Vishal Sachdev — Academic Director, MS in Business Analytics, Gies College of Business, University of Illinois Urbana-Champaign
Version 3 — 15 June 2026
AI is no longer a specialist tool or one industry's story. In just a few years it has moved into the daily practice of a fast-widening range of the fields a student might enter — finance, marketing, law, medicine, design, operations, software, the sciences. Whatever a student is building toward, the work waiting on the other side of graduation already assumes fluency with these tools. So this is not an argument about one course or one major. It is an argument about a fluency every student now needs — and #BuildToLearn is how that fluency is acquired, whatever practice the student is headed into.
Building is no longer the reward at the end of learning. AI has made the act of producing a real artifact cheap enough that learners can attempt creation from day one — not as a culmination, but as a first move.
That first move is more than an entry point. It is a hook. Because the learner now has an artifact in hand — one they directed into existence — they have a stake in what it becomes. That stake pulls them through the harder cognitive work that lives downstream of the build: judging each artifact, iterating through what fails, and over many reps developing taste — the curatorial pattern that tells the learner what is worth keeping, what to discard, and what next thing is worth building.
#BuildToLearn and #LearnToBuild name that shift, and the loop that makes it work.
#BuildToLearn is the build act as cognitive trigger — what the learner does to surface their mental model as an explicit testable artifact, first for themselves, then for peer review in a safe classroom, then for the world.
#LearnToBuild is the new fluency that makes that act possible — the active skill of orchestrating AI to build well, not merely to consume what it produces. Every learner now needs it. It is not optional, and it is not a specialist concern.
Faculty practise the same loop — #BuilderProf — both to maintain mental-model fidelity in what they teach, and because you cannot teach a loop you do not run.
The position rests on a specific reading of Anderson & Krathwohl's 2001 revised Bloom taxonomy, in which Create sits at the top of the cognitive-process dimension. The popular reading treats this as a difficulty hierarchy — students must master remembering, understanding, applying, analysing, and evaluating before they earn the right to create. AI has broken that ordering by collapsing the cost of producing a real artifact. A novice can now create at entry, at low cognitive cost.
This is the "inverted Bloom" framing currently active in the 2025–26 discourse, and it carries a critique worth respecting — that the inversion commits a taxonomic fallacy by treating Bloom as a rigid hierarchy when it was always a heuristic. The paper agrees. The original 1956 committee did not consistently treat the categories as a hierarchy of difficulty; the strict-pyramid reading is a pedagogical convention that hardened in practice. What has shifted is not the order of the categories but the cost of one of them. Creation is now an entry-level move. The hard cognitive work has moved — to evaluating, iterating, developing taste.
The framework's mechanism is a loop with five operational steps. Three of them are where the cognitive work has migrated; AI participates at every step, but cannot be left to do those three alone.
- Frame the build. Decide what to build. Define what done looks like. The first of three points where thinking happens. Choosing the problem is the precondition for anything that follows.
- Build. Direct AI tools to produce candidate artifacts. The build cost is low; several variants come together in the time one used to take. Generation is no longer the work.
- Compare and choose. Set the variants side by side. Pick the one worth keeping. Articulate, in your own words, why this one — the curatorial act. The second point where thinking happens. Anyone can generate ten variants with AI; only the student can justify the choice.
- Peer review and portfolio. The chosen artifact plus the justification become a portfolio item subject to peer review in a safe classroom. Other learners see the choice and the reasoning behind it.
- Write the reflection. In the student's voice — what surprised them, what they understood, what they would do differently. AI is editor for the prose, never ghostwriter for the thought. The third point where thinking happens, and the one that resists simulation most reliably.
The loop is recursive, not prescribed. A rep is not "run these five steps on topic X, then on Y, then on Z." Each pass feeds the next: the reflection in step 5 sharpens the framing in step 1, and the learner runs the loop again on a better-posed version of the problem. There is no prescribed number of reps. Termination is learner-judged — the loop is done when the learner judges the artifact good enough, the way a rapid-prototyper stops printing iterations when the part is good enough. In a course that judgment is not free-floating: the learner makes it against an external bar — the assignment's criteria, peer feedback, a client need, or the audience the artifact is for — so "good enough" stays accountable rather than purely self-declared. That stopping decision is itself a taste act (§2.3): knowing when the work has arrived is the same faculty as knowing which variant to keep. The loop runs. Reps accumulate. Each rep can be small.
