How this framework was built

Before this project began, we had been building a separate piece of infrastructure: an open-source library of skill files for using Claude in education. A skill file is a small markdown document that tells Claude how to do a specific kind of task well — how to author a Know/Understand/Do chart, how to design a single-point rubric, how to run a curriculum coverage audit, how to translate a learning target into developmental band statements. Skills encode patterns we’d worked out through trial and error so that they don’t have to be rediscovered every time.

The library matters here because the wellbeing framework was built using these skills. The KUD chart authoring skill, the rubric criterion generator, the framework crosswalk skill — all of them existed before this project started and were used, refined, and extended through it. Anyone using Claude in education can start with the same conventions rather than reinventing them. The repository is at github.com/GarethManning/claude-education-skills.

Working with Claude isn’t a tool relationship and it isn’t a research-assistant relationship. The honest description matters because the result on this site depends on it being done well.

Gareth brings the vision, the goals, and the responsibility. What this curriculum should do, who it’s for, what counts as success. The lived knowledge of what twelve-year-olds actually do in a classroom, what a wellbeing PLC argues about, what teachers will pick up and what they’ll abandon. The editorial judgment about whether a piece of writing sounds true. The contextual understanding of what REAL is, where it sits in the landscape of progressive schools, what its families and faculty value. The relational responsibility — being the one who looks the children in the eye, carries the consequences, defends the work to colleagues. The decisions about scope, direction, and when to stop.

Claude brings the building. Drafting at speed, comparing frameworks, running panel reviews against named scholars before content lands, holding the schema in working memory across an entire repository. Reading the literature at scale — the willingness to read three hundred pages of meta-analyses before answering a question. Gareth reads literature too, mostly specific articles he finds and the research syntheses Claude produces, but the breadth of what Claude reads is one of the working asymmetries.

Both of us are editors. Both push back. Claude routinely tells Gareth a draft isn’t ready, an idea is wrong, a scope is too large. Gareth follows the pushback most of the time and ignores it some of the time. Some disagreements are resolved by Gareth being right, some by Claude being right, some by both of us ending up somewhere neither of us started. Claude also reminds Gareth, with some frequency, to stop working at 8:30pm Budapest time. Gareth (usually) ignores this.

The infrastructure matters because trust between human and AI can’t rest on either side’s good intentions. Some of what we do came from painful and costly mistakes — a failed build that overwrote work, a research synthesis that turned out to invent citations, a content draft that didn’t match the actual schema and would have rendered wrong on a live page. Each mistake produced a discipline. Claude makes things up sometimes — so we use Second Brain, a persistent-memory MCP that lets Claude remember decisions, project state, and failure patterns across sessions, so the same mistake doesn’t repeat in different chats. Gareth, working alone, can lose track of what was decided yesterday — so we use STATE.md files in each repository to record what was actually built versus what was claimed. Drafts can land wrong — so we use audit-then-draft discipline, where Claude reads the actual data file verbatim before authoring against it. Builds can drift — so we use build plans drafted in chat before any code session begins, scope explicit and execution well-specified. Sessions can finish on optimism rather than fact — so we use halt-and-report patterns at the end of every build, with verification (typecheck, build, accessibility, mobile, visual confirmation) before anything is claimed complete. Substantive content authoring goes through panel review against named scholars before disk write, with score thresholds that trigger revision. None of this was working in week one. All of it accumulated as responses to specific things going wrong.

The work moves between two surfaces. Most thinking happens in Claude chat — drafting prose, planning builds, running panel reviews, working out what a section should say. Execution happens in Claude Code, a terminal-based agentic environment with file system access, where Claude can read the repository, write code, run tests, and commit. The two surfaces hand off to each other through prompts: chat produces a paste-ready Code prompt that names the changes, the conventions, the verification steps; Code executes and returns a halt-and-report; chat reviews the report and decides what comes next. Working across both surfaces is itself part of the discipline — chat for judgment, Code for execution, neither pretending to do the other’s job.

Neither of us gets to assume authority. The work is better for it.

Over the next year Gareth read deeply into how different kinds of knowledge are acquired and assessed. Not as a research project — he’s not a research psychologist — but as a teacher trying to build something that would work. The literature on competency development, knowledge structures, prerequisite sequencing, and dispositional learning is rich, and what emerged was a clearer picture: different kinds of knowledge develop differently, have different prerequisite structures, and need to be assessed differently. A curriculum that treats them all the same will fail at least one.

That’s where the T1/T2/T3 distinction comes from — hierarchical knowledge that develops in a clear sequence, horizontal knowledge that develops through reasoning across perspectives, and dispositional knowledge that emerges across time and contexts. The framework’s structure follows from that distinction. It’s described in detail on the methodology page.

The COBIS accreditation process gave the work a deadline. We had to put what the school does into a defensible form — not just for accreditors, but for ourselves. That forcing function is what turned a set of ideas into a curriculum.

About halfway through this work Gareth built a tool — a curriculum decomposition harness — that takes a published curriculum document (from another country, another framework, another tradition) and decomposes it into the same structure REAL uses: KUDs, learning targets, criteria. This is what made the comparison work possible. Without it, comparing REAL’s framework to RSHE, Welsh CfW, and CASEL would have meant manually re-reading three large bodies of curriculum and trying to align them by intuition. With the harness, each external framework gets decomposed into the same shape, and the comparison becomes something a reader can actually engage with.

