. Both facts became part of the published dives.
Quotes, and the pinned-commit rule
Quotes are verbatim, including punctuation we would never write ourselves. Each one names its file and section, and links to that exact commit and line range on GitHub. We do not link the default branch: the file moves, the line numbers stop matching, and a citation that used to be true becomes quietly wrong. The pinned link keeps working after the source changes, which is the whole point of a citation.
When we join two separated fragments of one passage we mark the gap. We never merge quoted text with punctuation the original does not contain.
Independence
Deep dives are never sponsored and never sold. Authors do not see a draft before it publishes and have no veto after. We name the sharp edges we find, including install flags that skip confirmation prompts, discovery rules that route readers toward the publisher's own catalog, and repackaged clones of a popular skill shipping under someone else's account.
Skillselion earns money from partner placements elsewhere on the site. No placement buys, changes, or blocks a dive.
Numbers
Install counts and stars in a dive are the same live figures the listing shows, rendered from one canonical source with the date they were synced. We do not freeze a count into prose, because a number written by hand drifts away from the page around it and readers end up choosing which of two figures to believe.
Frozen, dated statistics belong to Skillselion Research, which has its own methodology and its own freeze law.
Corrections and updates
Each dive carries the date its quotes were last verified against the source. When a file is materially rewritten upstream we revisit the piece, refresh the reading, and stamp a new verification date. Published quotes stay reachable through their pinned links either way. If we get something wrong, we fix it on the same URL and say what changed.
Deep dive methodology FAQ
How do you choose which skill or MCP server gets a deep dive?
By adoption first, then demand. We work down the most-installed skills in the catalog, and promote a skill out of order when search demand or a trending spike says readers are already looking for it. A deep dive covers established, widely installed work; a brand-new launch belongs in Skill of the Day, not here.
Do you install and run the skills you write about?
We read every line of the source and quote it verbatim. Where a claim is testable in our own workflow, we test it and say so in the piece. We never describe behavior we have not either read in the file or observed ourselves.
Why does each quote link to a commit instead of the main branch?
Because a main-branch link drifts. The file changes upstream and the line numbers stop matching the quote, so the citation quietly becomes wrong. Every quote links to the exact commit we verified it against, with the line range highlighted, which stays true forever.
Are deep dives sponsored, and do authors get a say?
No. Deep dives are never sponsored, sold, or shown to the skill author before publishing. We say what the file does, including the parts its author would rather we skipped, such as install commands that skip confirmation prompts or discovery rules that favor the publisher.
What happens when the source file changes after we publish?
The dive keeps its published quotes, which stay verifiable through their pinned commit links, and we revisit the piece against the new version. A material rewrite upstream earns a dated update on the same page, and the change itself usually becomes part of the story.
Read the dives
Every published deep dive is listed on the deep dives index, and each one renders on the listing page for the skill it reads. Questions about the method, or a skill you want read closely, reach us at info@skillselion.com.