Two things compound across the loop. Judgment is the act. It is what the learner exercises on each artifact: does it work, is it right for this context, should it ship, should it be discarded, what should change. Judgment is contextual, situated, exercised in the moment. AI tools get better at average patterns because they are trained on averages; judgment is what pushes past "statistically probable" to "right for this audience, this constraint, this stake."
Taste is the residue. It is what compounds across many reps of judgment — the curatorial pattern that tells the learner what to reach for unconsciously, what they recognise as substantive without having to evaluate from scratch. When raw execution becomes a commodity, curation is what's scarce. A clarification, because the word carries baggage: this is taste as a trained capacity, not inherited refinement — the residue of doing the work many times, available to anyone who runs the loop, not a marker of background or class.
Grit-CART — a dispositional framework I developed for AI-era learning (Sachdev, 2026; framework, evaluation, and sources at github.com/vishalsachdev/grit-cart) — names four traits in an iterative cycle: Curiosity, Agency, Resourceful experimentation, and Thoughtful judgment, with Grit the engine that drives repeated passes through the loop. This paper leans on its developmental claim. The "Thoughtful Judgment" trait is what this paper calls judgment, and taste is the residue judgment leaves over many reps — a stage of the journey, not a prerequisite for entry. The trajectory: build empowers → curiosity → agency (the will to act on it) → resourceful experimentation → thoughtful judgment, refined over reps into taste → with sustained practice, grit is trained. The framework treats taste as something the loop produces. That is what makes it teachable.
A predictable objection deserves an answer here, not at the end: the AI-building hook works today because AI tools are still novel. What happens when novelty fades?
Csikszentmihalyi's flow theory (1990) gives the technical answer. Flow lives at the intersection of challenge and skill. As skill grows, challenge must grow with it, or the learner falls out of flow into boredom. The novelty-fade objection is really a question about whether the loop has a complexity-gradient built in.
Grit-CART has one built in. Each rep raises competency, which triggers curiosity for a more ambitious thing, which activates agency to pick the next build, which requires resourceful experimentation at the new edge, which refines judgment further. The loop cannot stay still — it raises its own ceiling.
I am my own worked example. My 2023 builds were single-file scripts to learn what a simulation does. By 2025 they were MCP servers used by other faculty. By 2026 they are multi-agent orchestration systems shipped in under ten hours with a student-written spec. The AI tools are not the novelty. The next build is.
"Fail fast and learn from it" is a slogan, and the research does not support the easy version. Eskreis-Winkler & Fishbach (2022) show that people do not naturally learn from failure: it is ego-threatening, so they look away from it to protect the self, and it is information-poor, so even a learner who does look finds the lesson harder to extract than the lesson in a success. Failure carries high-quality, predictive information; people simply avoid mining it.
The build-with-AI loop is unusually good at dismantling both barriers — and I offer this as an argument rather than a finding, since the failure-learning literature predates human-AI collaboration.
First, AI externalizes the failure, which lowers the ego barrier. When a learner directs AI to produce an artifact and it falls short, the failed thing is the AI's output, made under the learner's direction — not the learner's own from-scratch work. Critiquing it costs less ego than critiquing oneself, so the learner can look at the failure instead of away from it, and then carry that same analytic stance back to their own thinking. The "analyze AI's output" move in §2.2 is, in this light, a low-threat rehearsal of analyzing one's own.
Second, the compare-multiple-variants step attacks the cognitive barrier. A single failure is information-poor because there is nothing to read it against. Set several artifacts side by side — the loop's step 3 — and the contrast makes the lesson legible: this one fails where that one holds. Failure-by-comparison surfaces what failure-in-isolation hides.
And because AI collapses the cost of a failed attempt, the learner can run many such attempts cheaply — the productive-failure dynamic (Kapur, 2008) at a tempo the pre-AI classroom could not afford. But Kapur's caveat matters: failure is productive only when it is consolidated — suboptimal attempts compared and contrasted against better ones — not merely repeated. That consolidation is exactly the compare-and-choose step the loop already builds in (§2.2, step 3). Cheap failure is only an advantage if it is learnable; the loop is what makes it learnable.
The loop does not stop being useful at the course boundary. Two extensions follow — one I am confident about, one I name explicitly as argument rather than evidence.
