Round one sounds fine, then answers drift. Don't hunt for a longer prompt—use three ways answers drift off-track—and four guardrails that stop it, and a copy-ready skeleton below.
In plain terms: the facts wander (the model cites chart details you never pasted), the method wanders (Zi Wei, BaZi, and Western rules get mashed together), and the question wanders (you asked about structure; it answers whether you should break up).
Prompts limit overreach—they don't make AI smarter
Reliable chart work has three parts:
- Calculation—time, calendar, rule set
- Data organization—complete, verifiable fields
- Interpretation—questions, reasoning, phrasing
Prompts constrain interpretation (and how chart data is used)—they cannot replace calculation or organization. For charting and verification, see the full AI reading workflow.
Why follow-ups go wrong
Answers skew over multiple turns when boundaries loosen—not because the model got dumber:
Data drift
The model cites fields not in your packet—inventing stars, cycles, or transits, or substituting early-chat impressions for pasted data.
Method drift
Zi Wei blends with BaZi, Western astrology, or pop psychology—mixed rules, uncheckable conclusions.
Question drift
You ask about palace structure; the model answers “is this the right relationship for me?” Scope widens; generic filler fills gaps.
A good prompt is not about making AI talk more—it is about making AI do fewer unsupported things.
Four boundaries a good prompt sets
A prompt does not add charting skill—it limits how existing data gets used. Examples use Zi Wei; swap terms for BaZi or astrology.
Data
Weak: “Analyze my chart.”
Better: “Use only the Zi Wei data I provide. Do not re-cast or add stars, transformations, or timing not present.”
System
Weak: “Read my chart.”
Better: “This is Zi Wei—do not mix BaZi or Western astrology rules.”
Reasoning
Weak: “Analyze my love life in detail.”
Better: “Spouse Palace and related palaces only. List fields used first. No specific partner, marriage date, or cycles not in the packet.”
Question
Weak: “Cover career, love, and the next three years.”
Better: “Career direction only. If annual or major-limit data for a year is missing, do not analyze that year.”
Use a skeleton, not a magic prompt
The four boundaries fold into one copyable skeleton. It does not ask AI to “play astrologer”—it defines which data to use, how to combine signals, how to separate time layers, handle conflicts, and when to stop.
Paste your chart data first, then copy the block below—replace only the bracketed discipline, timing layers, and question. List timing layers that match what you actually pasted.
[System & data scope] The following data belongs to: [discipline name]. Use only that system's rules, terms, and logic. Do not mix other chart systems, astrology, or generic psychology. This packet includes these timing layers: [list: natal / luck cycles / major limits / annual / monthly / transits / progressions, etc.] Do not infer or invent timing layers that were not provided. [This question] I want to understand: [specific question] Scope for this reply: [topic, time range, or subject] Do not answer outside this scope. [Analysis principles] Do not conclude from a single field, symbol, or placement alone. Major claims should rest on two or more related signals when possible; if a claim relies on one signal only, state that limitation. Identify data directly relevant to this question first, then analyze relationships—not a field-by-field tour of the whole chart. [Time layers] When both long-term structure and timing data are present, separate them: 1. Long-term, stable tendencies; 2. Themes amplified in a given phase; 3. Triggers or shifts at specific points in time. Do not turn long-term tendencies into dated events, or let short-term signals override overall structure. [Evidence vs inference] Label each part as: 1. Facts stated in the data; 2. Synthesis from multiple related fields; 3. What the data cannot confirm. For each important claim, cite which fields support it— not conclusions alone. [Conflicting signals] If signals contradict, keep both and explain: - Different layers they may act on; - Long-term vs short-term difference; - Different conditions or expressions; - When dominance cannot be decided yet. Do not merge into one neat answer for completeness. [Uncertainty & stop conditions] If the data cannot support a specific claim, state: - How far you can go with what exists; - What is missing; - What must not be inferred further. Do not invent fields, dates, people, or events not in the packet.
Matches Astralium's in-app reading instructions: constrain usage and reasoning, not preset judgments. You can keep only the [System & data scope] and [Uncertainty & stop conditions] blocks if you already have your own question format.
Stable data before the prompt
Prompts shape reading—they cannot fix a wrong or incomplete chart. Astralium handles charting before that step: engine charting and structured, copyable packets—not fixed conclusions, but one shared starting point for any model. See our brand stance, the packet explainer, and the same-birth charting comparison.
Cast → copy data → add question and boundaries → any AI
Accurate charts. Open interpretation.
Chart first, then paste a packet with clear boundaries—not a longer prompt.
Create a free chart →