A developer hands you a one-page program brief on a Tuesday: 220 units on a tight infill site, structured parking below, a community room facing the park, a pro forma that only closes if the gross-to-net efficiency clears 82 percent. They want three schemes by the design-charrette Friday. The old way was a weekend of foam-core massing studies and a junior designer pulling all-nighters in Revit. The new way is Autodesk Forma with the 2026 Neural CAD for Buildings foundation model: you describe the program and the constraints, and within minutes you have massing options scored on solar access, view corridors, and unit yield, each ready to push to Revit at LOD 200. The trap is mistaking that speed for finished judgment. Forma runs the generative-design engine, which optimizes exactly what you tell it to and is blind to everything you did not say, so the options it returns are candidates, not decisions, and the schematic that advances is the one you selected, verified, and own. This lesson designs the concept-to-schematic workflow where Forma generates the options and you, at the design-intent gate, decide which one becomes the project.

Why Concept-to-Schematic Is a Generative-Design Step, Not a Drafting Step

The early-design phase is underdetermined: at concept there are many ways to arrange 220 units on a site, and the question is not "draft the one we already have" but "explore the space of possibilities and find the ones worth developing." This is exactly what the generative-design engine is built for, the fourth of the program's four engines: where generative AI drafts language, computer vision reads the jobsite, and predictive ML forecasts the schedule, generative design searches a space of geometric options against an objective and returns the ones that score well. Autodesk Forma with the 2026 Neural CAD for Buildings model is such a tool: you give it the site, the program, and the objectives (maximize daylight, maximize yield, respect the setbacks), and it generates massing and layout options that optimize them faster than a human team could draw by hand.

The engine determines the discipline. A drafting tool produces what you specified, so verification asks "did it draft what I asked?" A generative-design tool produces options that optimize a stated objective, so verification asks the harder question: "does the option that scored best actually serve the project, including everything I did not state in the objective?" The engine's characteristic failure, which the program names the objective-function problem, governs this entire workflow: the tool optimizes the objective function it was given, ignoring every constraint not encoded in that function. Forma will return a massing that maximizes solar gain and unit count while quietly violating the neighborhood's contextual scale, ignoring the client's preference for a courtyard typology, or proposing a floor plate no contractor would frame economically, because none of those were in the objective.

So the human's role is the role generative design always demands: define the objective well, read the candidates, and select the one that serves the project against the full set of constraints, including the unstated ones the objective function did not capture. The speed is real and valuable, fifty massing studies in the time it took to draw three, but speed produces candidates, and the designer turns candidates into a decision.

The Three-Stage Pipeline: Massing, Environmental Analysis, Options

Forma's workflow runs in three connected stages, better seen as a pipeline than a single button. The first stage is massing: from the program brief and the site, Forma generates building volumes that fit the buildable envelope, respecting the setbacks, height limits, and lot coverage you encode. The 2026 Neural CAD model lets you drive this with natural language and rapid iteration, so "220 units, structured parking, community room toward the park, 82 percent efficiency target" becomes a starting set of massing studies in minutes, each a real volume you can manipulate, not a sketch.

The second stage is environmental analysis, where Forma earns much of its reputation: it runs fast solar, daylight, wind, and noise analysis on each massing option, so instead of waiting days for a consultant's daylight study you see relative daylight and solar performance across the options immediately. This lets you compare schemes on environmental performance while the massing is still cheap to change: a daylight problem caught at concept costs a sketch to fix, the same problem caught at CD costs a redesign. The third stage is options and metrics: Forma scores each scheme on the metrics you care about, daylight, energy proxy, and unit yield, and presents them so you can compare candidates side by side, which is the comparison matrix the applied problem will ask you to produce.

The pipeline is powerful because the stages are connected: a change to the massing immediately reflows the environmental analysis and yield metrics, so exploration is fast and consistent. But notice what it optimizes: solar, daylight, yield, the things it can measure and that you told it to maximize. It does not measure whether the massing reads as a good neighbor on the street, whether the unit mix matches the market, whether the implied structural grid is buildable, or whether the client will love it. Those are the unstated constraints, and the pipeline's fluency at the stated metrics can lull you into treating the top-scoring scheme as the answer when it is only the answer to the question you happened to encode.

Forma optimizes the objective you give it, daylight, yield, solar, and is blind to everything you did not encode, the context, the market, the constructability, the client's taste; so the option that scores best is the best answer to the question you asked, not necessarily the best scheme for the project, and the designer's job is to ask the missing questions before the scheme advances.

The Objective-Function Problem, Made Concrete in Massing

The objective-function problem is abstract until you watch it happen on a real site. Suppose you tell Forma to maximize unit yield and daylight on the infill site. It returns a scheme taller toward the rear lot line, because that maximizes daylight to the units and pushes yield up. The metrics are excellent: highest daylight score, highest yield, energy proxy in range. But the rear lot line abuts a two-story residential street, so the scheme puts a tower against neighbors' back yards, which will draw a contextual-scale objection at the planning hearing and may be politically impossible regardless of zoning. Forma did nothing wrong; it optimized exactly what you asked, and the contextual-scale constraint was never in the objective function, so it was invisible to the optimizer.

