Read-Only vs. Implementer: The Two Tribes of SEO Skills, Scored

We run every skill in the directory through a six-dimension review — trigger, structure, workflow, content, engineering, security, 0–10 each. Somewhere around the fortieth SEO review, a split showed up that the composite scores don't capture: some skills will never touch your site, and some exist to rewrite it.

The split is architectural, not cosmetic

Ask one question of any SEO skill: what exists after it finishes running?

For one tribe, the answer is a file. A report, a prioritized list, a scored audit. The site itself is untouched. agentic-seo-skill is the purest example we've scored — 16 sub-skills and 89 scripts, all of it strictly read-only, and the skill says so explicitly rather than leaving it for you to discover. It crawls, it measures, it writes findings. That's the contract.

For the other tribe, the answer is a diff. fire-your-seo-agency, an 8.1 on our scale, is built to act on its own findings: once you approve the plan, it edits the codebase directly. And geo-optimizer-skill goes further — it has an --apply mode that batch-rewrites content files across a site.

These aren't two quality levels of the same thing. They're two different products that happen to share a keyword.

The asymmetry nobody prices in

Here's the argument I want to make: the failure cost of these two tribes is not comparable, and most people choosing between them treat it as if it were.

When a read-only auditor is wrong, you lose tokens and time. agentic-seo-skill misjudges a canonical tag, the report tells you to fix something that was fine, you spend an hour chasing it, you get annoyed. The blast radius is one afternoon and some credibility.

When an implementer is wrong, the blast radius is the site. The failure mode I keep coming back to is 301 consolidation done in the wrong direction. An implementer skill decides that URL A should redirect to URL B, when the link equity actually lived on A. It applies the change sitewide. Now you've pointed your strongest pages at a weaker target, and the damage is quiet — rankings don't fall that day, they erode over weeks, which means you don't even get a clean rollback moment. You get an archaeology project.

Batch content rewrites have the same shape. --apply across a content directory means one miscalibrated template decision propagates into hundreds of files at once. Reversible in git, sure — if you notice, and if you knew what "correct" looked like before the skill rewrote it.

The interesting middle: what discipline looks like

What separates a good implementer from a dangerous one isn't intelligence, it's gating. fire-your-seo-agency scored 8.1 partly on process, not cleverness: it refuses black-hat tactics outright instead of negotiating about them, and it stamps re-test dates on its changes — treating every edit as a hypothesis that has to be re-validated later, not a completed action. That re-test discipline is the closest thing I've seen to an implementer acknowledging the asymmetry in its own design.

The read-only tribe has its own discipline to maintain. technical-seo-checker, our highest-scored SEO skill at 9.3, earns a lot of that on workflow: findings you can hand to a human who decides what to do with them. The moment an auditor starts auto-applying its own recommendations, it migrates tribes and inherits the other tribe's risk profile — whether its security design followed or not.

Choose by blast radius, not by score

Our scores are deliberately comparable across both tribes — a 7.8 read-only skill and an 8.1 implementer sit near each other on the leaderboard. That's correct for measuring skill quality and misleading for choosing one.

The decision rule I'd actually use: run auditors on any site, any time, cheaply and often. Graduate to an implementer only when the site's failure modes are ones you can detect quickly and roll back cleanly — when you have traffic monitoring that would catch a bad 301 consolidation inside a week, and a git history that makes the rollback boring. If you can't describe how you'd notice the implementer being wrong, you don't have an implementer problem, you have a detection problem, and no skill score fixes that.

geo-optimizer-skill at 6.7 overall is a cautionary datapoint here: it has the --apply ambition, but it depends on external CLIs it never declares installing — meaning the blast radius of a batch rewrite extends into dependencies you didn't agree to. The pattern holds across the directory — the appetite to change your site is common; the gating that makes it safe is rare. Buy the gating, not the appetite.