TL;DR
Source: a 105-second Facebook reel by Jake Van Clief (creator of Clief Notes and the Interpretable Context Methodology, ICM), captioned "everyone wonders why I did my master's thesis on the psychometrics of language models". Published around September 2026.
Claim: over the next decade, psychology (how people frame the problem) will matter more than programming. He says he was flown to Dubai for a three-week contract with large companies to teach this one idea.
Anchor: Gerald Weinberg's The Psychology of Computer Programming (1971). Weinberg estimated that under 20% of the code in the systems he studied went to the actual problem, and that much of the rest could be removed by a change in how the programmers thought, not by better machines.
Application to AI: "agent" is a naming convention. Teams write hundreds of lines of code just to steer one agent to the right place at the right time. Copilot-style assistants and custom GPTs, he says, were the psychologically wrong frame. Drop "agent" from the goal and ask which workflow you are automating; the routing comes out completely different.
Mechanism: route with
CLAUDE.md/AGENTS.mdto a series of instructions kept as plain Markdown files. Use the model as the runtime for those instructions ("like a compiler on top of the software we've built") rather than as a generator of new code.Verdict for a staff engineer: the reframe is sound and matches what the stronger entries in this index converge on (fixed routing, model as executor, scripts for steady steps). But the video calls it a workflow, not a deterministic one, and an LLM interpreting Markdown is not deterministic. Determinism lives in what you take away from the model: the route, the file contracts, the scripts and the checks. Weinberg's 20/80 is a 1971 estimate, not a measurement, and the Dubai engagement is unverifiable marketing.
What the video says
The reel has no written post beyond the one-line caption; the argument is in the spoken track, burned in as one-word captions. In order:
Thesis (0:00 to 0:04). "For the next 10 years, psychology is going to be more important than programming."
Credential (0:04 to 0:10). He is in Dubai for three weeks, contracted by "some big companies" specifically on this concept.
Historical hook (0:10 to 0:43). Programming and software engineering have "always been about psychology". He reads from Weinberg's 1971 book: programming is not only human behaviour but complex human behaviour; under 20% of code is aimed at the problem itself; much of the other 80% could go entirely, through psychological rather than physical change.
Translation to agents (0:43 to 0:54). Psychologically, "agent" is a naming convention. People write hundreds of lines of code just to get one agent to the right place at the right time.
Diagnosis (0:54 to 1:07). Microsoft Copilot and custom GPTs were "the rage" last year, but were the psychologically wrong thing to build: the frame you adopt changes what you are even programming for.
Reframe (1:07 to 1:21). Remove "agent" from the goal and think about automating a workflow; you route completely differently. Many people use
CLAUDE.mdorAGENTS.mdto route to a series of instructions that are just Markdown files. The screen shows a workflow diagram with the rule every arrow needs a verb: reads, uses, writes, and a node labelled "Owner reviews" pointing to agaps.mdfile.Mechanism (1:21 to 1:35). Instead of coding, or using AI to create new code, use the AI as the runtime for the instructions, "like a compiler on top of the software we've built".
Close (1:35 to 1:45). Switch your psychological frame of where the information is and why you are using the tool, and you stop building the wrong things and start building things that last.
All quotes are short excerpts from the captions; the rest is paraphrase.
The Weinberg source
Gerald Weinberg's The Psychology of Computer Programming (Van Nostrand Reinhold, 1971) is the founding text of the field: programming as a social and cognitive activity, "egoless programming", and the idea that teams, not tools, set the limits. The passage on screen walks through a programming example and concludes that most of the code existed because of how the problem was understood and divided, not because the problem needed it.
What it supports:
Framing drives volume. How you name and partition the problem decides how much code you write. This carries over directly: if the unit of design is "an agent", you build orchestration, memory and handoff code around a persona; if it is "a workflow", most of that is replaced by a route and some files.
Human factors are first-order. Weinberg's point that people, not machines, are the binding constraint fits a world where the model is cheap and the human's model of the work is scarce.
What it does not support:
The numbers. "Under 20%" and "80% removable" are Weinberg's estimates from one worked example in a 1960s setting (tight memory, hand-tuned code), not a measured ratio. Treat them as rhetoric.