The harness is open-source. It’s the same tool that — in the longer run — could let any school decompose any curriculum into a structure that supports rigorous comparison and adaptation. That’s not the work this site documents, but it’s adjacent to it.

Some of this work could happen because of how REAL is organised. The school operates as a teal organisation — flat(er) hierarchies, distributed decision-making, real autonomy for teachers to take initiative on what they think the school needs. Building a curriculum framework over two years while teaching full-time isn’t possible in most schools because the institutional structure doesn’t support it. At REAL, the assumption is that teachers will build things. Barna Baráth, the school’s founder, has consistently backed slower, deeper work over the easier alternative. Without that protection, this framework would not exist.

That’s worth naming because it’s part of what the framework is and how it got made. It’s also part of what the site is trying to demonstrate — that this kind of work is possible when teachers are trusted to do it.

The work on this site points toward something the curriculum can’t quite do alone. The dispositions evidence is unambiguous that teacher conditions are the highest-leverage point in school-based wellbeing programmes — Jennings and colleagues’ CARE for Teachers trial showed that teacher emotion regulation and classroom emotional support can be developed and sustained; Roorda and colleagues’ meta-analyses of nearly 130,000 students confirmed that warm and contingent adult relationships are among the strongest replicated predictors of student engagement; Oberle and Schonert-Reichl’s cortisol study found that teacher burnout measurably propagates into the bodies of the children in the room. The full case lives on the methodology page. The implication for AI work is what we want to develop here.

If teacher dispositional state is the lever, and if current school structures systematically dysregulate teachers — workload, time pressure, administrative load, the demand to be present and warm with thirty children when you’ve slept badly — then AI’s contribution to wellbeing in schools is not primarily through student-facing applications. It’s through reducing the burden that pulls teachers away from the part of the work that actually moves the dial: noticing children, holding warmth under pressure, modelling the dispositions the curriculum names.

This is what the harness and the skills library are quietly working toward. The harness reduces the time it takes to align curriculum work with established frameworks from months to days. The skills library encodes the disciplinary patterns so a teacher doesn’t have to re-derive them every time. Together they hand teachers back time and capacity — not by replacing the relational work, which AI can’t do, but by removing the work that displaces it. A teacher who isn’t drowning in curriculum design is more present with the children in front of them. A school whose wellbeing PLC isn’t reinventing every framework from scratch can spend its meeting time on what’s happening in classrooms.

Our work is not directed at children directly. It’s directed at the conditions under which adults work with them. That distinction is load-bearing. AI applied directly to children — companion apps, anthropomorphic tutors, emotional-support bots — has a thin and contested evidence base and serious risks the recent literature is starting to document. AI applied to teacher capacity has the entire dispositions evidence base behind it as a leverage argument, and the absence of those risks. The work on this site is one school’s attempt to build it properly.

This site is the public articulation of one strand of work that’s still in progress. The framework is more developed in some bands than others. The relational and daily-habits side of the wellbeing programme — Community Circle, Reflection 360, the Play of Painting, Sit Spots, Solo-time, the practices Mary-Ann is documenting — isn’t yet captured here at the depth it deserves. We plan to introduce Sue Roffey’s Circle Solutions more systematically across the school next year. The agentic method itself is still being refined.

What Gareth would like a reader to take from this page is that the framework wasn’t handed down from a research institute or pulled from a textbook. It was built — by a teacher in a particular school, with colleagues, with AI, in conversation with the published research — over two years, with all the rough edges that implies. We think that’s a good way to build curriculum. We’re trying to figure it out as we go.

Mary-Ann Ridgway, co-principal, has done two distinct things that shaped this work. She challenged the design assumption that competency-based assessment alone could capture wellbeing — the conversation that sent him into the research on knowledge structures and led to the T1/T2/T3 distinction. She has also led the documentation of REAL’s lower-school wellbeing practices for COBIS, capturing the lived texture of the programme that this framework tries to be true to.

The wellbeing PLC — Emese Hasznos, Diana Abou Naccoul, and Margareta Merenyi — provided the conversational ground in which the thinking developed. The questions that pushed the framework forward came from this group as much as from anywhere else.

Barna Baráth, the school’s founder, has consistently backed slower, deeper work over the easier alternative. Without that protection, this framework would not exist in the form it does.

The teachers at REAL are the people who do this work in classrooms every day. What students experience comes from them, not from a curriculum document.

The framework also draws on substantial prior scholarship: the statutory PSHE/RSHE programme of study (UK Department for Education); the Curriculum for Wales 2022; CASEL’s social-emotional learning competencies; Sue Roffey’s Circle Solutions; and a wider literature credited on the methodology page.

Manning, G., & Claude (Anthropic). (2026). Wellbeing at REAL School Budapest. Available at [URL]. Licensed under CC BY-NC-SA 4.0.

If you adapt this for your own school under the licence, we’d be glad to hear about it.

gareth.manning@realschool.hu