scholar-distill
Build source-supported types, memberships, and controlled vocabulary from a term table
- License 许可
- MIT
- 所属仓库
- tony/skills
- 最近更新
- 2026-11-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- scholar-distill
- metadata
- { "source": "plugins/scholar/skills/distill/SKILL.md", "argument-hint": "[--out=<dir>]" }
- description
- Build source-supported types, memberships, and controlled vocabulary from a term table
- allowed-tools
- [ "Bash", "Read", "Grep", "Glob", "Edit", "Write", "AskUserQuestion" ]
- disable-model-invocation
- true
02 资源文件 6
- 参考文档agents/openai.yaml288 B
- 参考文档references/evidence-tiers.md1.7 KB
- 参考文档references/output-contract.md4.0 KB
- 参考文档references/prose-rules.md4.5 KB
- 参考文档references/stage-gates.md4.2 KB
- 脚本references/term_metrics.py47.8 KB
03 SKILL.md 正文
this skill
Stage 3. Turn a flat term table into an ontology. Three layers, in order. Any one of them alone is not an ontology.
Read references/evidence-tiers.md for what each claim must carry and
references/stage-gates.md for this stage's exit condition. The analyzer
is references/term_metrics.py; <analyzer> below is its absolute path
resolved from this skill's directory.
User arguments: $ARGUMENTS
Read references/output-contract.md before writing study files.
Layer 1: types
The is-a hierarchy. What kinds of thing the corpus contains.
Write it to evidence/hierarchy.md, citing each type declaration and is-a
claim. Each type is also a row in terms.jsonl with kind: type, its
corpus-supported definition, and parent set to another type or the literal
string root.
Write each type's name onto its members' rows as parent. A type row you create
carries "schema": 1 and no cites; the gate exempts it
(references/stage-gates.md), and the citations live in hierarchy.md. A type the corpus
already names (a class) keeps the row extract wrote: change its kind from
class to type in place, cites included.
Each non-root parent needs a source-supported is-a or member relation.
Sharing a file, directory, or prefix does not establish that relation.
root is the special anchor for top-level types, not an invented type.
Layer 2: supported memberships
Record the corpus's basis for membership in evidence/hierarchy.md. A code
declaration supports a named type even when the corpus gives no behavioral
membership test. Keep its documented role separate from that declaration.
Use a necessary-and-sufficient discriminator only when the source establishes both directions. A docstring describing a type's role does not establish that everything with that role belongs to the type, or that its members have no other roles. Keep the extracted definition and citations; do not strengthen a role description into such a test.
An analyst-proposed type without a supported membership relation remains unsupported. Record the gap rather than inventing a type or parent.
Layer 3: the controlled vocabulary
Write evidence/vocabulary.md: the corpus's terms, their senses, preferred
spellings, observed aliases, and any source-declared deprecations.
Preserve supported synonymy, multiple senses, and overlap. Cite each asserted
equivalence or disjointness constraint; sibling placement does not imply
disjointness. When the source leaves a relationship unknown, record that
limit without asserting either overlap or exclusion. Lexical similarity
proposes a question for contest, not a semantic relation.
The preferred term is the one the corpus already uses most, not the one that reads best. A vocabulary that renames what the corpus says has stopped describing it.
The orphan gate
Run the term gate and the hierarchy gate from
references/stage-gates.md; both must exit zero:
$ python3 <analyzer> notes/ontology/<subject>/ --check-terms
$ python3 <analyzer> notes/ontology/<subject>/ --check-hierarchy
Fix each cycle by re-reading the corpus's own is-a claim. An orphan may be a valid term whose parent the corpus does not establish. Retain it and report the failed gate. Do not drop it or promote it without source support to clear a finding. Never invent a type to house an orphan or split a type to tidy its fan-out.
Rules
- A term keeps the spelling
extractrecorded. The vocabulary names a preferred term; it does not rewrite the rows. - Deprecated variants stay in the table. They are what a reader of the corpus will actually encounter.
Output
Per references/output-contract.md. Hero: ✓ <n> types, <n> terms, <n> preferred terms, ⚠ Gate failed: <n> orphans, <n> cycles, or ⚠ Blocked: <reason>. A failed gate lists each orphan or cycle on its own line, then one
sentence naming the terms whose corpus evidence must be re-read before
assigning parents (no quoted corpus text, not even in parentheses):
⚠ Gate failed: 1 orphan, 1 cycle
- Orphan: checkpoint
- Cycle: phase -> stage -> phase
Re-read the corpus evidence for checkpoint, stage, and phase before assigning parents, then rerun the gate.
Sections:
## Types: the hierarchy and each type's supported membership basis.## Vocabulary: preferred terms with their deprecated variants.## Unresolved: terms that resisted placement, and what that suggests about the corpus rather than about the table.
Next-step options: run contest, revisit a type, stop here.
Write it per references/prose-rules.md.
Portability notes
$ARGUMENTS— the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.- Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.
04 同仓库更多 Skills
serene-bliss
Use when asked to make a CLI --help, docs page, error message, or tooling UX more serene or calmer for developer experience; returns text, no edits.
review
Weave code review: run independent adversarial reviews in parallel, then synthesize findings
refine
Use when asked to refine a draft, iterate on code until cleaner, or polish a README through repeated critique and improvement.
prompt
Weave prompt: compare independent adversarial implementations, then adopt the single best one
plan
Weave planning: compare independent adversarial plans, then synthesize the best one
fix-review
Fix weave review findings: validate, add test coverage, fix, and commit each as atomic changes