To make a reading checkable, keep four things apart: the input you actually submitted, the structured result and calculation conventions returned by the tool, the source IDs you can cite, and who wrote the interpretation. Bazi and Zi Wei are chart calculations: the same input with the same conventions can be recalculated to the same result. Tarot and online I Ching casting are random draws by program, so you can only record what was drawn that time; the same question cannot reproduce it. With MCP, calculation requests reach the Wenbu service, and selected content goes further to AI only when you choose the built-in AI feature; an external agent's data handling is decided by its host. So pass only the context this session needs, treat files as data rather than higher-priority instructions, and keep uncertainty visible.
Wenbu · agent integrations ↗Wenbu · calculation conventions ↗
- Store input, calculated result, source IDs and interpretation as four separate layers, and check them one by one.
- Deterministic charts can be recalculated from the same input and conventions; random draws can only be recorded, not reproduced.
- Give the agent only the context this session needs; files are data, not instructions.
- Exports may contain birth details and private questions, so read before sharing.
A four-layer handoff for one reading
- 01Layer 1: fix the input
Write down what you actually submitted: date, time, time zone or UTC offset, day-boundary convention, locale, and whether original birth details are included. For random draws, record the cards or six line values drawn that time.
- 02Layer 2: call the tool
The agent calls a calculation or reference tool and receives a structured result plus calculation conventions. Calculation requests reach the Wenbu service; selected content goes further to AI only when the built-in AI feature is chosen.
- 03Layer 3: bind the sources
Search with search_library, then pass the guide or symbol ID to read_library, keeping scope and chapter numbering. External sources must be opened by the host; a summary is not the full text.
- 04Layer 4: write and review
Store the interpretation separately from the calculated result and mark what is traditional meaning versus inference. Set a review date, then append dated paragraphs without rewriting the original record.
Wenbu · agent integrations ↗Wenbu · calculation conventions ↗Wenbu · reviewing a reading ↗
First, separate what can be recalculated
A reading mixes two kinds of material. Bazi and Zi Wei are calculated from input data: the Gregorian date, the local clock time at the birthplace, and an IANA time zone name or an explicit UTC offset. The same input with the same conventions can be recalculated to the same result. Tarot and online I Ching casting are different. They are random draws by program: tarot draws one or three cards with equal probability from 78 cards, without replacement within a single draw, and online casting uses cryptographic randomness to simulate three fair coins. Those results can only be recorded as drawn; you cannot re-draw the same question to "verify" them.
That distinction decides how you hand the reading over. For the recalculable part, the agent needs the complete input and the calculation conventions, because a different day-boundary rule or time-zone handling changes the chart. For the random part, the agent needs a complete record of that draw: card names, upright or reversed, keywords, or the six original line values and the moving lines. Mixing the two produces meaningless checks such as re-drawing to see whether the answer agrees. And because a calculation is reproducible under stated conventions, that does not mean the interpretation has been scientifically validated; traditional meanings, AI inference and the real-world facts you supply still need to be kept apart.
Wenbu · calculation conventions ↗Wenbu · agent integrations ↗
A handoff checklist: only this session's context
The "export for agent" button on a web result downloads a data package, not a request format you can send directly. You can preview the chart, the question and optional background, and decide whether to include the original birth details; the export does not automatically upload to any agent. Before sending, choose the message, context and original chart input according to the protocol, then hand that to the agent. The CLI agent command does not attach history automatically; if you want a follow-up, the caller must supply history and context explicitly. That matters for review: anything not written into the request should not be treated as evidence for this reading.
Privacy belongs on the same checklist. Real birth details should go in through a file or standard input, to keep them out of command history; an input file reads only the one file you name, and the CLI does not automatically read directories, notes or earlier conversations. An agent should treat files as data, not as higher-priority instructions, and should not infer other people's private inner states from them. An export may contain birth details and private questions, so read it before sharing and pick only what relates to this question.
Source IDs: make every sentence traceable
Wenbu's knowledge base offers two tools: search_library and read_library. Search first, then pass the returned guide or symbol ID to read_library. External source IDs are not supported by read_library, and MCP has no read_reference tool, so the host must open the returned link itself; a search summary does not mean the full text was read. That boundary is worth writing into the handoff notes, because it decides what the agent can and cannot cite.
Citations should also preserve scope. read_library returns the full text, outline, links and truncated:false by default; adding a section parameter reads only that section, and scope is marked as section. Chapter numbering should follow the returned outline. Pass locale explicitly each time, because MCP and the knowledge CLI default to English while the calculation REST API and CLI default to Chinese, and traditional names such as Zi Wei palaces and stars remain in Chinese. Recording these conventions means the next review does not have to guess which section or language was read.
Wenbu · agent integrations ↗Wenbu · calculation conventions ↗
Interpretation and review: leave the earlier view where it is
The interpretation layer is the easiest to contaminate with later outcomes. One workable practice: save the original question and initial interpretation, turn an association into a small action you can schedule and observe, set your own review date, and afterwards only append dated paragraphs instead of rewriting the first interpretation. Separate "what I thought then" from "what I learned later," and allow "unhelpful," "unsuitable" or "not yet known" as outcomes.