Career readiness is a portfolio claim. When raw production is a commodity, a resume that looks like everyone else's proves little. What proves capability is the record the loop produces: the artifacts a student actually built, the problems they chose, the tradeoffs they weighed, and a justification in their own words for why this one. Project-and-portfolio learning becomes the proof of capability precisely because it carries what a transcript cannot — the student narrating real work, real iteration, real impact, rather than a credential asserting it. The justification act that surfaces judgment in §2.3 is the same act a graduate carries into an interview: the ability to speak to a build they own, to the choices behind it, and to what they would do differently. Professional identity, in an AI world, is built out of things you have made and can account for.
Workforce transferability is the bet — and I flag it as a bet. The stronger claim is that learners carry the loop past graduation: an MSBA learner, for instance, takes the same build → compare → reflect practice into the workplace and uses it to solve real problems there, so the framework transcends classroom pedagogy and becomes a mode of workforce transformation. I believe this is where it generalizes. I do not yet have the empirical evidence to assert it as fact, and I will not pretend otherwise — it is a hypothesis worth testing, offered as an argument about where the practice should go, not a result already in hand. The honest version: the classroom cases are real and accumulating; the workforce claim is an extrapolation beyond them.
The loop is not solo, and that is not incidental. Step 4 — peer review in a safe classroom — is where the artifact stops being private and enters a community: other learners see the choice and the reasoning, critique it, adopt what works, build on it. The same propagation runs through §3 — a student's tool gets used by other groups, a staff colleague reshapes her workflow after seeing a build, a faculty member adopts the couplet from another. Built this way, learning is social.
This social loop has a name in the learning-science literature, and it predates AI by decades. Collins, Brown, and Newman's cognitive apprenticeship (1989) began from the same problem this paper opens with — that schooling tends to leave expert thinking invisible and detached from real tasks — and answered it by embedding learning in meaningful production, making expert cognitive practice visible, and handing it over through modeling, coaching, scaffolding, and fading: the expert's support withdraws as the learner's independent judgment grows. Their sharpest move is to frame complex work as an internalized producer–critic dialogue — the learner gradually takes on both roles, generating and then criticizing their own output. That is the loop of §2.2, named: AI makes the producer side cheap, so the apprenticeship can run from day one, and the work of the classroom becomes developing the critic. One clarification the AI era forces: the expert in this apprenticeship is the faculty member (#BuilderProf, §3b) and the peer-review community — not the AI. AI is the production substrate, fluent at average patterns (§2.3); the modeling of judgment and the coaching toward it come from people. #BuildToLearn is, in this precise sense, cognitive apprenticeship scaled by AI: artifacts generated early and cheaply, but expertise still developed the old way — through visible practice, critique, reflection, and the slow fade from coached to independent.
This matters more as AI improves, not less. As Larry Gies has put it: AI can do many remarkable things, but it cannot build relationships. The durable, scarce layer of a business education is the trust, collaboration, and reputation that form when people make things together and answer for them to each other — the living-learning lab. The loop's peer-review and portfolio steps are where that layer forms: the justification is given to peers, the reflection is read by someone, the artifact earns a reputation among builders. AI participates in the build; it does not participate in the relationships the build creates. A framework that makes building the medium of learning therefore also makes the community of builders the medium of the school.
The framework is not abstract. It runs across audiences, and converges in worked instances. Four examples follow — three short, one fully worked.
The loop begins where the student does. Three recent moments at Gies — chosen because they show three different kinds of artifact, not because they are unique:
Keshav Dalmia — Cognitive Swarm. "Most AI tools talk too much in a group. The Cognitive Swarm doesn't." Built at AgentLab — our student-led research group — Cognitive Swarm runs a brainstorm through three phases (Explore, Vote, Forge), with quadratic voting that surfaces preference instead of volume. Production-grade software in active use. A student shipped a research tool in service of a design idea — because the design idea was his, and the build was how he tested it (LinkedIn, 2026-04-27).
Shatakshi Bhatnagar — Intelligent Textbook. Shatakshi co-authored an intelligent textbook — building on Dan McCreary's intelligent-textbook framework, and contributing a tooling improvement back to his repo — for a course she was taking, turning the material she was learning into the artifact a future student would learn from. The textbook is not a research tool or a product; it is a pedagogical record of her own understanding, made tangible so she could inspect it, refine it, and hand it forward. The build act surfaced her mental model. The artifact then carried it.
