Why Learning Brain exists.

What it actually is

Learning Brain is a plug-in for AI tools. A lightweight service your AI talks to when you ask it for learning-design help. It doesn't replace your AI. It sits alongside, supplying the learning-science expertise that general-purpose models don't have.

When you ask Claude (or Codex, or ChatGPT) to design a course, audit a module, or write assessment items, your AI connects to Learning Brain automatically. What comes back is three things:

• Curated evidence. The specific research relevant to the task, with cited sources and an evidence-strength rating on every claim.
• A quality rubric. The review criteria a senior learning scientist would apply to this kind of work.
• Instructions for the AI to follow. The structure it should use when producing the output.

The AI produces the finished work from that package. Learning Brain makes sure the work is structurally sound.

Thirty-two tools cover the main jobs of instructional design: writing objectives, designing modules and courses, building question banks, scripting facilitator guides, auditing existing work, predicting whether training will transfer to the workplace. Five expert personas organise them: Learning Scientist, Curriculum Architect, Instructional Writer, Delivery Coach, and Course Doctor.

Behind the tools sits a hand-curated knowledge base of 195 research notes across nine domains of learning science, 360+ cited studies, and fourteen quality rubrics modelled on the review criteria a senior learning scientist would bring to a colleague's work.

You keep using the AI the way you already use it. The conversation doesn't change. The quality of what comes back does.

How it pushes back

Most AI tools for L&D produce fluent output that looks evidence-based. Learning Brain is designed to resist that in three specific ways.

1. Evidence only, never invention

Tools cite from the knowledge base or they refuse. The knowledge base is hand-curated: each note is a complete, cited argument with an evidence-strength rating (strong, moderate, weak, or theoretical). When a tool has nothing substantive to say on a topic, it says so rather than falling back to general AI knowledge. Popular myths — learning styles, left/right brain, the 10,000-hour rule, the Cone of Learning, "digital natives" — get flagged the moment they appear.

2. Eighteen specific refusals

Some requests don't get built, regardless of how confidently they're asked for. Ask for a ninety-minute module covering 35 topics and Learning Brain explains why it won't. Ask for training that will transfer to wildly different situations without specific practice and it refuses. "Everyone at the company" as an audience, refused. A cramming schedule dressed up as learning, refused. An auto-graded quiz for a skill only a human can judge, refused. The refusal list started at five and grew as new failure patterns turned up in testing. It now stands at eighteen.

3. Expert-review rubrics

Nineteen of the thirty-two tools carry a rubric the AI is instructed to answer to before returning anything. Aspirational goals get rewritten, not softened. Question banks too narrow for the objective get flagged. Modules missing a demonstration phase get caught. Every rubric includes explicit instructions for the AI: name structural flaws directly, refuse hedge words and compliment sandwiches, keep pushing back when the request is bad.

The first two mechanisms hold regardless of which AI you're using. Rubric compliance is highest on Claude and Codex, which is why those are the recommended clients. ChatGPT chat follows the rubrics less consistently on complex audits. The tools still work there, but the best experience lives in Claude and Codex.

How the tools check themselves

Nineteen of the thirty-two tools come with a built-in review loop. They don't just produce an output and hand it back — they produce a draft, score it against their own rubric, name what's missing or weak, revise, and only then present the result. Every design artefact arrives already reviewed against the criteria a senior learning scientist would apply.

An objectives list isn't just generated — it's checked for specificity, measurability, cognitive level, and alignment, and rewritten where it fell short. A module audit isn't a narrative opinion — it's twelve structural criteria each graded pass/fail with evidence, plus the ten instructional illusions flagged by name. A question bank isn't just a pile of items — it's items checked for construct validity against the underlying objective.

Most AI tools try to do this with a prompt that says "follow these criteria." Models often don't: they interpret criteria generously, or skip them when the context gets long. Learning Brain makes the self-review a structural step in the tool contract, so the model has to answer to it before anything reaches you.

Why a curated knowledge base

Most AI products that claim to know a domain use a simple retrieval step: take a question, search through a pile of stored documents, hand the model the most relevant chunks. That works for tasks where the source material is structured and clean. It doesn't work well for learning science, where the primary material is dense academic papers full of caveats, methodological nuance, and contested findings.

Learning Brain uses a different approach. The knowledge base is 195 hand-curated notes, each written as a complete argument: a defensible claim, the research behind it, the conditions under which it holds, and the design implications. Each note cites primary sources. Each carries an evidence-strength rating. Each links to related notes so the tools can reason across topics. When a course-design question touches cognitive load, the retrieval-practice notes and the multimedia notes are one link away.

The result: the tools don't just have information about learning science. They have a practitioner's opinionated, evidence-graded structure over it. That's hard to replicate by scraping the literature.

How it was tested

Three kinds of testing, all still running.

Testing designed to fool it

The five highest-risk tools — the ones that audit work or push back on briefs — each faced five inputs specifically crafted to trip them up. Polished modules with no demonstration phase. Objectives that sound active but promise nothing. MCQs where the longest answer is always correct. Engagement-first workshop briefs from sponsors who'd already signed the budget. Two different AI models ran the tools against each input. The result: zero misleading outputs across all fifty tests. Every flaw was caught. Every fix was specific, not hedged.

A structural test suite

All thirty-two tools run against a 311-case test suite covering input handling, refusal conditions, citation discipline, and output structure. The full suite passes at 100%. Every code change runs against it before deploying.

Ongoing quality judging

A separate test uses an independent AI judge to score tool outputs across five dimensions: directness (does it push back, or hedge?), rubric compliance, citation accuracy, honesty about what the knowledge base covers, and professionalism. This catches regressions and identifies which rubric instructions the calling models follow reliably versus inconsistently.

When the quality judge finds a pattern that slips through, tuning the rubric isn't enough — a specific refusal gets added. That's where the eighteen refusals came from. They're guaranteed: they don't depend on the AI deciding to catch a bad request. They catch it before the AI sees it.

Known gaps

There's no longitudinal study showing that courses designed with Learning Brain produce measurably better learning outcomes than courses designed without it. The structural differences are clear in head-to-head comparisons. The direct causal chain from "used this tool" to "learners retained more" takes time and access to real organisations to measure.

The knowledge base was curated by one practitioner. It's principled and comprehensive, but it's one perspective. The strongest domains are cognitive load, retrieval practice, multimedia design, and instructional design frameworks. Adjacent topics — learning analytics, xAPI, compliance-specific regulations — are thinner. When the knowledge base is silent on something, the tools say so rather than bluffing.

Who built it

Built by Laurie Harrison. Twenty-plus years designing learning experiences across educational publishing, corporate L&D, and edtech.

Learning Brain began as a personal knowledge base. I started it in 2020, during covid. I came across the concept of Zettelkasten and started building a library of notes on the research behind every design decision I was making: the studies I kept citing, the principles I kept returning to, the findings that had actually shifted my practice. Frankly, I'd spent years making learning-design decisions I couldn't fully back up, and this was my way of trying to fix that. When AI tools started producing confident, fluent-sounding, but vacuous learning-design output, I connected the notes to them to see what would happen. Learning Brain is what came out. And it worked so well that I decided to turn it into a tool and make it freely available. Since the “thinking” is done by the end user’s LLM, it costs almost nothing to run.

Learning Brain is an independent personal project. The knowledge base is my own synthesis of published learning-science research. Not affiliated with any organisation. Not intended as a commercial product. That’s why it’s free for you to use.