ELT design and editorial review, inside your AI.
Learning Brain ELT plugs into the AI you already use (Claude, ChatGPT, Codex, Cursor) and gives it an SLA-grounded team of ELT design, writing, and audit tools. Produce CEFR-coherent lesson skeletons, vocabulary sets, reading and listening tasks with construct-valid items, graded grammar practice, speaking tasks with CEFR-aligned rating descriptors, and synchronous speaking-session designs. Audit reading texts, listening scripts, dialogue models, and full coursebook pages against the rubric a senior ELT editor would apply — consistently, with substrate-cited reasoning per flag.
AI-generated ELT material reads fluently. The senior editor's pass catches what it doesn't do.
The current generation of AI tools produce ELT material with high surface plausibility. Dialogue that scans. Reading texts that pitch roughly to a stated level. Vocabulary sets that look reasonable. Comprehension exercises with four plausible options each.
What they miss is what a senior commissioning editor at OUP / CUP / Pearson catches in their first read: dialogue that forces present perfect into unnatural exchanges to demonstrate the target grammar. Reading texts where the over-level items aren't glossed or contextually supported. Comprehension exercises labelled "reading for gist" that actually test specific information. Vocabulary sets that promise productive use of items most learners can only handle receptively. Character rosters that wouldn't pass the publisher's modern inclusion checklist.
You can spend an hour catching these by hand on every page. Or you can give your AI an SLA-grounded audit tool that catches them in seconds, consistently, with substrate-cited reasoning per flag.
Eight ELT-specific tools, on top of nine cross-cutting ones. Seventeen in total.
The design tools. arch_design_elt_lesson produces a single-lesson skeleton (60–90 min) with a coherent shape: PPP, TBLT, text-based, dogme, or auto-selected. Each stage carries duration, purpose, predicted comprehension and production demand, and explicit substrate grounding. Refuses content-shaped goals ("cover present perfect") and aspirational ones ("become more fluent"). coach_design_speaking_session designs a synchronous (in-person or live online) speaking session with interaction patterns matched to class size and mode, CCQs for any target language, named monitoring focus per stage, and a dedicated feedback slot.
The writer tools. write_vocabulary_set produces a vocabulary set with selection rationale grounded in frequency / coverage / learner need, full per-item depth (word, lemma, POS, IPA, CEFR level, frequency band, collocations, example sentence), and a spaced-recycling plan. write_reading_task and write_listening_task produce skill-focused tasks with construct-valid items (no MCQ-for-gist, no T/F for unlicensed inference), distractor rationales, and explicit text- or audio-evidence per item; both refuse texts that exceed lexical-coverage thresholds for the claimed level. write_grammar_practice_set produces a graded set moving controlled → contextualised → freer, with L1-aware expected error profile and stage-matched correction strategy. write_speaking_task produces a task with real outcome, real information / opinion gap, and (for assessment intents) a CEFR-aligned rating descriptor anchored to intelligibility.
The audit tool. doctor_audit_for_input_quality — the differentiator. Takes an ELT text / dialogue / exercise / full lesson page and audits it for naturalness, CEFR level coherence (vocabulary and grammar), construct validity, cultural and contextual assumptions, inclusion, teach-vs-test mismatch, and pedagogical coherence. Returns per-flag locations, severities, substrate-grounded rationales, and concrete suggested revisions. The audit a senior ELT editor performs, applied consistently at scale.
lb_elicit_learner_context captures the fields ELT needs — CEFR level, L1 group, target skill focus, instructional setting, exam target, class size and mode — before any design tool will run.
Three audiences. Same product, different jobs.
Freelance ELT writers
Get the structure publishers want, with substrate citations attached. The lesson tool produces the skeleton; the vocab tool produces the items; both run the rubric self-check before returning. What publishers reject pages for — mis-pitched vocabulary load, no spaced-recycling plan, content-shaped goals — gets caught before you submit.
Commissioning editors
Run the input-quality audit on every freelancer submission, consistently across the team. The audit reads like the red-pen pass you'd do by hand: six categories of flag with location, severity, why-it's-wrong grounded in the SLA literature, and a concrete suggested revision. Substrate-cited reasoning the freelancer can engage with rather than dismiss as taste.
Teacher-trainers
CELTA / DELTA / CertTESOL trainees produce intuition-led plans. Every Learning Brain ELT output carries a substrate_grounding array of note IDs. Trainees can read the cited research and trace the reasoning — the gold standard for "show me your reasoning" assessment.
Three end-to-end jobs. One install, all flows.
Design a B1 reading lesson with a matched comprehension exercise
Freelance writer commissioned for a 60-minute B1 lesson on "starting a new job". Pastes the supplied source text. Asks Claude (with ELT pack installed) for the lesson skeleton, a vocabulary set, and a matched reading task.
