Designing a facade with solar logic is one of those tasks where AI can do the math in seconds, as long as you feed it the right location, date, and constraints, and keep the conversation from drifting into old assumptions.
- Solar azimuth at a specific location and date determines where the sun is in the sky at any given hour.
- Subtracting ninety degrees from the azimuth gives the fin rotation needed to block direct sun rather than align with it.
- Chain-of-thought prompting and hard resets are essential for keeping earlier context from bleeding into new calculations.
This lesson is a preview from our AI Workflows for AEC Professionals Course Online. Enroll in a course for detailed lessons, live instructor support, and project-based training.
Once a facade table is set up, it is tempting to treat it as a static output. The real power shows up when you start asking the AI to recalculate values based on design goals that matter to the project, like blocking morning sun on east-facing glass at different times of day on different floors. That is exactly the kind of problem where a chatbot that knows how to compute solar angles can save an hour of manual work.
Moving Beyond Random Math
Earlier exploration with the facade table relied on mathematical principles like random values, combined with addition and subtraction to nudge shade angles. That is fine for studies, but a real project deserves a more rigorous approach. The natural next step is to think about solar angles directly. The AI chatbot can compute them if the prompt includes the right information.
A clean update prompt starts with a clear instruction. Update table. Then add a location and date. Assume the location is Ironwood, Michigan, and the date is summer solstice. The chatbot does not need you to convert summer solstice to a calendar date. It knows that June 21 is the summer solstice and will factor that into the solar calculations.
Setting Different Times by Floor
Different floors often have different daylighting goals. The basement might tolerate direct morning sun longer than an upper level, or the opposite might be true based on program and views. A useful way to encode that in a prompt is to set a target block time per level. For an east-facing wall, the prompt might read as follows:
- Basement, block maximum direct sun until 11:30 AM, so the fin rotation is 11:30 AM azimuth minus 90 degrees.
- Level one, block earlier at 9:30 AM, so the fin rotation is 9:30 AM azimuth minus 90 degrees.
- Level two, block at 10:30 AM, so the fin rotation is 10:30 AM azimuth minus 90 degrees.
The minus ninety is important. If the fin were set parallel to the sun, it would let the light pass through. Rotating the fin ninety degrees off the solar azimuth orients it perpendicular to the sun, which is what actually blocks direct rays. Building that logic directly into the prompt keeps the model from guessing at the geometry.
Maintain the Table Format
Chatbots have a habit of reformatting outputs when new information arrives. A single word can prevent that drift. Add maintain to the prompt, as in maintain the 3 × 20 markdown table. That tells the model to keep the shape and structure consistent with the earlier version, so downstream scripts and spreadsheets continue to work without rework.
It is also worth setting limits on the calculation. Shade fins cannot rotate through the glass or into the building, so bounding the values with max value 90 and min value 0 is a safe precaution. The model should return values in that range for this kind of calculation anyway, but the explicit guardrails prevent occasional outliers caused by the prompt history or a misinterpreted constraint.
Hallucinations and Lingering Context
Even with a clean prompt, the first result might still look like it is using an older rule such as the earlier random logic. That is not quite hallucination. It is the chatbot blending new instructions with prior context, which is common in long threads. A quick way to check is to look at the table values and compare them against what a real solar calculation would produce. If the numbers still look random, the model has not fully reset its understanding of the rule.
The fix is chain-of-thought prompting. Copy the original update prompt from the sandbox or chat, paste it again, and before any of the new logic add a line that forces a reset. Something like, begin by updating all values of the table to zero, works well. That instruction tells the model to wipe the table in memory before applying the new solar logic, which strips out the random math that was lingering from earlier prompts.
Run the Corrected Prompt
After adding the reset instruction, submit the updated prompt. The chatbot zeroes the table in memory, applies the new rules, and returns a clean set of rotation values. Each floor shows the correct angle for its target block time, within the zero to ninety degree range. The table format stays consistent, which means it can be dropped back into the existing workflow without reformatting.
From here, the design team can review the numbers, adjust target times by floor, or layer in additional prompts such as alternate dates like winter solstice or equinox. Each new version of the prompt is a small iteration on the same structure, which is one of the strengths of keeping prompts in a sandbox document rather than typing them freshly each time.
Why This Pattern Generalizes
This particular facade exercise is specific, but the pattern shows up everywhere in AEC work. Anytime a table is driven by a rule that depends on location, time, or other external data, the same technique applies. Write the rule explicitly. Add guardrails on the valid range. Use the word maintain to preserve the format. When the model drifts, reset the state with a zero-everything instruction before introducing the new logic.
Each of those habits pays off across projects, because it keeps the AI grounded in the current task rather than mixing in leftover assumptions from earlier prompts. That discipline is what turns a promising chatbot experiment into a reliable design tool.
AI can compute solar azimuth for a specific location and date, which makes it a natural fit for facade studies that depend on sun position. Build prompts that name the location, date, and target block times by floor, subtract ninety degrees to orient the fins perpendicular to the sun, and use maintain to keep the table format consistent. When earlier context bleeds into new results, reset the table to zero before applying the new rule. Done well, the combination produces a design-ready facade table in minutes and keeps the conversation honest from one iteration to the next.