The leap to AI. Weinberg argues for changing the programmer's mind, not for handing execution to an interpreter that reads prose. The video's conclusion (use the model as runtime) is Van Clief's, not Weinberg's.
"Psychometrics of language models". The caption's hook is a different discipline (measuring a model's traits with psychological instruments). The video never connects it to the argument.
Architecture: agent frame vs workflow frame
The video's core move is a change of design unit. In the agent frame you build a persona and then the code that keeps it on track. In the workflow frame you write down the route and the step contracts, and the model executes one narrow step at a time.
The right-hand side is what the on-screen diagram in the reel shows: steps as Markdown files, arrows labelled with verbs, an owner review that writes gaps to a file. What turns it from "a workflow" into a deterministic one is everything the model does not decide: the route table, the scripts and the checks.
Tradeoffs and failure modes
Where determinism actually comes from. A model reading Markdown instructions is a stochastic interpreter: the same file and the same input can take different paths, skip a step, or write to the wrong file. A workflow becomes deterministic only to the degree you remove decisions from the model:
| Layer | Deterministic if | Typical leak |
|---|---|---|
| Route (which step runs next) | encoded as a fixed table, state machine or engine, not inferred from prose | the model decides it is "done" early or loops |
| Step contract | each step names its inputs and its single output file ("every arrow needs a verb") | step writes extra files or edits upstream ones |
| Repeated work | turned into a script the model calls | the model re-derives the same transform each run, slightly differently |
| Exit check | a validator, test or schema check that fails the step | an LLM self-check that always passes |
| Human gate | placed at irreversible or judgment-heavy steps ("Owner reviews") | gate on every step, so people rubber-stamp |
The video's "compiler" metaphor overstates the first row. A compiler is a deterministic function of its input; the model is closer to a junior operator following a runbook. The design goal is to make the runbook so narrow that variance does not matter.
Where the reframe wins.
Bounded, repeatable processes (content pipelines, research digests, intake and triage, report generation) where the steps are known and the value is consistency.
Teams that over-built: multi-agent orchestration, personas and handoff code for what is really a sequence of five steps.
Change cost: editing a Markdown step is cheaper than redeploying code, and non-engineers can review it.
Where it loses.
Open-ended work where the route is the hard part (debugging, research with unknown structure). Forcing a fixed workflow there hides the real decision inside one oversized step.
Scale and audit. Markdown routing has no types, no versioned interface and no test harness by default; drift between
CLAUDE.mdand the step files is silent.Cost and latency. An LLM as runtime for every step is slower and dearer than code for steps that are already understood. The right end state for a stable step is a script, with the model only at the edges.
Evidence. This rests on anecdote and a client story; there is no before/after measure of code removed or reliability gained.
Review checklist
Use this when someone proposes "an agent" for a job:
Can the job be written as a list of steps with named inputs and one output each? If yes, it is a workflow; design the route first and the prompts second.
Is the route held outside the model (a table in
CLAUDE.md/AGENTS.mdat minimum, an engine or state machine for anything audited)?Does every arrow in the diagram have a verb (reads, uses, writes), and does each step write only its declared output?
Which steps have run the same way three times? Those become scripts the model calls.
Does each step end in a check that can fail without asking the model's opinion (schema, test, diff, file exists)?
Are human gates only where a step is irreversible or needs judgment, with what the reviewer is checking written down?
How much code would disappear if "agent" were removed from the design doc? Personas, handoff protocols and memory layers built for one linear process are the Weinberg 80%.
Is there a before/after measure (pass rate, rework, time, cost) rather than a story?
Related entries
R012, AI crash projects for public good. A commenter's line there, the model can phrase but cannot decide, is the same split between a fixed route and a model executor.
R006, harness terms; R005, five-layer map. The workflow frame moves "who decides" (R005's loop layer) out of the model and into the route; R006's guardrails become the exit checks.
R014, Intro to AI course; R012's learning RCT. The human side of the psychology claim: how people frame the tool decides what they get from it.
Source notes
Facebook reel by Jake Van Clief, 105.5 s. Published about 2026-09-06 by page metadata (not shown on the page). About 1,200 reactions, 61 comments and 114 shares when read.
The Weinberg passage is paraphrased from the page shown in the video, so it follows his wording only loosely.
The Dubai engagement and the companies involved are unverified.