A skill is a markdown file with frontmatter. The description decides when it loads, so it is written as a trigger condition rather than a summary. The same file works in Claude Code (~/.claude/skills/), on claude.ai under Settings → Capabilities, and committed into a repo for a whole team.
These are the ones I wrote. All free, all copy-paste, no packaging step.
source-finder
Mine
Returns the researchers, not the biggest channels.
When it fires
Any 'who should I read on X', any credibility check on a named person, and proactively before citing a popular figure in health, fitness, nutrition, longevity, finance or psychology.
Why it exists
Search ranks by reach, and reach correlates with production quality far more than with accuracy. This ranks by what kind of claim a source can actually support — six tiers, plus a disqualification pass that does most of the work.
Drafting or editing any prose a person will read, and whenever asked to make text sound less generic or less like AI wrote it.
Why it exists
LLM prose is recognisable because it is even — same sentence length, same rhetorical move every paragraph, a hedge before every claim. Human writing is lumpy. This removes the evenness and adds back the specificity. Explicitly not for disguising authorship of submitted work.
Course content that matches the evidence on how people learn.
When it fires
Writing a lesson, designing a quiz or exam, building review scheduling, choosing a progress metric, or reviewing existing course material.
Why it exists
The techniques that work feel worse than the ones that don't, so anyone building a course from intuition builds the wrong thing — and every competitor's marketing cites a learning pyramid whose source study has never been produced. This is the lesson shape, the question rules, the scheduling model, and an explicit do-not-build list, with the evidence behind each one.
How to build a paid course on this platform without reinventing what already exists.
When it fires
Building or rebuilding a course for any subject, porting a course from a stale or independent branch, or adding assessment, certificate, practice-bank, or spaced-repetition infrastructure to one that already exists.
Why it exists
The failure mode this exists to prevent: a course build that reinvents commerce or auth the platform already has, spends its whole budget on one subject before validating the lesson pattern on a single pilot, or treats retrieval practice as prose pasted into markdown when the platform needs typed data to grade against it. Distilled from rebuilding Meridian end to end in one long session, including a production-deploy crisis root-caused live against the database and a copyright flag on an official question bank handled explicitly rather than skipped.
When a workflow of AI agents beats one agent, and the failure modes that show up once it does.
When it fires
Writing or reviewing a multi-agent workflow script, deciding whether a task needs multiple agents at all, an agent silently failing on a content-heavy task, or coordinating background workflows that might touch the same files.
Why it exists
Built from 23 real workflow runs against this codebase, kept and indexed rather than left to age out of a session transcript. Covers why splitting research from implementation beat every combined pass tried, why adversarial council framing catches false findings a single grader misses, the exact two-part fix for a silent output-token-ceiling failure that took three attempts to diagnose correctly, and why a fixed agent count is wrong for a search space of unknown size.
The exhaustive version of course-builder — every reusable prompt, the full rubric, and the research base, for a course meant to be the best in its category.
When it fires
Building a flagship or premium course relaunch, a build whose entire pitch is provenance or rigor, or any course work explicitly asked to be the best version rather than a correct-and-shippable one — course-builder stays the fast default for routine course sessions.
Why it exists
course-builder is eighteen rules and a process — enough to run a normal build without reinventing commerce or skipping the pilot step. This is what's underneath it: 27 reusable, bracket-templated prompts mined from real Meridian and Veridian builds, a reconstructed quality rubric with unconfirmed items explicitly marked, nine course-building-specific agent-orchestration patterns, and a full competitor/UX/accessibility research base — one standalone document instead of three cross-referenced ones.
Turns any Claude chat into an on-demand SAT tutor, marked against a named misconception library instead of a generic trap list.
When it fires
Reviewing a missed SAT practice question, wanting an explanation that also teaches the general mechanism rather than just the one question, or running a drill on a specific recurring misconception from an error log.
Why it exists
Rebuilt after six of my own practice tests showed an identical 45/54 Reading & Writing raw score scoring 540 on one and 700 on the other — proof, not theory, that Module 1 routes the ceiling and is worth roughly 3x Module 2. The mark-and-explain loop now runs simple-answer-first, then the system behind it, so a miss teaches the whole category instead of just the one question, and every named trap maps to an actual lesson and a five-second test instead of a generic label. v3 wraps that loop in six modes (TEACH, MARK, DRILL, PLAN, SIMULATE, COACH) and stops at a checkpoint after every explanation — another question, a simpler pass, or a standalone guide plus a link to the full lesson — instead of assuming one pass fixed it.