Education Agent Skills (Gareth Manning) — source summary

A 165-skill pedagogy skill pack on the agentskills.io standard, by Gareth Manning (CC BY-SA 4.0; 545 stars, 91 forks; created 2026-03-01, last pushed 2026-08-04). Delivered via Telegram, routed 2026-08-05. As with cybersecurity-skills and agentic-seo-skill, the domain is not why it is here — but unlike those, this one changes something about the form.

What it is

165 SKILL.md files across 20 domains, counted from the repository tree rather than the README: original-frameworks 17, ai-learning-science 14, curriculum-assessment 13, student-learning 13, wellbeing-motivation-agency 12, down to inclusive-design 3. It ships as a cross-vendor plugin — .claude-plugin/, .codex-plugin/ and .agents/plugins/ marketplace manifests, a CI validation workflow, and a hosted MCP server with OAuth (mcp-server/) so Claude.ai can reach the skills without a local install. Targets named in the README: Claude Code, Claude.ai via MCP, OpenAI Codex, and Hermes.

Why it's interesting here

Skills that cite their evidence. This is the new thing. Each skill's frontmatter carries evidence_strength and a list of evidence_sources naming real literature. From adaptive-hint-sequence-designer:

evidence_strength: "strong"
evidence_sources:
  - "VanLehn (2011) — The relative effectiveness of human tutoring, intelligent tutoring
     systems, and other tutoring systems (meta-analysis, effect size 0.76)"
  - "Shute (2008) — Focus on formative feedback"

Every other pack the wiki holds asserts its own quality and stops there — the benchmark gap named in Open questions. This does not close that gap: nobody has measured whether the skill works, and a citation behind the technique is not evidence for the prompt that encodes it. Corrected 2026-09-05: the first half is no longer true. skillsbench measures whether skills work — 87 tasks, matched no-skill and curated-skill runs over 18 model-harness configurations, average pass rate 33.9% → 50.5% — and it was in this spoke, unread, when the sentence above was written on 2026-08-05. The second half stands and is now the whole point: SkillsBench measures curated skills in general, and nobody has measured this pack, whose citations are still evidence for the technique rather than for the prompt that encodes it. SkillsBench's own finding sharpens that — self-generated skills gain nothing, so which pack you load is exactly what decides the outcome, and it is exactly what nobody grades. What it does is make the claim auditable. A reader can check whether VanLehn says what the skill says, which is more than pm-skills or designpowers offer, and it is the first pack to put a falsifiable provenance claim in the machine-readable layer rather than the README.

Typed schemas, independently arrived at. Skills declare input_schema / output_schema with per-field types and descriptions — the same skill-as-typed-function move as pm-skills-arsenal, from an unrelated author in an unrelated vertical. Two independent instances make it a pattern in the standard's use rather than one team's idiosyncrasy, and both pair it with an MCP server that enumerates and invokes skills.

A vertical that points at the end user, not the professional. Every previous vertical serves the practitioner: PM, designer, marketer, security analyst. Domain 20 (student-learning, 13 skills) is written for live student interactions during study sessions — the person the professional's work is about. The pack is aimed at teachers, curriculum designers, school leaders and EdTech builders, and then a fifth of it addresses the student directly.

Caveats

• The evidence framing is the author's own and unaudited. "Evidence-grounded" is a claim about the cited research, not about the skills, and no third party has checked either link.
• GitHub reads the licence as unrecognized (NOASSERTION); the LICENSE file is Creative Commons Attribution-ShareAlike 4.0, © 2026 Gareth Manning.
• Authority is self-described — 20 years of international-school teaching across 27 countries, per the README, with nothing else corroborating it.
• Star count is a 2026-08-05 snapshot on a repository five months old.

Tier

T2 — first-party and self-interested, but the structure is verifiable: skill count, domain distribution, frontmatter fields and the plugin/MCP layout were all read from the repository rather than taken from the README. The pedagogical claims are not verified here.

Related

agentskills-spec · agent-skills · pm-skills-arsenal · cybersecurity-skills · designpowers · hermes-agent · model-context-protocol · gareth-manning