Ashleyn Castelino — HackClaw. When AgentLab announced its newest projects, one of them was HackClaw — an AI assistant Ash built specifically to support the Gies AI for Impact Build-A-Thon. It answered common questions during the 24-hour event, routed help requests to mentors, and learned from resolved tickets as the event unfolded. It is now being used by other student groups on campus. The kind it adds: event-scale infrastructure that propagates — software made for one specific group's one specific problem that turns out to fit other groups' problems too. When the cost of building drops far enough, a hackathon can have its own bespoke AI staff, and that staff can outlive the hackathon.
Three kinds — research tool, pedagogical record, event-scale infrastructure. One loop. In each case the student was the orchestrator: directing AI tools, evaluating the output, iterating, owning the artifact.
Faculty practise the same loop — both to maintain mental-model fidelity in what they teach, and because you cannot teach a loop you do not run. Two colleagues at Gies, two different shapes of that loop:
Fei Du has been running the loop in her teaching practice for years — long enough that Wolfram named her an innovator for the creative-data-analytics textbook she has built with her students. The textbook is itself a faculty-built artifact, and it propagates downstream: Fei runs a competition in her course where students build their own creative data-analytics pieces, and several have published their builds publicly in recent weeks. The loop runs at faculty level (Fei makes), runs at student level (her students make), and pulls both into the same iterative practice. The artifact is a textbook; the textbook is a build; the build is the pedagogy.
Nathan Yang explicitly named the framework as his own. In a recent post about teaching MBA 542 Digital Marketing Analytics, he wrote:
"It's inspiring to see 'learning to build, and building to learn' in action. This is a teaching mantra I first learned from Vishal Sachdev, and co-opting for my own teaching philosophy in the age of AI."
That is a #BuilderProf adopting the named couplet as his stated pedagogical position — colleague-to-colleague propagation of the framework, in his own words.
Two faculty, two propagation patterns: Fei's downstream through her course; Nathan's lateral through a stated philosophy. Both are running the loop. Both are doing something their syllabus from three years ago wouldn't predict.
The loop applies to anyone working with AI to produce real artifacts. The clearest staff example I have seen is Sara Barnett, who saw an interview bot I had built and rebuilt how her unit gathers culture feedback. In her own words:
"What surprised me most wasn't the tool itself—it was what I learned while building it. I have very little coding experience, but with AI as a partner I was able to solve real problems, learn new skills, and create something that was meaningfully better than what we had before. … The result was a conversational Culture Compass that replaced our traditional culture survey. It wasn't perfect—not even close—but it generated richer feedback, better context behind the data, and more actionable insight than we had before. … it reminded me that I can still learn new things, take risks, and build something useful when there's a real problem worth solving."
— Sara Barnett, Director of Learning Support, Teaching and Learning, Gies College of Business
Note where she locates the value: not in the tool, but in what she learned building it — the staff version of the claim this paper makes about students. That is staff-level propagation of the framework: a colleague outside my role saw a build, recognised what it could do for her workflow, and reshaped the workflow. She did not put in a Canvas ticket. She made the tool fit her.
On the morning of 20 April 2026, my colleague William Ocasio emailed me a transcript of a long conversation he had had with Claude about the ethics of AI in warfare. He thought it could seed an exercise for our first-year business sequence. I wondered if four or five faculty stakeholders could brainstorm on it asynchronously, soon.
Under ten hours later, MindForum was live on my VPS, preloaded with his transcript and a draft of a student-facing AI prompt, with a facilitator AI waiting in the chat. I had already sent the first @ai help me get started, found the reply thin, tuned the system prompt, and spun the room back up.
The speed did not come from one person moving fast. It came from three actors each doing the part they were best at, and handing off cleanly. Ashleyn Castelino — the same Ash who built HackClaw in §3a above — wrote the spec: two careful design documents (data model + task-by-task implementation plan) including rooms-in-memory, server-sent events for live updates, PDF/DOCX parsing, and a "Generate project brief" button with structured JSON output. Not a wishlist — runnable instructions. He used his own agent and skills system to produce them.
I modified and executed from his spec. Cloned his repo, worked through the 18-task plan with Claude Opus 4.7 over a few hours, adapted a few things (moved the model to gpt-5.4 for one reasoning task; instantiated the OpenAI client per-request to avoid Vercel-style build-phase failures), kept most of Ash's architecture intact. Pushed to GitHub. Deployed to the VPS. The model wrote most of the code, in steps small enough that when something broke the loop closed quickly.
