AI agent skill · by @yaojingang

Yao Bayesian Skill

Convert uncertain real-world choices into an auditable Bayesian evidence-to-action report with priors, evidence grading, posterior update, action thresholds, sensitivity checks, multi-turn decision logs, and Markdown plus bilingual HTML output. Do not use for Bayes theorem tutoring, homework, generic brainstorming with no report, or final licensed medical, legal, or financial advice.

Archived 2026-10-23 · Source updated 2026-08-28 · License: MIT. Check the original repository for the latest version.

SKILL.md instructions

# Yao Bayesian Skill

## Use This Skill For

- structure a vague choice into hypothesis, time horizon, success metric, and actions
- set a prior, grade evidence, update posterior, compare action thresholds, and recommend next information
- start from incomplete input with a weak prior, then improve the judgment through multi-turn questioning
- export one synchronized Chinese-first `markdown` plus bilingual `html` report

## Do Not Route Here

- Bayes theorem tutoring or homework-only calculations
- broad research or brainstorming with no explicit decision report
- final professional medical, legal, or investment advice

## Default Workflow

1. Use `references/intake-contract.md` to convert the request into one structured decision brief.
2. If input is incomplete, read `references/multi-turn-dialogue-loop.md`; start with a weak prior and ask the minimum next questions.
3. Use `references/evidence-prior-playbook.md` to grade evidence and choose the lightest valid update path.
4. Run `references/prior-hygiene-checklist.md`; show only the 3-5 principles most relevant to this case.
5. Maintain the round log: user input, remaining gap, update path, probability change, and decision readiness.
6. Run `scripts/bayesian_decision_report.py` for canonical JSON or `scripts/generate_report_bundle.py` for `markdown + html`.
7. Finalize with `references/decision-report-contract.md`, `references/report-export-pipeline.md`, and `references/sensitivity-and-safety.md`.

## Iteration And Implementation Constraints

When extending this skill: state assumptions before coding, keep the smallest valid workflow, touch only files required by the request, and define user-visible success checks before editing. Typical checks: incomplete input yields a weak prior plus follow-up questions; each round is logged; the report explains belief changes; HTML/Markdown still render the intended guidance.

## Output Contract

- Produce a decision report, not a formula dump; mark numbers as observed, estimated, or assumed.
- Put the plain-language conclusion and action recommendation before technical sections.
- Include weak evidence, dependence risk, sensitivity, prior-hygiene checks, and high-risk disclaimers when relevant.
- For multi-turn use, log prior, posterior, readiness, gaps, and formula/update path for each round.
- Reports default to Simplified Chinese; HTML also supports Chinese/English switching, sticky navigation, collapsible advanced sections, and top-right `Print` / `Save as PDF`.
- Printing or saving HTML as PDF should expand folded sections first.

## Reference Map

- `references/intake-contract.md`: request-to-brief conversion
- `references/multi-turn-dialogue-loop.md`: incomplete-input handling and iterative questioning
- `references/evidence-prior-playbook.md`: evidence tiers, priors, update-path selection
- `references/prior-hygiene-checklist.md`: default judgment priors for checking priors, evidence, and action intensity
- `references/decision-report-contract.md`: required report sections and schema alignment
- `references/report-export-pipeline.md`: automatic HTML/Markdown generation and bilingual HTML rules
- `references/sensitivity-and-safety.md`: sensitivity analysis and high-risk disclaimers
- `scripts/bayesian_decision_report.py`: canonical V0/V1 calculation
- `scripts/generate_report_bundle.py`: Chinese-first Markdown plus bilingual HTML