A Claude Skill That Filters Out the Popularisers

Every field has two populations of expert: the people who do the work, and the people who explain it. The second group is vastly larger, vastly more visible, and optimised for something other than being right.

I wrote about the filter I use to tell them apart. This is the automated version — a Claude skill that runs the filter without being asked, so I stop having to retype the calibration in every research conversation.

#What a skill actually is

A skill is a markdown file with frontmatter. The description field is the part that matters: Claude reads it to decide when the skill is relevant, and loads the body only then. So a well-written description is a trigger condition, not a summary.

Mine fires on: "who should I read on X", "is this person credible", and — the useful one — proactively, whenever a popular figure in health, fitness, nutrition, longevity, finance or psychology is about to get cited. Check the tier before citing, not after.

#Installing it

Claude Code — save it as ~/.claude/skills/source-finder/SKILL.md. That's it. Available in every session on that machine, across every project. Restart isn't needed.

Claude.ai and Cowork — Settings → Capabilities → Skills → upload the same file. It syncs to your account, so it works on web, desktop and mobile.

A project or team — put it in .claude/skills/source-finder/SKILL.md inside the repo and commit it. Everyone who clones gets it.

The same file works in all three. There's no build step and no packaging.

#The file

Copy everything below into SKILL.md:

---
name: source-finder
description: Find and rank the most credible sources in ANY domain — researchers and verifiable operators over popularisers, course-sellers and SEO farms. Use whenever the user asks who to read/learn from/follow/trust on a topic, asks for "the best sources / experts / people" in a field, asks whether a named source is credible, or is about to make a decision resting on a claim whose provenance has not been checked. Also use proactively when about to cite a popular figure in health, fitness, nutrition, longevity, finance, business, psychology or supplements — check the tier before citing.
---

# SOURCE FINDER

Rank sources by **whose claims can be trusted, and for what** — never by popularity, follower count, or production quality. The goal is to surface the people the popularisers are quietly reading.

The canonical calibration: asked about hypertrophy, the right answer is **Chris Beardsley** — not the largest "science-based" YouTube channel. If the output resembles a list of the biggest channels in a space, it is wrong; restart.

## The six tiers

| Tier | Definition | What it can support |
|---|---|---|
| **1 — Primary** | Peer-reviewed literature, preregistered trials, regulatory filings, the field's own reference reviews | Anything, within its stated limits |
| **2 — Researcher** | A working scientist publishing *in this specific field* who also communicates publicly | Mechanism and interpretation |
| **3 — Researcher-practitioner** | Works directly from primary literature, publishes their own analysis or preprints, engages with evidence rather than citing it decoratively. **The target tier.** | Synthesis, and the "why" |
| **4 — Operator** | Verifiable, numbers-backed track record over a period long enough to include a bad year | What works in practice. Never a mechanism claim |
| **5 — Press / company** | Journalism, aggregators, or anyone with a commercial interest in the conclusion | That an event occurred |
| **6 — Surface** | Content marketing, SEO farms, course funnels, engagement-optimised video | Nothing |

**"Operator" requires disclosed, dated, verifiable figures.** A physique, a lifestyle, a client roster, or a good framework does not qualify. Say so by name when someone fails this.

## Disqualification pass — run on every candidate before including them

Drop anyone who:

1. **Cites without engaging** — names a study, never gives effect size, sample, limitations, or what would falsify it.
2. **Sells the certainty** — confidence outruns the literature, and there is a course, coaching package or supplement line at the end that the confidence was manufacturing demand for.
3. **Has drifted outside their field** — expertise does not transfer. A brilliant molecular biologist opining on nutrition policy is tier 5 in that moment.
4. **Never publicly changes their mind** — a spotless record in a live field means not paying attention, not being right.
5. **Claims operator status without numbers.**
6. **Only ever agrees with their own audience** — audience capture is indistinguishable from expertise from the inside.

## Two overriding calibrations

**"Well-known" is not "weak."** If the strongest evidence points at something unglamorous everyone has heard of — sleep, sunscreen, protein, index funds, walking, flossing — say so plainly and explain the mechanism. Skepticism belongs on evidence quality, never on popularity.

**Depth means rigour, not contrarianism.** Never manufacture a counterintuitive answer to seem sophisticated. Reflexive iconoclasm is the same failure as credulity in better clothes.

## Output format

For each source:
- **Name · tier · the one-line reason for that tier** — specifically what makes them tier 3 rather than tier 5.
- **What they're good for**, and what they cannot support.
- **Where their best material is** — and honestly whether it's paywalled. Never imply you read gated content you could not access.
- **Their blind spot or bias.** Everyone has one; if you cannot name it, you have not read enough of them.
- **Who credibly disagrees**, and on what. A field where every good source agrees is a field being seen poorly.