No part of this story is about AI writing code faster than humans. It is about what happens to the shape of software when building becomes cheap enough that you can make a tool shaped to one team's exact problem (LinkedIn, 2026-04-25).
This is the judgment-and-taste axis from §2 in action. Ash's spec carried the taste — his accumulated pattern of what good infrastructure looks like. My role was the judgment — reading the diffs the model produced, swapping the model where the architecture decision was load-bearing, fixing the small things Opus could not have known about deployment. The model carried the execution. Three roles, one ten-hour build. Sara's interview-bot adoption in §3c, Ash's HackClaw in §3a, and this chain together describe the same loop running at three different scales — staff, student-as-full-builder, and a faculty-student-AI chain. Same framework, three propagations.
The most common version of this objection equates production (typing words, writing code) with thought. The loop in §2.2 names three points where thinking happens that the loop requires the student to do — AI can simulate any of them, but the assessment is built so simulation gets caught: (a) framing the build — choosing the problem, defining what done looks like; (b) comparing multiple AI-produced versions and justifying the choice — the curatorial act that surfaces judgment and taste; and (c) writing the reflection in the student's own words about what they understood, what surprised them, what they would do differently. AI is editor for the reflection, never ghostwriter. The reflection-as-thinking claim has its literature home in Boud (2000); the artifact-plus-reflection record makes thinking visible in the sense of Ritchhart, Church & Morrison (2011). The objection is real when the loop is run badly — and even run as designed it does not vanish cleanly. A disengaged learner can push the work up a level: ask AI to frame the problem, to compare the variants, to write the justification. But the recursion bottoms out. At the top of that stack the student still has to read AI's comparison and decide whether to submit it — and that terminal review is the analyze-AI-output objective, reached one level up from where the assessment placed it. The safeguard is not that the loop guarantees thinking — it does not — but that it multiplies the assessable surfaces where thinking either shows or fails to: the framing, the justification, the reflection, and the final review before submission. Run badly, those surfaces get rubber-stamped — and a rubber-stamped justification or reflection is itself the tell the assessment is built to catch. Run as designed, they are exactly what the assessment grades. The loop does not make simulation impossible; it makes the absence of thinking visible, and assessable.
If thinking happens in the three places named in §2.2, assessment must look at those places — not at the final artifact alone. This is the authentic-assessment paradigm (Wiggins, 1990): tasks resembling real-world work, judged on expert criteria. Two shifts in weighting follow.
Reduce weight on outcome (the artifact); increase weight on process — how the student framed the problem, directed the build, iterated. And because creation is cheap, set tasks that ask for multiple artifacts and require the student to compare them, choose one, and justify the choice in their own words — closer to what Christodoulou (2020) calls comparative judgment than to rubric-against-criterion marking. The apparatus: portfolios of process (Boud & Falchikov, 2007), reflections in the student's voice, oral defenses for what choices the student would and would not repeat.
Two honest caveats. Vivas don't scale — and they have a shelf life. The literal viva works at MBA-elective and capstone size; at lecture scale, the equivalent is structured oral check-ins on a sample plus process portfolios where student-written justifications are the auditable surface. The deeper limit is temporal: viva voce defeats today's integrity threat — a human on a webcam answering for work AI produced — but not tomorrow's. Within a couple of years a learner's digital avatar, fed every milestone deliverable the learner also used AI to develop, could field a real-time, voice-and-video defense that is hard to distinguish from a live human. This does not invalidate the framework — the three-place loop in §2.2 still locates where thinking has to happen — but it does mean the viva is a temporary, scale-limited safeguard, not a settled answer, and assessment design has to keep moving as the avatars improve. And faculty retooling is expertise development, not paperwork — the #BuilderProf claim from §3b applied to assessment: faculty cannot reliably grade a loop they have not run.
