Jake Van Clief ("Clief") is a USMC veteran and Edinburgh-trained founder whose company Eduba sells AI training, consulting and a multi-party AI workspace. His work rests on one idea: AI is the interpreter inside a program people design, so structure, intent and human review matter more than agents.
Who he is
"Clief" is Jake Van Clief, who runs the Clief Notes Skool community and is Founder and CEO of Eduba (Crunchbase).
Military: US Marine Corps, about 2015 to 2020, avionics and cryptographic systems on F-35 and F-18 aircraft (Eduba team; alphaXiv profile, which marks the dates unverified).
Education: BA in Political Science, Flagler College, 2024, as a nontraditional student (Flagler); MSc Future Governance, University of Edinburgh Futures Institute, 2024 to 2025.
Other roles: AI governance and international research advisor at ICR Research (London), 2025 onward; earlier Van Clief Media LLC, which is now the legal entity behind Eduba.
Research: first author of Interpretable Context Methodology (ICM, March 2026, with David McDermott) and co-author of The Ethics Engine (October 2025), a no-code pipeline for psychometric testing of LLMs.
Channels: YouTube (@JEVanClief), a Substack, Facebook reels (see R015), and the icm-architect skill on GitHub (user RinDig).
Eduba
Eduba is Van Clief Media LLC (Palm Coast, Florida), a "veteran-owned AI consulting and training firm" with one software product and a learning ladder that feeds it (eduba.io; Terms, updated 14 July 2026). Crunchbase places the founder in Edinburgh.
| Part | What it is | Notes |
|---|---|---|
| Consulting and training | Workflow teardown by department, then training staff to build | Self-reported: Pacific Life and Colgate-Palmolive, 600+ trained, "95% adoption after 30 days"; KPMG UK, 40+ executives (works) |
| chat.eduba.io | "A workspace for thinking with humans and agents in the loop" | Units are incubators: a topic, a system prompt, Markdown context files and participants, who can be people or agent personas (features) |
| Enterprise tier | Org admins define personas over admin-approved model providers | Role-based access per org and incubator, full message history, per-org subscription with an "ask allowance and credit balance" (enterprise) |
| Edubaware, Ledger | Named in the Terms: a platform for "structured AI context systems" and a talent placement site | Unclear whether Edubaware is the chat product |
| The Lyceum | Twelve-week, capstone-based certification, cohorts by skill level | lyceum.eduba.io |
| Clief Notes (Skool) | Free community and courses: Foundation, Archive, Implementation Playbooks, Building Your Stack, Vault | "12,500+ members" on eduba.io, versus 30,000+ claimed in his course; neither verified |
The workspace has no public price list and names no models. Explore (hierarchical knowledge), card actions and flashcards generated from your own incubators are its listed features. The Terms exclude HIPAA, FISMA and GLBA data and say hosting is US-only.
For engineers, the product is ICM made multi-party: an incubator is a scoped context (system prompt plus Markdown files) shared by humans and model personas, the same "right files for this job" idea as a routed folder.
Central theses
Six claims recur across the paper, essays, videos and Eduba's sales copy. They stack: the first is the premise, and the others follow from it.
AI is part of a program, not the program. A program is instructions, something that runs them, and surviving state; the model is only the runner. "We are turning English into a programming language whether we meant to or not" (Substack, 20 Oct 2026). So the engineering lives in the instructions and state, not the model.
You rarely need agents; you need structure. "Agent" is a naming convention for a bundle of prompt, tools and context. One model with a read/write/search harness, numbered stage folders, a routing
CLAUDE.md/AGENTS.mdand per-stageCONTEXT.mdcontracts replaces a multi-agent framework for sequential, human-reviewed work. Principles: one stage per job, plain text as the interface, layered context loading, every output an edit surface, "configure the factory, not the product" (ICM paper; video).Augment, don't just automate. Drawn from Engelbart's 1962 Augmenting Human Intellect and Weinberg's 1971 "Levine the Tailor": software has made people fit it; AI is the first layer that can bend software to the person. One-pass efficiency gains are "barren"; value comes from regenerative human-tool interaction (video; Engelbart walkthrough).
