AI Prompt Engineering for Feng Shui Analysis: A Practical Guide
The intersection of traditional Chinese spatial analysis and modern large language models has produced something most practitioners either dismiss entirely or overvalue. Both positions miss the actual utility. The system works as a reasoning scaffold, not as an oracle. I have been running these workflows for roughly three years across residential and commercial assignments, and the results are useful when you understand what they actually produce. Feng Shui Prompts Modern refers to structured prompt templates designed to make LLMs perform spatial analysis using classical formulas — primarily the Eight Mansions method, the Flying Star system, and Ba Zhai room placement. The prompts constrain the model to output structured readings: sitting direction, facing direction, Kua number mapping, sector-by-sector star ratings, and cures based on elemental theory. The value is not in the AI knowing feng shui. No AI knows anything about feng shui. The value is in the prompt forcing the model to walk through each computational step explicitly before generating a conclusion. That reduces hallucination by a substantial margin because the model has to show its work.
A typical workflow takes about twelve to fifteen minutes from floor plan to structured report when you have a clean text input. If you are pasting OCR'd blueprints with garbled room labels, that jumps to forty-five minutes of cleanup. Worth noting: the prompt chain itself usually runs in under thirty seconds on most current models, but the bottleneck is always input preparation, not inference.
Setting Up the Core Prompt Structure
The prompt needs four sections minimum. Input geometry, orienting data, analytical framework selection, and output format specification. Here is the basic skeleton: First section: provide the building's sitting and facing degrees in a clear text statement. Include the construction year if known. This matters because Flying Star calculations reset every twenty years and the period determines which stars are active. If the owner does not know the build year, use the earliest major renovation as a proxy. Do not guess. Second section: specify the analytical school. Eight Mansions is simpler and works reliably for quick residential reads. Flying Star is more precise but requires accurate compass data and build period. Some users run both in parallel and compare. That usually surfaces contradictions worth investigating rather than ignoring.
Get the Full Details
Third section: define the output structure. Ask for sector breakdowns by degree, star ratings per sector, element associations, and recommended cures with elemental justification. Request the output as a table followed by narrative interpretation. Tables force the model to be specific. Narrative sections after the table are where hallucination creeps back in, so keep those brief.
My Experience With a Specific Edge Case
Last autumn I processed a townhouse in London that had been extended twice in different centuries. The ground floor was nineteenth century, the first-floor extension dated to 1972, and the rear kitchen annex was built in 2004. Each section had a different sit-fact orientation because the extensions were angled differently against the terrace. A single prompt run treated the whole building as one unit and produced a clearly wrong reading because the composite sitting direction averaged out to something that did not correspond to any actual door or window axis. The workaround was straightforward but not obvious if you have never dealt with composite buildings: run separate prompts for each structural section, assign each its own sitting and facing degrees, then merge the results by floor level rather than by sector. I learned this the hard way after the first read suggested the master bedroom sat on a wealth star that the actual compass reading contradicted. Once I segmented the prompts by architectural period, the output aligned with what the physical compass confirmed within two degrees. The merged report took about twenty extra minutes but was dramatically more accurate.
Advanced Nuance Most People Miss
Most prompt templates treat each directional sector as independent. They do not account for external form objects affecting nearby sectors simultaneously. A tall building to the northeast will suppress the Kun sector's energy, but it also creates a sha qi channel that affects the central palace if it is aligned with a main entrance. Standard prompts rarely surface this interaction because they are optimized for interior readings only. You need to add an explicit instruction for external feature mapping if your analysis should include it. Another subtlety: Flying Star calculations assume the property was built or substantially renovated during the period being analyzed. A house built in 1990 that was fully gut-renovated in 2018 operates under the 2018 Period 9 star chart, not the original 1990 chart. I see this mistake constantly in online forums. The prompt will happily compute the 1990 chart if you do not specify otherwise. Always verify the renovation date and override the default unless the structure is genuinely untouched.
When This Approach Completely Fails
It fails when the input data is vague or when the assignment requires nuanced environmental assessment that no text prompt can capture. I ran a prompt chain for a lakeside cabin last winter where the water body was directly in front but set back behind a tree line. The model classified the water as a wealth activating feature based on its directional position alone. It missed the filtering effect of the trees entirely. A human practitioner would notice the visual and energetic screening immediately. The AI could not. You need to manually correct that classification in the prompt by specifying the distance and obstruction details. It also fails poorly when clients expect health predictions or life outcome guarantees. The system can map stars to body parts and life areas using traditional correspondences, but those are symbolic mappings, not diagnostic tools. I had one client who became anxious because the prompt flagged the northwest sector with a illness star adjacent to the master bedroom. The reading was technically correct according to the formula, but the star combination was inactive due to the currentPeriod and the sector contained no actual door or bed placement. The anxiety was unwarranted, but the prompt output without the period context looked alarming. Always include period status in the output so the reader can distinguish active from inactive stars.
Downloadable Prompt Template
You can copy this structure and adapt it for your own use. I keep a master version updated with each model release because the reasoning quality shifts between versions. Build name and location: [insert] Construction year: [insert or unknown]
Last major renovation year: [insert] Sitting degree: [insert from compass or floor plan] Facing degree: [insert]
Analytical school: [Eight Mansions / Flying Star / both] Key features: [doors, windows, beds, stoves, toilets, water features, external structures within 100 meters] Output format: sector table with degree range, star rating, element, activity association, cure recommendation with elemental justification. Followed by brief narrative summary limited to three paragraphs. Flag any sectors where the star is in-period versus out-of-period.
Avoid making predictions about health outcomes or financial results. Stick to spatial analysis and traditional cure recommendations. That template runs cleanly on most current models. The output is usually structured enough to paste directly into a client report with minor editing. The editing step typically takes five to ten minutes depending on how many external features you need to add manually. If you skip the manual verification step, you are trusting the model to interpret physical descriptions accurately, and that introduces enough error to make the whole exercise questionable for professional work. Use it as a first draft generator, not a final authority. The real time savings comes when you have multiple units to analyze in sequence — apartment buildings, multi-tenant commercial spaces, or historical properties where you need to compare period charts. Running twenty prompts through this workflow takes about three hours instead of the six to eight hours a full manual analysis would require. The quality tradeoff is acceptable for preliminary assessments and internal reference. For client deliverables where accuracy is non-negotiable, I still verify every output with a physical compass and on-site observation before sending anything out.
