Zhuzimap is an auditable Agent Skill that first secures decision-critical information, then maps a modern trade-off to source-backed Chinese philosophical cognitive models and turns disagreement into an actionable, testable, and revisable next step.
Skill ID and explicit invocation: $zhuzimap.
Repository: github.com/952JO/zhuzimap
Release: v0.3.0
This is not historical role-play and it does not force every thinker to speak. The Skill first checks whether a council is useful, then applies an information sufficiency gate. If critical facts are missing, it asks 2–5 focused questions and stops before model selection. Only after the gate passes does it decompose the problem and select models that change the reasoning, boundaries, or action. Direct factual questions exit the council; philosophy never replaces professional evidence in high-risk domains.
The current V0.3 release contains 60 models across 15 thinkers. V0.3 adds the information-intake gate; the V0.2 thinker and model set is unchanged.
Codex and Cursor both support project-level .agents/skills/. Run this from
the target project's root:
$source = Resolve-Path ".\skills\zhuzimap"
$targetRoot = ".agents\skills"
$target = Join-Path $targetRoot "zhuzimap"
if (Test-Path -LiteralPath $target) {
throw "Destination exists; compare or back it up first: $target"
}
New-Item -ItemType Directory -Path $targetRoot -Force | Out-Null
Copy-Item -LiteralPath $source -Destination $target -RecurseWhen installing into another project, replace $source with the absolute path
to this repository's Skill directory. Start a new agent task or reload the host
window after installation.
For a global install, set $targetRoot to
Join-Path $env:USERPROFILE ".agents\skills". Codex still supports the legacy
~/.codex/skills location, but .agents/skills is the preferred cross-host
location. The directory follows the
Agent Skills specification.
Cursor 2.4 officially added Agent Skills
support to its editor and CLI.
Use $zhuzimap. I am considering an internship.
Ask for enough decision-critical information before selecting thinkers.
Only begin the council after the information gate passes.
Hosts may also select the Skill automatically, but that behavior depends on the host and model. Explicit invocation is recommended for important decisions.
flowchart LR
A["Modern problem"] --> B{"Council applicable?"}
B -- "Direct fact/tool query" --> C["Exit and answer or verify directly"]
B -- "Contested trade-off" --> D{"Information sufficient?"}
D -- "No" --> E["Ask 2-5 decision-changing questions"]
E --> D
D -- "Yes" --> F["Facts / gaps / assumptions"]
F --> G["2-5 decision dimensions"]
G --> H["Select 2-4 thinkers by model"]
H --> I["Independent judgments"]
I --> J["Compare and synthesize"]
J --> K["Reversible action + evidence + stop/escalate rule"]
The council does not vote. It distinguishes:
- same conclusion, same reason;
- same conclusion, different reasons;
- same goal, different paths;
- different conditions;
- surface conflict with deeper complementarity;
- irreducible value conflict.
| Thinker | Four current models |
|---|---|
| Laozi | limits to expansion; reducing forced control; flexible strength; starting with the small |
| Confucius | rectifying names; harmony without conformity; resisting haste and small gains; calibrating excess and deficiency |
| Mozi | testing claims with three standards; stopping costs without benefit; mutual benefit; capability over status |
| Sunzi | comparing conditions before commitment; securing against defeat first; limiting prolonged campaigns; adapting to changing form |
| Wang Yangming | unifying knowing and doing; practice in real affairs; conscience with external experience; commitment that governs trade-offs |
| Zhuangzi | questioning fixed perspectives; following structure rather than force; preserving unused capacity; respecting finite attention |
| Mencius | securing conditions before demands; protecting shared purpose from local profit; public welfare above office; cultivating without forcing growth |
| Xunzi | building capability through accumulation; shaping behavior through institutions; separating conditions from response; correcting single-angle blindness |
| Han Feizi | applying rules across status; matching claims to performance; designing beyond personal trust; closing small control gaps |
| Guan Zhong / Guanzi | securing material preconditions; aligning policy with lived needs; calibrating plans to capacity; long-horizon capability |
| Zhu Xi | accumulating inquiry before synthesis; learning in sequence; testing competing interpretations; depth over breadth |
| Zhang Zai | relational responsibility; learning through doubt; mapping opposition toward resolution; transforming adversity without glorifying harm |
| Wang Fuzhi | grounding names in facts; renewal through continuity; integrating knowing and doing; grounding principles in concrete things |
| Huang Zongxi | office as public trust; mission over patron; institutions before heroes; independent public review |
| Dai Zhen | distinguishing needs from excess; grounding principles in lived facts; testing rules by reciprocity; preventing power from monopolizing reason |
See the full model index and source-backed model cards.