Behind the scenes: elicit-context (CEFR + L1 + skill focus) → design lesson skeleton (text-based shape, 5 stages summing to 60 min) → write vocabulary set (8 items, mixed receptive/productive, with spaced-recycling plan) → write reading task (5 items, construct-validity-matched, with distractor rationales and text-evidence per item). Each step rubric-self-checks before returning.
lb_elicit_learner_context · arch_design_elt_lesson · write_vocabulary_set · write_reading_taskDesign a B2 speaking session for a 14-learner class
Teacher trainer designing a 75-minute live online session on negotiation. Asks Claude for a speaking task and a session plan.
Behind the scenes: elicit-context (CEFR + class size + mode) → write speaking task (ranking-and-decision task with real opinion gap, expected language, anticipated breakdowns, CEFR-aligned rating descriptor for formative use) → coach speaking session (run sheet with interaction patterns matched to online mode, CCQs for target language, named monitoring focus per stage, dedicated feedback slot).
lb_elicit_learner_context · write_speaking_task · coach_design_speaking_sessionAudit a draft coursebook page
Commissioning editor reviews a freelancer's B1 dialogue + comprehension exercise draft. Pastes the material. Asks for an audit.
Returns: overall verdict (clean / minor / substantial / send-back), six substantive flags (naturalness of forced grammar, mis-labelled gist exercise, register-incoherent discourse markers, thin inclusion, pedagogical mismatch), each with location, severity, substrate-grounded reasoning, and concrete suggested revision. Revision priority order. Senior-editor-grade output, in seconds.
doctor_audit_for_input_qualityBoth scenarios are documented end-to-end with literal prompts and illustrative outputs in the demo doc (GitHub).
Common questions
What's the difference between this and Learning Brain (core)?
Learning Brain (core) is the corporate-L&D product at learningbrain.ai — 32 tools across Architect / Writer / Coach / Doctor / Learning Scientist personas, designed for instructional designers, training managers, and course creators.
Learning Brain ELT is a sibling product, same architecture, with ELT-specific tools and an SLA-grounded substrate. Audience: freelance ELT writers, commissioning editors at ELT publishers, teacher-trainers. The two share the cross-cutting and learning-scientist tools (9 of 17 here) but the design / write / audit / coach tools are ELT-specific.
Different domain, different audience, different evidence base. Install whichever matches your work. They can coexist in the same Claude account — the skill files have distinct names (learning-brain and learning-brain-elt).
What's the substrate based on?
314 notes across 24 domains. The 195 cross-cutting notes (cognitive load, retrieval, multimedia, assessment, instructional design, course design, learner psychology, etc.) are the same substrate that powers Learning Brain core — built from ~360 primary research sources over four years.
The 119 ELT-specific notes span 15 domains: SLA foundations (input / interaction / output / noticing hypotheses, interlanguage, focus on form, skill acquisition, implicit-vs-explicit knowledge, age effects), vocabulary acquisition, grammar instruction, the four skills (including phonics / decoding, reading fluency, academic-writing feedback), assessment and language testing (Bachman-Palmer validity, ALTE standards, computer-adaptive testing, automated scoring, fairness and bias, remote proctoring, AI in assessment), methodology and materials, learner factors, sociolinguistic-and-contextual (World Englishes / ELF, translanguaging, native-speakerism, ELT-as-global-industry), CEFR and frameworks (Companion Volume 2020 first-class additions, English Profile), materials and publisher workflows, digital and AI in ELT, inclusion and accessibility (dyslexia / SpLDs, neurodiversity, multilingual learners, refugee / migrant, low-literacy adults, trauma-informed), teacher development (cognition, classroom management, wellbeing, novice-vs-expert, reflective practice), discourse and pragmatics (pragmatics, interactional competence, discourse markers, genre / register, politeness, intercultural competence, corpus pragmatics), and specific contexts (CLIL, EMI, EAL, ESOL, business / BELF, EAP, ESP, young learners, teen, exam prep, vocational).
Every note grounds in primary literature with explicit source citations. v0.2 substrate (2026-05-26).
Does it cost money?
No. Free to use. The MCP server runs at near-zero marginal cost per call (substrate query + prompt scaffold → HTTP response, ~10ms of compute), and the actual reasoning happens in your existing Claude / ChatGPT / Codex subscription — not in our infrastructure.
That zero-marginal-cost architecture is intentional — it's what lets us keep the product free without a business model dependency. If you find it useful, send the link to a colleague.
How do I install it?
Two things to install: the MCP connector (the 17 tools) and the workflow Skill (how your AI uses them). On Claude Desktop / Cowork, both install at the account level and propagate to Claude Chat, Claude Cowork, and Claude Code's Desktop tab from a single setup.
Full step-by-step per AI tool on the connect page. Typical install: ~3 minutes.
Will the audit tool replace my editorial team?
No. It runs the consistent rubric pass that senior editors run by hand, in seconds, across every page. That removes the bottleneck of "we have one senior editor who can audit, and they're a queue" without removing the judgment work editors do downstream — deciding which flags to push back on, negotiating revisions with freelancers, calibrating the bar across the unit and the series, and making the editorial calls only humans can make.
The realistic outcome: every page gets the audit pass; editors apply human judgment to what the audit surfaces; freelancers get faster, more consistent feedback; the editorial team scales without quality drift.