The engine is a faithful optimizer of the stated objective, which makes it powerful and dangerous in the same breath: powerful because it finds high-performing options you would not have drawn, dangerous because they are high-performing only on the stated axes and may be unacceptable on the unstated ones. The unstated constraints in early design are numerous and often the ones that decide the project: neighborhood context and entitlement politics, the client's aesthetic and brand, the real unit mix the market will absorb, the constructability of the floor plate and structural grid, the phasing if the site must stay partly operational, the accessibility and egress logic the massing must eventually accommodate. None is automatically in Forma's objective, and several cannot be cleanly encoded as an objective at all.

So the defense is twofold. First, encode as much as you legitimately can into the objective and constraints so the optimizer solves closer to the real problem: the setbacks, the height limits, the efficiency target, the daylight goal. Second, and more important, hold every unencoded constraint in your own head and apply it to the candidates yourself, reading each top-scoring option against the context, the market, the constructability, and the client, because those are the judgments the tool structurally cannot make. The candidates are the tool's contribution; the selection against the unstated constraints is yours.

Candidates, Not Decisions, and the Designer Who Owns the Selection

The program's rule for every generative-design step is that the tool produces candidates and the human produces the decision, and nowhere is the distinction more load-bearing, because the schematic sets the project's DNA for the next two years. Treat Forma's top-scoring scheme as the decision and you have let an optimizer blind to context, market, and constructability set the project's direction, under the illusion that the high scores meant the scheme was good. Treat the schemes as candidates and you read all of them, understand why each scored as it did, apply the unstated constraints, and select the one that serves the project, possibly the second- or third-ranked scheme because it reads better on the street or frames more economically.

This is responsible charge applied to early design. The licensed architect or designer in responsible charge owns the schematic; their judgment selected it; their name is on it to the client and eventually in the permit set. Forma generated the options and ran the analysis, real and valuable work, but generation is not authorship, and the architect does not get to point at the optimizer if the selected scheme proves a bad neighbor or an unbuildable plate. The decide-then-draft discipline applies in its purest form: use Forma as a tireless generator and analyst, generate widely, then the designer decides which scheme is the project, bringing the context, client, market, and constructability the tool could not, and only then does it get developed into the schematic set. Letting Forma's ranking decide and then drafting whatever it ranked first inverts the order and surrenders the judgment the responsible-charge role exists to exercise. The output is a selected, owned scheme, not a tool ranking.

The Design-Intent Gate Before the Schematic Advances

The program's five verification gates are design intent, code, contract authority, dollars, and life-safety, and the concept-to-schematic step sits squarely at the design-intent gate, with code verification close behind. The design-intent gate asks whether the option expresses what the project is meant to be: the program the owner gave you, the experience the building is meant to create, the relationship to its context, the client's intent. A Forma scheme can score beautifully on daylight and yield and still fail design intent, because intent includes everything from "does this feel like a place people want to live" to "does this honor the program the owner actually wrote," judgments the optimizer's metrics do not contain. Passing the gate means a human in responsible charge has read the selected scheme against the program and the intent and confirmed it expresses them, not merely that it scored well.

Code verification rides alongside, and the playbook is precise about the sequence: generate massing options that pass zoning, then layouts that pass code. At concept, code verification is preliminary, confirming the massing can plausibly accommodate egress, accessibility, and the occupancy's requirements, not running a full IBC 2024 Chapter 10 egress check, because the plan is not detailed enough yet. But you must confirm the scheme is not code-infeasible: that the floor plate can hold compliant egress paths, that parking and access work, that the massing does not assume something the code will forbid, because a scheme that cannot be made code-compliant is not a candidate at all, no matter its daylight score. The cardinal rule applies in its early-design form: verify against intent and preliminary code before the scheme advances to the schematic the client signs off on and the team builds upon, because the schematic is what every downstream decision inherits.

The verification is consequence-proportioned. A massing scheme that will set the project's direction for two years and feed the entitlement, the pro forma, and the eventual permit set warrants real scrutiny. The schematic that advances has been read against intent, checked for preliminary code feasibility, and selected with the unstated constraints in mind, which is the gate that turns Forma's fast candidates into a defensible schematic design.

The LOD 200 Revit Handoff and What Travels Downstream

The workflow ends with a handoff to Revit at LOD 200, the level of development where elements are generic systems with approximate size, shape, location, and orientation, appropriate for schematic design. This is where the early-design work becomes the foundation of the production model, and it deserves attention because what travels downstream travels with authority it may not have earned. A massing from Forma's optimizer carries the optimizer's blind spots into Revit unless the designer has corrected them: if the contextual-scale problem was not caught, it is now in the schematic model; if the floor plate is uneconomical to frame, that assumption is now baked into the geometry the structural engineer starts from.