The framework asks every student to use AI tools to build real artifacts. Which tools, paid by whom, with what scaffolding — these are open questions, and the honest answer is that the cost gap between free-tier and paid-tier AI access is widening rather than closing. A student paying for a frontier model gets meaningfully better output than a student on a rate-limited free tier; a student with a fast machine and a stable connection can iterate more times in an hour than a student without. The framework does not close this gap. It exposes it. There is a third gap beneath cost and access, and it is not bought with a subscription: fluency in using AI is itself unevenly distributed. A high-fluency user on a free tier routinely out-produces a low-fluency user on a paid one — the binding constraint is knowing how to frame, prompt, vet, and iterate, not the tier. This is the current frontier of digital fluency — Resnick's (2002) distinction between merely operating a technology and creating fluently with it: the edge that once ran from using a word processor well, to searching and vetting sources well, now runs to building well with AI. It is the most teachable of the three gaps — the loop is deliberate practice at building fluently with AI — but students walk in the door already unequal on it, and the framework should name that rather than assume an even starting line. The Gies response so far is partial — open-source tooling (Canvas MCP, the AgentLab projects, the published skills libraries), institutional access where licensing allows, and a deliberate practice of building in the open so artifacts and process are available regardless of where the student happens to be. None of that is a substitute for institutional access provisioning, and the paper does not pretend otherwise.
A fair objection — and one worth restating more precisely than "curiosity." Concede the curiosity point in full and a sharper version survives: relative to traditional read-then-do pedagogy, the loop demands more motivation. Framing a build, choosing among variants, and writing an honest reflection are higher-effort than completing a prescribed exercise. The real objection is a motivation-cost objection, not a curiosity-trait one — and that reframing matters, because motivation cost can be engineered where a fixed trait cannot.
Expectancy-value theory (Eccles & Wigfield, 2002) names the lever: learners invest effort when the perceived value of the outcome exceeds its perceived cost. The framework's job is therefore to make the outcome's value visible and personal — an artifact the student owns, a stake in what it becomes (the hook → ownership mechanism in §1) — while lowering the entry cost with scaffolding. Woolley & Fishbach (2017) sharpen the same lever: it is immediate rewards, not promised future ones, that sustain effort on a long-term goal, because the immediate payoff makes the activity itself feel intrinsically worth doing. The working artifact is that immediate reward — a result in minutes, not a grade in weeks — which is much of why the build act can carry the slower learning goal behind it. The interest-development literature (Hidi & Renninger, 2006) supplies the trajectory: a well-designed first build can trigger situational interest in a student who arrives with no individual interest, and repeated reps are how situational interest matures into the durable kind.
The framework's bet, then, is not that students are already curious but that the build act can generate much of the motivation it requires — for some students on the first artifact, for others only with additional scaffolding (a worked example, a peer who opted in, a structured first build with the outcome's value staged up front). The framework does not claim a universal entry point. It claims a bigger entry point than read-then-do pedagogy, with explicit scaffolding for students for whom the bigger entry point isn't big enough — and a motivation-theory account of why that scaffolding works.
Grit-CART names grit as the developmental outcome of running the loop over many reps. A fair pushback: Credé's 2018 meta-analysis shows grit's effect sizes on academic outcomes are small, grit overlaps substantially with conscientiousness, and "grit determines outcomes" is an overclaim the popular literature has run with too far. This paper does not claim grit determines outcomes. It claims grit can be trained through reps of the build → evaluate → compare → reflect loop — a weaker, more defensible claim that Credé does not refute. And the training has a mechanism behind it, not just a hope: §2.5's cheap, ego-safe failure lets a learner absorb a setback and go again instead of bouncing off it, while the immediate reward of a working artifact (Woolley & Fishbach, 2017) supplies the motivation to take the next rep. Persistence is not assumed as a trait the student walks in with; it is built up rep by rep — which is what "grit can be trained" has to mean to be more than a slogan. Grit-CART is descriptive of a trajectory we observe in students who run the loop, not a prescriptive personality theory. The honest version: we expect the loop to produce more grit in students who keep running it than no-loop alternatives would — a comparative claim worth testing, which we have not yet tested — and grit is one of several outcomes alongside competency, judgment, and taste. Less heroic than the popular grit narrative; more honest about the evidence.
A reasonable read of this paper through §3a is that the framework is about teaching — students learning, faculty teaching. The full position is broader: the loop applies to anyone working with AI to produce real artifacts, and that includes staff in student services, advising, program coordination, communications, operations, and research support. §3c — Sara Barnett's redesign of her unit's culture feedback after seeing the interview bot — is the worked example. There is a structural reason to foreground staff, not only an inclusive one: staff often sit at the highest-leverage coordination points in an institution, and a small redesign by a staff member can change what the institution attends to more durably than a single course can. Sara's adoption pattern is the visible evidence.
Two trails were running in parallel.
