Michael Polanyi

A writing skill inspired by Michael Polanyi's ideas of tacit knowledge, personal knowledge, and integrative judgment.

Use it when you want AI responses to sound less like a polished summary and more like an experienced practitioner: grounded, directional, bounded, and practically useful.

[!IMPORTANT]
This project is a prompt-engineering distillation inspired by selected ideas in Michael Polanyi's work.
It does not attempt to reproduce his philosophy in full, and it does not simulate Michael Polanyi as a person.

Quick start

Install the skill:

mkdir -p ~/.claude/skills
cp -R skills/michael-polanyi ~/.claude/skills/

Restart your client if it only loads skills at startup.

Then try a prompt like:

Use Michael Polanyi to answer this like someone who has done the work: this project keeps reworking the same features — should I fix process, documentation, or staffing first?

A good result should:

• lead with a judgment
• surface practical signals
• show boundaries and failure conditions
• end with one concrete next step

Why this skill exists

Many model answers are:

• correct but generic
• balanced but evasive
• polished but thin
• "deep" in tone but weak in decision value

This skill pushes answers toward:

• clearer overall judgment
• stronger practical signals
• better fact / interpretation / hypothesis separation
• visible trade-offs, limits, and failure conditions
• one concrete next step

Before / after

Weak answer

It depends on the situation. You should evaluate the process, team structure, and documentation quality, then decide based on the overall context.

Stronger answer in this skill's style

我的判断是：先别急着换人，也别先补大文档，先补一层最薄但必须执行的流程。项目总在返工，通常不是“没人努力”，而是目标、边界、验收和拍板责任没有在开工前钉死。先看三个信号：需求是否经常中途改口、不同角色是否理解不一致、做完后是否总出现“不是这个意思”。如果这三点里已经中了两点，问题更像流程前门失守，而不是单点执行力差。下一步建议：先硬上四个最小约束——唯一拍板人、单页需求边界、变更记录、每周短演示——连续看两个迭代，再决定是否要动关键岗位。

Who this is for

Use Michael Polanyi for:

• experienced advice
• practitioner-style critique
• trade-off analysis
• complex decision support
• ambiguous situations with incomplete information
• responses that should sound like they come from someone who has done the work
• prompts like “正确但很空”, “更像真正做过事的人”, or “avoid pseudo-depth”

Do not use it for:

• strict JSON or schema output
• deterministic extraction tasks
• exact code transformation
• legal or compliance wording
• tasks where rigid formatting matters more than judgment

Compatibility

This repository follows the common file-based skill layout used by Claude Code and Codex-style agent tooling.

• Claude Code: ~/.claude/skills/michael-polanyi/SKILL.md
• Codex-style setups: ~/.agents/skills/michael-polanyi/SKILL.md

The public title is Michael Polanyi.
The technical skill slug is michael-polanyi.

[!TIP]
In practice, this skill works best as an explicitly invoked judgment/writing skill.
Do not rely on blind auto-discovery for every prompt. Asking for Michael Polanyi by name is the most reliable path.

Install

Claude Code

mkdir -p ~/.claude/skills
cp -R skills/michael-polanyi ~/.claude/skills/

Expected result:

~/.claude/skills/michael-polanyi/
  SKILL.md
  examples.md
  polanyi-notes.md
  references/
  scripts/

Codex

mkdir -p ~/.agents/skills
cp -R skills/michael-polanyi ~/.agents/skills/

Expected result:

~/.agents/skills/michael-polanyi/
  SKILL.md
  examples.md
  polanyi-notes.md
  references/
  scripts/

If your client only loads skills at startup, restart the session after installation.

How to use it

Option 1: Name the skill directly

Ask for the skill by name when you want a more grounded judgment style:

• Use Michael Polanyi to evaluate whether this architecture should go to production.
• Use Michael Polanyi to critique this recommendation and tell me where it sounds inexperienced.
• Use Michael Polanyi to answer this with a clear judgment, practical signals, and one next step.

Option 2: Ask for the pattern without naming it

You can also request the response contract directly:

• Give me a directional judgment, not a balanced preamble.
• Separate facts, interpretations, and hypotheses.
• Surface the subtle practical signals an experienced operator would notice.
• Show trade-offs, failure conditions, and what would change your mind.
• Avoid pseudo-depth or fake wisdom.

30-second self-check

After installation, try one of these prompts:

• Use Michael Polanyi to answer this like someone who has done the work: the team is stuck and not shipping.
• Use Michael Polanyi to tell me why this answer sounds correct but inexperienced.

The output should usually contain:

• an upfront judgment
• subtle but practical signals
• explicit boundaries or flip conditions
• a concrete next step

Maintainer workflow

This repository also includes maintainer tooling under workbench/.
If you are iterating on the skill rather than just installing it, start with:

• workbench/README.md
• workbench/evals/protocol.md

The published GitHub Wiki is versioned in this repository under docs/wiki/.

Troubleshooting

The skill does not appear to load

Check:

• the directory name is exactly michael-polanyi
• SKILL.md exists at the top level of that directory
• the name field in frontmatter is also michael-polanyi
• your client has been restarted if it only loads skills at startup

The skill is installed but not automatically discovered

That can happen.
This repository is best treated as an explicitly invoked judgment/writing skill.
Use the skill by name when reliability matters.

The output sounds more literary, not more useful

That is drift, not success.
Re-check the response against the expected traits:

• clearer judgment
• better practical signal density
• sharper boundaries
• stronger next-step usefulness
• less pseudo-depth, not more

Repository map

skills/michael-polanyi/
  SKILL.md            # main runtime skill entrypoint
  README.md           # runtime package guide
  examples.md         # high-value before/after examples
  polanyi-notes.md    # conceptual grounding
  references/         # deeper runtime references
  scripts/            # runtime/lightweight quality helpers

workbench/
  README.md           # maintainer workflow
  evals/              # prompts, rubric, eval set, protocol
  agents/             # blind comparison and grading prompts
  scripts/            # assertion and benchmark helpers
  eval-viewer/        # HTML review generator

docs/wiki/
  *.md                # versioned source for the published GitHub Wiki

Evaluation

Do not judge this project by branding alone. Judge it by output quality.

Use:

• workbench/evals/prompts.md for prompt sets
• workbench/evals/rubric.md for scoring
• workbench/evals/protocol.md for the actual test method

A good "after" answer should become:

• more grounded
• more directional
• more bounded
• more useful

It should not merely become:

• more literary
• more philosophical
• more confident without grounds

Limitations

This project is:

• a prompt-engineering skill
• a judgment and writing pattern
• a practical distillation, not an academic reconstruction

This project is not:

• Michael Polanyi simulated as a persona
• a universal answer improver for all tasks
• a guarantee of automatic skill discovery
• a tool for strict structured output or compliance-sensitive wording

Language note

The repository README is written mainly in English.
Several examples and eval prompts are currently in Chinese because they were developed around Chinese judgment-style test cases.
The pattern itself is not language-specific.

Contributing

See:

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• SECURITY.md
• CHANGELOG.md

Why the name

The project uses Michael Polanyi as the public-facing name because the skill is explicitly inspired by Polanyi's ideas of tacit knowledge and responsible judgment.

The implementation is practical, not academic:
it translates selected ideas into prompt-engineering rules for more experienced-sounding, more useful output.