Then:
- **Disqualified list** — 2–4 prominent names that fail the pass, with the specific failure. These are what the user would otherwise find first, so naming them is as useful as naming the good ones.
- **Gaps** — if no trustworthy communicator exists for a sub-area, say so and point to primary literature. Naming a gap beats nominating a mediocre substitute. (Skincare is the standing example: there genuinely isn't one.)
- **Single best entry point** — one paper, book or article. Prefer the primary source over any summary.

Keep it to five names that survive scrutiny rather than twenty that don't. Never pad to look thorough.

## Cached results

Domains already run, so they need not be re-derived. Re-check annually — sources get captured, drift, or start selling something.

| Domain | Tier 2–3 | Operator | Disqualified |
|---|---|---|---|
| Hypertrophy | Chris Beardsley; Milo Wolf; Pak Androulakis-Korakakis; Brad Schoenfeld; Eric Helms; Greg Nuckols | Keenan Malloy (cites PMIDs); Elijah Mundy; Yo Talks; TJR | Mainstream "science-based" YouTube |
| Sprint & power | Jean-Benoît Morin; Pierre Samozino; Ken Clark; Peter Weyand | Cal Dietz; Chris Korfist; Frans Bosch | — |
| Endurance | Stephen Seiler; Iñigo San Millán; George Brooks | — | Most Zone-2 podcast content |
| Mobility | Primary stretch-tolerance literature | Matthew Smith (Olympic S&C) | Generic follow-along routines |
| Longevity | Matt Kaeberlein (Optispan); Morgan Levine; Nir Barzilai; Andrew Steele | Fight Aging!; Lifespan.io | gethealthspan.com (telehealth content marketing) |
| Hair | Rodney Sinclair | — | Compounding-pharmacy content |
| Skin | **Gap.** Primary only: Pinnell/Duke (vitamin C); Tanno 2000 (niacinamide); Voorhees/Fisher (retinoids) | — | Nearly all "peptide guide" content |
| Relationships | R. Chris Fraley; Paul Eastwick; Eli Finkel; Harry Reis | — | The entire pickup genre |
| Wealth / ecommerce | — | Baremetrics Open Benchmarks; Pieter Levels; Andrew Youderian; Patrick McKenzie | Course-sellers; framework-publishers who never disclose figures |
| Supplements | Examine.com | — | Every brand-owned "research" blog |
| Philosophy | Stanford Encyclopedia of Philosophy → then the primary paper | — | Summary/explainer channels |

## When checking a single named source

Give the tier, the reason, what they can and cannot support, their blind spot, and one better source at a higher tier if one exists. Be fair — many popular sources are genuinely good at tier 4 and only fail when treated as tier 2. The failure is usually the reader's framing, not the source's existence.

#Why the disqualification pass is the important half

Ranking sources is easy and mostly useless. The filter earns its keep by dropping candidates:

• Cites without engaging — a study named, effect size never given, limitations never mentioned. The most common failure and the hardest to spot, because it looks exactly like rigour.
• Sells the certainty — confidence outruns the literature, and something is for sale at the end that the confidence was manufacturing demand for.
• Drifts outside their field — expertise doesn't transfer.
• Never publicly changes their mind — a spotless record in a live field means not paying attention.
• Claims operator status without numbers — a physique or a framework isn't a disclosed, dated record.
• Only agrees with their own audience — audience capture is indistinguishable from expertise from the inside.

#The two calibrations that stop it going wrong

A filter that penalises popularity will, left alone, start selecting for obscurity — because contrarian answers feel like the product of rigorous filtering. That's the same error as credulity in better clothes.

So two rules override everything else:

Well-known is not weak. If the strongest evidence points at sleep, sunscreen, protein, index funds or walking, say so plainly and explain the mechanism.

Depth is rigour, not contrarianism. Never manufacture a counterintuitive answer to seem sophisticated.

And one more that took me longer to add: name the gaps. Sometimes the honest output is that no trustworthy communicator exists — consumer skincare is the standing example. A filter that always returns a name will eventually return a bad one.

#An honest note on judging people versus artefacts

The first version of this skill had a table listing "mainstream science-based YouTube" as disqualified. I've since changed it, because it was the wrong shape of claim.

The failure mode — citation without engagement — is a property of individual pieces, not of everyone producing them. Several popular educators in that space also co-author peer-reviewed reviews and meta-analyses, and those specific outputs sit at primary tier regardless of what their channel looks like.

Judge the artefact. A blanket dismissal by name is the same lazy move as blanket trust, and it's more embarrassing because it feels like discernment.

---

The full source-tier rubric, applied per domain, is in the protocol at /protocol/sources. The essay behind the method is Finding the Real Experts.