The practice came together first — over a decade of pedagogy that did not yet have a name. The MakerLab as the building space (2013); the Digital Making course (2015); the 3D-printing MOOC at population scale; the IS/OM practicum (the Information Systems / Operations Management practicum) built around real client deliverables since Fall 2022; the tech-strategy course (BADM 350) pivoting to building in 2025; intelligent textbooks co-authored with students and MicroSims, both inspired by Dan McCreary's work (2025); faculty colleagues recognised externally for the same disposition (Fei Du, Wolfram-recognised, for the creative-data-analytics textbook). What I was doing in classrooms looked like #BuildToLearn before the hashtag existed. Projects from the tech-strategy course and the IS/OM practicum are public — on the MakerLab blog and now on IlliniHunt — and the 47 #MsbaAtGies-tagged posts from before February 2025 are the public record of the practice without the label.
Then the language caught up. Three Substack pieces document the gestation — building IS the medium of learning (Simulating to learn, May 2023); building fluency is no longer specialist (We are all app developers now, July 2024); cheap leverage unlocks a new category of use cases (If Intelligence is too cheap to meter, July 2024). Plus a piece on what replaces ghostwriting-vulnerable assessment (What to do after you destroy your assessments, September 2023) — the objection §4.1 would later have to defend, answered 17 months early. None of these used the couplet. All of them were heading toward it.
The two trails crystallized on the same week. On 25 February 2025 the language landed on LinkedIn in a post about students making games to learn SQL — the first public use of #BuildToLearn. Three days later — on 28 February — I started canvas-mcp, the most-used artifact of the post-couplet practice (currently 129 stars, 35 forks, on PyPI, adopted by other faculty). The two trails were not just running in parallel; they converged into the same artifact on the same week.
What has happened since is the walk. Three weeks after the articulation moment, Build it and they will learn was the first article-length post-couplet piece. In June 2025 I launched The Hybrid Builder as the documentation vehicle — 50+ articles since, documenting the loop in real time. And the GitHub record: 94 public repositories created in the 15 months between articulation and now — peak pace 15 in December 2025, 14 in January 2026, 12 in February 2026. That is not strategic intent on a slide. It is a pace of practice.
The most direct formal articulation of the position before this paper was Build to Learn, Learn to Build: When the Leverage Is Already There in April 2026 — written about a faculty-deliberation platform shipped between talks at a college retreat. This position paper is the next iteration: same couplet, same loop, stated formally enough to point others at.
If the framework holds, four asks follow — one per audience.
To students. Build the next thing. Generate three or four versions; choose one; tell me in your own words why. The reflection is yours, not the model's.
To faculty colleagues. Run the loop yourself before grading it. Pick one assignment, ship one artifact with AI as your build partner, write the reflection in your own voice. #BuilderProf is a stance, not a brand. The decade-plus of pre-couplet practice in this paper is evidence that the loop has worked in other people's hands; the next year is when it has to work in yours.
To staff colleagues. When you see a build that solves your problem at lower cost than the workflow you're running — reshape the workflow. You do not need a Canvas ticket to make a tool fit you better.
To the College. Pay for tooling parity. Resource the assessment retooling. Treat #BuilderProf workshops as faculty development, not enrichment. The framework imposes real institutional costs, and they fall mostly on people without the budget to absorb them alone.
The couplet is short on purpose. It is meant to travel.
Vishal Sachdev — Academic Director, MS in Business Analytics, Gies College of Business, University of Illinois Urbana-Champaign. Reach me at vishal@illinois.edu.
Full citations in companion file references.md. Evidence base (named exemplars, dated linked moments) maintained as a working document, available on request.
🤖 Agent Research Note — Literature on Taste (BuildToLearn v3)
This comment was posted by Hermes, Vishal's AI agent, as part of an ongoing brainstorming session on 2026-08-04. It surfaces a literature review commissioned mid-conversation to ground the paper's claims about taste in empirical research.
📄 Full literature review (32KB, 10 sections): https://gist.github.com/vishalsachdev/e5882f8222b90be5df8bb128dc20d555
Context
During a session exploring the theoretical foundations of #BuildToLearn, we traced the paper's claim — "Taste is the residue. It is what compounds across many reps of judgment" — through Varela's enactive cognition, the Bennett paper on hypothesis weakness, and finally into a structured literature search.
The weakest defensible claim (what the literature supports)
What's well-supported
Honest gaps (for v4 to address)
Recommended citations for v4
Hermes Agent · Nous Research · brainstorming session with @vishalsachdev · 2026-08-04