Most workflows do not belong to AI end to end. Break each department's work apart and sort every step into traditional software, rule-based logic, AI, or "do not build this". Steps that stabilize become scripts; a check (person, script or small model) ends every step (eduba.io; icm-architect).
Value migrates up the abstraction ladder. Three layers of AI use: prompts to outputs (the "book"), built flows (the "movie"), and flows improved from reviewed results (the "game"). Raw output commoditizes, so value moves to process, context and judgment; beware building what the platform will absorb next release (video).
Framing and intent outrank code. Psychology of programming matters more than programming for the next decade; human flexibility "abstracts up. Every time", and "the need for why never goes away". Teach principles that survive platform changes, and put governance in the build, not after rollout (R015; Substack; eduba.io).
The stance is anti-hype by design: Eduba's tagline is "We start with truth, not hype", and its KPMG case asks "when should the answer be no?". Older Substack titles ("The Orchestration Imperative", "The Rise of Computational Orchestration", 2025) show the same line before ICM: AI rollouts fail because nobody orchestrates the work around the model.
Assessment for engineers
The architecture claims are sound and match what strong harness teams already do; the evidence behind the bigger claims is thin and self-reported.
What holds
Scoped context per stage is well supported: ICM's own counts put a stage at about 2,000 to 8,000 tokens against 40,000+ for a monolithic prompt, and the long-context literature it cites backs the direction.
Files as state give inspectability, diffability and cheap resumption. Every intermediate artifact is editable before the next stage runs, which is a real review affordance.
"Sort each step: software, rules, AI, or don't build" is a good default for enterprise rollouts and pushes against agent sprawl.
What is unproven
ICM's evaluation is practitioner reports from an invite-only group of 52 (33 using multi-stage workspaces; 30 of 33 reported heavy editing at the first and last stages). There is no controlled comparison against monolithic prompting or a framework, and only the Claude family was tried.
Adoption, hours-saved and membership figures on eduba.io and in the course are self-reported, and the membership counts disagree (12,500+ versus 30,000+).
"Deterministic workflows" overstates it: a model reading Markdown is a stochastic interpreter. Determinism has to come from fixed routes, scripts and failing checks (the gap R015 already flagged).
Failure modes to expect
Scope ceiling: the paper itself rules out concurrency, real-time agent-to-agent traffic and complex branching. Teams will hit it when a workflow fans out.
Context rot by hand: routing tables and
CONTEXT.mdcontracts drift unless something tests them. icm-architect's "walk test" (a memory-less agent must orient from files alone) is the right check but is manual.Persona proliferation in the workspace product: admin-defined personas over several providers recreate the "one agent per job" pattern the theses argue against, unless incubators stay narrow.
Commodity risk for the method: Agent Skills, AGENTS.md and vendor project memory now ship the same primitives, so the durable value is the teaching and the teardown, which his ladder thesis predicts.
Review checklist
Use these when a team proposes an ICM-style folder workflow or a tool like Eduba's workspace.
Is the workflow really sequential with human review, or does it branch, fan out or need concurrency (outside ICM's stated scope)?
Does each stage name its inputs, its output file and the check that ends it (person, script or small model)?
Which steps are stable enough to be scripts instead of prompts?
Is there an automated walk test: a fresh session orients and reports status from files alone, scored, on every change to the routing file?
Are claimed savings measured against a baseline, or only reported by users?
Would the next platform release absorb this step (the ladder trap)?
For shared workspaces: who owns each persona's system prompt, and is data class (HIPAA, regulated) allowed by the vendor's terms?
Related
R015 Psychology of programming and deterministic workflows: his reel on Weinberg and "agent is a naming convention".
R020 The model proposes, the harness enforces: where the determinism R015 claims has to come from.
Sources
Crunchbase: Jake Clief; alphaXiv profile; Flagler College profile
Interpretable Context Methodology, arXiv:2603.16021; The Ethics Engine, arXiv:2510.11742
AI is part of a program. It is not the program. and the Substack archive
Videos: Do You Actually Need AI Agents?, What Do We Build Now That AI Can Do the Work?, Engelbart's 1962 paper, The Ladder That Explains Every AI Failure