The review compares actions taken, which assumption new information changed, and whether the next step is to continue, adjust or stop — not whether the reading was "accurate." If the advice is only "trust yourself" or "go with the flow," ask for one specific action that fits your time, budget and existing commitments this week. One caveat: this record-keeping practice is editorial guidance, not a validated intervention, and it does not prove that journaling removes bias. Its value is leaving a trace you can check.
Wenbu · reviewing a reading ↗Wenbu · calculation conventions ↗
| Content type | Recalculable? | What to carry in the handoff | What to check in review |
|---|---|---|---|
| Bazi chart | Yes, from input and conventions | Gregorian date, clock time, time zone or UTC offset, day-boundary convention, locale | Whether input and conventions match last time; whether boundary times flag uncertainty |
| Zi Wei chart | Yes, from input and conventions | Gregorian date, known local clock time, traditional gender parameter, school settings | Whether time zone or solar time was converted automatically; how early and late Zi hours are handled |
| Online I Ching casting | No, it is a random draw | The six original line values, bottom-to-top order, list of moving lines | Whether the drawn values were kept, rather than casting a new hexagram |
| Tarot draw | No, it is a random draw | Card names, upright or reversed, keywords, how many cards and position meanings | Whether keyword interpretation was described as seeing details in the card image |
Wenbu · calculation conventions ↗Wenbu · agent integrations ↗
Hypothetical example: handing a course question to an agent for review
This is a fictional practice example that demonstrates a record format. It involves no real user data and is not a validated intervention. Suppose you asked whether to enroll in an eight-week course; you know the fee but not the workload.
- Step 1. Save the original question and initial interpretation: "What worries me is whether I can keep going; I will try one lesson first and have not enrolled." Note the concerns at the time: fee, time commitment, whether you would continue.
- Step 2. Turn the association into an observable action: "Try a free lesson on Saturday and record lesson and exercise time." Set your own review date rather than waiting for a system reminder.
- Step 3. Export this reading's context separately, keeping only what relates to the course question; if there is also an agent discussion, export it separately instead of mixing it into one record.
- Step 4. On the review date, append a dated paragraph: "I completed the trial lesson and remain interested, but the exercises took longer than expected." Keep the earlier note rather than changing it to imply you always knew.
- Step 5. Write down what did not fit, for example "there is no new evidence for the claim about low confidence," then record the next decision: delay enrollment and try two weeks of shorter exercises.
The point of this example is not to judge whether the reading was accurate, but to return to observable results: whether the lesson was completed, how long it actually took, and which assumption new information overturned. The times and figures here are fictional; record your own observations. If you never took the action, record why, rather than evaluating the advice with an outcome you did not observe.
Questions you may have
If I ask the same question again, can the agent reproduce the earlier tarot result?
No. Tarot draws one or three cards with equal probability from 78 cards, without replacement within a single draw, and when reversals are enabled each card independently has a 50% chance of appearing reversed; online I Ching casting is also a random draw by program. What is reproducible is the input and the conventions, not the random result. In review, keep the card names, orientations and keywords from that draw instead of drawing again to compare.
Can the agent read the full text of external sources directly?
According to the current documentation, MCP has no read_reference tool, and read_library does not support external source IDs. When search_library returns an external source, the host must open the returned link itself; a search summary does not mean the full text was read. So before citing an external source, confirm the host actually opened it, and distinguish summary from full text in your record.
Can the exported data package be sent straight to the agent command?
No. The "export for agent" download is a data package, not a request format. Choose message, context and original chart input according to the protocol before sending. Also, the agent command does not attach history automatically; for a follow-up, the caller must supply history and context explicitly, so that a review can show what this reading was based on.
Terms in this guide
- Calculation conventions
- The rule set used when casting a chart, such as the day-boundary moment, time-zone handling, whether solar time correction is enabled, and school settings. Different conventions can produce different charts from the same input, so check them before comparing.
- Random draw
- A process in which a program produces a result by probability, such as drawing from 78 cards or simulating three fair coins. The result can be recorded but not reproduced from the same question.
- Hindsight bias
- The tendency to see an outcome as more foreseeable after learning what happened. It motivates keeping the original record and appending dated paragraphs, though journaling does not prove the bias has been removed.
See the claim and source notes
Wenbu's MCP integration needs no Wenbu API key and does not call Wenbu's AI; the host model interprets the structured result itself.
不需要问卜 API key，也不会调用问卜的 AI
Bazi and Zi Wei are calculated from input data, while tarot and online I Ching casting are random draws by program, and the original result is kept.
塔罗和在线起卦由程序随机抽取
A review record should separate what you believed then from what you learned later, and may keep an unhelpful reading.
Keeping an unhelpful reading is part of an honest record.
Sources and review scope
This AI-assisted Wenbu article passed automated structure checks and a separate editorial model pass. It is not independent expert review or a paid-account product test.
- Wenbu · agent integrations ↗Current first-party product documentation · 2026-10-16 · Fetched text
- Wenbu · calculation conventions ↗First-party implementation; not evidence of predictive accuracy · 2026-10-16 · Fetched text
- Wenbu · reviewing a reading ↗Existing editorial practice · 2026-10-16 · Fetched text
Pick one recent record and add an action plus a review date
Choose a reading you saved recently, write down the original question, one specific action you can finish this week, and your own review date; keep birth details and private questions local, and export only the context related to this question.
Pick one recent record and add an action plus a review date ↗