feynman-style

Files (56)

• .gitignore
• SKILL.md
• ai-vocabulary-supplement.json · assets/
• .gitkeep · assets/feynman-corpus/
• index.json · assets/feynman-corpus/
• feynman-delta-profile.json · assets/
• feynman-rules.json · assets/
• .gitkeep · assets/system-prompts/
• pedagogical-walkthrough.md · assets/system-prompts/
• popular-science.md · assets/system-prompts/
• technical-exposition.md · assets/system-prompts/
• conftest.py
• pyproject.toml
• .gitkeep · references/
• before-after-examples.md · references/
• feynman-corpus-map.md · references/
• feynman-style-guide.md · references/
• feynman-vitality-guide.md · references/
• negative-triggers.md · references/
• __init__.py · scripts/
• build_delta_profile.py · scripts/
• delta_math.py · scripts/
• lint_ai_vocabulary.py · scripts/
• lint_common.py · scripts/
• lint_concreteness.py · scripts/
• lint_conversational.py · scripts/
• lint_curiosity_markers.py · scripts/
• lint_latinate_diction.py · scripts/
• lint_reading_grade.py · scripts/
• lint_sentence_rhythm.py · scripts/
• pdf_extract.py · scripts/
• preserve_argument.py · scripts/
• score_feynman_delta.py · scripts/
• sibling_skills.py · scripts/
• style_pass_report.py · scripts/
• system_prompt_loader.py · scripts/
• skill_api.py
• __init__.py · tests/
• conftest.py · tests/
• .gitkeep · tests/fixtures/
• russell_output.md · tests/fixtures/
• test_build_delta_profile.py · tests/
• test_corpus_index.py · tests/
• test_lint_concreteness.py · tests/
• test_lint_conversational.py · tests/
• test_lint_curiosity_markers.py · tests/
• test_lint_latinate_diction.py · tests/
• test_lint_reading_grade.py · tests/
• test_ports.py · tests/
• test_preserve_argument.py · tests/
• test_rules_schema.py · tests/
• test_score_feynman_delta.py · tests/
• test_skill_api.py · tests/
• test_skill_integration.py · tests/
• test_style_pass_report.py · tests/
• test_system_prompt_loader.py · tests/

SKILL.md (first lines)

# feynman-style

Second pass after `russellian-style`. Russell makes prose correct, atomic, and hedge-free but dense; this skill warms the surface so it clicks for a reader. The argument is fixed by Russell and must survive untouched.

## The two-layer contract

- **Feynman owns the prose surface.** It may re-introduce what Russell strips: rhetorical questions, asides, contractions, direct address, "now you might ask—", honest-doubt framing.
- **Russell owns the argument skeleton.** Feynman must not change claims, claim accuracy, logical structure, or atomic argument order. `preserve_argument` enforces this as a hard gate.

## What it owns

- Surface readability: analogy, concreteness, reading grade, conversational diction, curiosity.
- The Feynman voice prompts and rule registry.
- The Surface/Integrity linter partition (which Russell checks may run on Feynman-final text).
- `preserve_argument` — the claims/structure survival gate.

## What it does NOT own

- Source ingestion, claim ledgers — `book-knowledge`.
- Chapter drafting, release bundles — `book-compose`.
- The Russell discipline itself — `russellian-style`.
- Generic AI tells (em-dash overuse, rule-of-three) — `humanizer`.

## Linters

- `lint_reading_grade.py` — Flesch-Kincaid grade load; flags too-hard sentences.
- `lint_conversational.py` — rewards direct address, contractions, rhetorical questions.
- `lint_latinate_diction.py` — flags Latinate jargon with a plain Anglo-Saxon alternative.
- `lint_concreteness.py` — analogy markers + concrete-instance density; flags ungrounded abstraction.
- `lint_curiosity_markers.py` — rewards honest-doubt / puzzle framing.
- `lint_sentence_rhythm.py` — cadence/burstiness, ported from russellian-style (Feynman threshold recalibration deferred).
- `lint_ai_vocabulary.py` — AI-slop vocabulary guard (ported).

## Calibration

Generation is prompt-driven. Linters use absolute thresholds from `assets/feynman-rules.json`. An optional relative-frequency distance scorer (`score_feynman_delta.py`) compares against `assets/feynman-delta-profile.json` using an L1 distance over word frequencies; the profile is hand-set by default and can be rebuilt offline by `build_delta_profile.py` from copies the user drops in a git-ignored `corpus-raw/` folder. The skill ships and runs with zero network access.

From CharlesHoskinson/russellian-book-suite, MIT.

More from CharlesHoskinson/russellian-book-suite