So the handoff is a verification checkpoint, not a file transfer. Before the LOD 200 model becomes the team's shared starting point, the designer confirms that the scheme handed off is the selected, intent-verified, code-feasible one, not merely the highest-scoring Forma output, and that the model candidly represents the level of development it claims, not a precision the concept work does not support. Overstating the LOD sets up the downstream disciplines to build on sand. The handoff should also carry the basis: the program it serves, the objectives that drove the generation, and the judgments that drove the selection, so the schematic team and consultants understand not just the geometry but why this scheme and not another.

This connects the step to the project's information chain, the same interoperability reality the program returns to: Forma to Revit at LOD 200, then onward to consultants who may be on other platforms, with IFC 4.3 as the eventual exchange and the LOD discipline keeping everyone honest about how resolved the design actually is. The handoff is the moment the candidates become the project's committed direction, so it is the moment to confirm, one last time, that what is advancing is the scheme a human selected and owns, developed to a level of detail it can candidly support.

The Applied Problem: Three Viable Schemes and the Comparison Matrix

Here is the exercise, drawn from the playbook. For a 220-unit residential infill site, use Forma with the Neural CAD model to produce three viable schemes and the comparison matrix that puts them side by side on daylight, energy, and yield. The deliverable is not "the scheme Forma ranked first"; it is three candidates you have read, the matrix that makes their trade-offs legible, and your selection with the reasoning that selected it, including the unstated constraints Forma's metrics did not contain.

Produce three things. First, the generation setup and candidates: the program and objectives you encoded (the efficiency target, the daylight goal, the setbacks and height limits as constraints), and three viable massing-plus-layout schemes, each carried far enough to be a real candidate. Second, the comparison matrix: the three schemes scored on daylight, energy proxy, and unit yield from Forma's environmental analysis, presented so the trade-offs are visible, the high-yield scheme that scores lower on daylight, the high-daylight scheme that loses a few units, the balanced scheme. Third, and most important, the selection memo: which scheme you advance and why, naming explicitly the unstated constraints you weighed (contextual scale and entitlement politics, the real unit mix and market, the constructability of the floor plate, the client's intent and the program), and confirming the design-intent and preliminary-code verification that lets the scheme advance to the LOD 200 Revit handoff.

The lasting product is a workflow you can run on any site: Forma generates the candidates and the analysis at a speed no manual process can match, and you, in responsible charge, select against the full set of constraints and own the schematic that advances. The professional who masters this gets the engine's full leverage, fifty studies in the time of three, environmental analysis at the moment it is cheap to act on, without surrendering judgment to the optimizer, because they understand that Forma optimizes the objective it was given and the designer supplies everything the objective left out. That is the whole discipline: candidates from the engine, the decision from the human, at the design-intent gate, before the schematic becomes the project.

Key Takeaways

• Concept-to-schematic is a generative-design step, the fourth engine: Autodesk Forma with the 2026 Neural CAD for Buildings model searches a space of massing and layout options against a stated objective and returns the high-scoring ones, a fundamentally different activity from drafting that demands a different verification discipline.
• The pipeline runs in three connected stages, massing from program and site, fast environmental analysis (solar, daylight, wind, noise) on each option, and scored metrics, and its power is that a massing change reflows the analysis and yield instantly, making exploration fast at the moment design is cheapest to change.
• The objective-function problem governs the step: Forma optimizes exactly the objective it is given (daylight, yield, solar) and is blind to every constraint not encoded in it, so a top-scoring scheme can quietly violate contextual scale, ignore the real market, or imply an unbuildable floor plate.
• The defense is twofold: encode what you legitimately can (setbacks, height limits, efficiency and daylight goals), and hold every unencoded constraint (context, market, constructability, client intent) in your own head and apply it to the candidates yourself.
• The schemes are candidates, not decisions: the designer in responsible charge reads all of them, understands why each scored as it did, and selects the one that serves the project, possibly not Forma's top-ranked scheme, owning the schematic the way the decide-then-draft and responsible-charge principles require.
• The step sits at the design-intent gate with code verification close behind: passing means a human confirmed the scheme expresses the program and intent (not merely scored well) and is preliminarily code-feasible (the massing can hold compliant egress and access), per the playbook's sequence of zoning-passing massing then code-passing layouts.
• The workflow ends in an LOD 200 Revit handoff that is a verification checkpoint, not a file transfer: confirm the selected, intent-verified scheme is what advances, that the LOD candidly represents the concept's level of development, and that the basis (program, objectives, selection reasoning) travels with it.
• The artifact: three viable schemes for a 220-unit infill site, a comparison matrix scoring them on daylight, energy, and yield, and a selection memo naming the unstated constraints you weighed and confirming design-intent and preliminary-code verification, the product of an engine that generates candidates and a human who decides.