- Information intake: known facts, decisive gaps, and 2–5 focused questions. It contains no thinker selection or substantive conclusion.
- Concise: judgment, key dimensions, disagreement, next action, stop rule.
- Standard: adds facts/gaps/assumptions, model selection, strongest objection, reversal conditions, and uncertainties.
- Traceable: adds primary text, section, short quotation, modern interpretation, and current-case inference.
Ask for “concise mode” or “trace every model to its source” in the prompt.
- Should a personal tool become a platform?
- Should a low-traffic internal system become microservices?
- How much production authority should a code agent receive?
These are actual frozen-evaluation outputs, not idealized examples written separately for the README.
- Quotation marks are reserved for text verified in a loaded model card.
- Sunzi is not a license for deception or adversarializing ordinary cooperation.
- Wang Yangming's liangzhi is not personal intuition or factual evidence.
- Numbers must come from user facts, explicit calculations, reliable sources, or be labeled as provisional thresholds requiring calibration.
- Medical, legal, financial, safety, and other high-risk conclusions remain governed by current evidence, controls, and qualified judgment.
python -m pip install -r requirements-dev.txt
python -m unittest discover -s tests -v
python skills\zhuzimap\scripts\validate_models.py --jsonThe historical V0.1 frozen cases are in evals/cases.yaml, with the rubric in evals/rubric.md. The 10-thinker expansion has 20 additional representative/misuse cases in evals/v0.2-expansion-cases.yaml. The V0.3 information gate has dedicated multi-turn and bypass-pressure cases in evals/v0.3-information-gate-cases.yaml. The repository retains ordinary baselines, anonymous A/B pairs, blind reviews, and a reproducible unblinding summary. Internal release thresholds are engineering gates, not statistical claims.
In one frozen V0.1 evaluation, blind reviewers found non-redundant incremental value in 15/15 positive cases; all 5 direct-fact controls exited the council; forced mapping was 0/20; and both high-risk cases preserved evidence and professional boundaries. Four thinkers were conservatively distinguishable across multiple cases. Wang Yangming appeared in only one positive full-suite answer, so stable distinctiveness is not claimed. See the full evaluation summary for method and limitations.
Those V0.1 scores cover only the original five thinkers. They are not evidence that the ten new thinkers are runtime-distinctive. V0.2 currently has source/schema checks and dedicated forward-test cases; a comparable frozen runtime blind evaluation remains a separate release gate.
The host validation record documents a fresh Codex CLI install, explicit invocation, protocol/model-card reads, and a first result in 107.96 seconds. Runtime validation on a second host is still pending.
Build a reproducible distributable ZIP:
python scripts\package_skill.py `
--skill-dir skills\zhuzimap `
--version 0.3.0 `
--output-dir distThe command prints the absolute archive path, file count, and SHA-256 digest.
The archive preserves the top-level zhuzimap/ directory.
- Skill is not visible: verify the exact path is
.agents/skills/zhuzimap/SKILL.md, with no extra nested copy, then start a new task or reload the host. - Automatic invocation did not occur: invoke it explicitly with
Use $zhuzimap. .... - Explicit invocation still fails: confirm that
SKILL.mdis uppercase and the folder matches the frontmattername, then run the validators below. - Windows validation is garbled: run
quick_validate.pywithpython -X utf8. - Model validation fails: fix the reported YAML indentation, missing field, duplicate ID, or non-HTTPS source and rerun the command.
- Host version is too old: check that host's current Agent Skills support. Format compatibility does not prove runtime discovery is enabled.
For a project-level install, resolve and inspect the exact target before deleting it:
$skillPath = Join-Path (Get-Location) ".agents\skills\zhuzimap"
Resolve-Path -LiteralPath $skillPath
Remove-Item -LiteralPath $skillPath -RecurseFor a global install, set $skillPath to
Join-Path $env:USERPROFILE ".agents\skills\zhuzimap".
Start a new agent task afterward. Always inspect Resolve-Path; the command
must target an installed copy, not this repository's skills/ source.
- Product charter
- Competitive and mechanism boundary
- Model card guide
- New model template
- v0.2 expansion audit
- v0.3 information-gate audit
- v0.1.0 completion audit
- Release notes
- 中文发布说明
- Contributing
Licensed under the MIT License.