terminology-normalizer（术语规范化器）

Terminology Normalizer

Purpose: make the draft read like one author wrote it by enforcing consistent naming (canonical terms + synonym policy), without changing citations or meaning.

Role cards (use explicitly)

Taxonomist (canonicalizer)

Mission: decide one canonical term per concept and a light synonym policy.

Do:

• Prefer taxonomy node names (outline/taxonomy.yml) as canonical labels when available.
• Define a short synonym policy only where readers expect it (use sparingly).
• Keep headings and tables aligned with canonical terms.

Avoid:

• Renaming proper nouns (paper titles, benchmark names, model names).
• Over-normalizing away meaningful distinctions (e.g., collapsing two different mechanisms into one word).

Integrator (apply without drift)

Mission: apply replacements consistently without changing meaning or citations.

Do:

• Keep replacements local and conservative; reread sentences that become ambiguous.
• Preserve citation placement and subsection boundaries.

Avoid:

• Introducing new claims while rewriting for terminology.
• Moving citations across subsections.

Role prompt: Terminology Editor (one voice)

You are normalizing terminology in a technical survey draft.

Your job is to make the draft read like one author wrote it by enforcing consistent naming.

Constraints:
- do not add/remove citation keys
- do not move citations across ### subsections
- do not introduce new claims while renaming

Method:
- pick a canonical term per concept
- define allowed synonyms (optional, minimal)
- apply consistently across headings, prose, and tables

Inputs

• output/DRAFT.md
• Optional (read-only context):

- outline/outline.yml (heading consistency) - outline/taxonomy.yml (canonical labels)

Outputs

• output/DRAFT.md (in place)
• Optional: output/GLOSSARY.md (short appendix/glossary table, if useful)

Workflow

Use the role cards above.

Steps:

1. Build a glossary candidate list from the draft (10–30 key terms):

• core objects (agent, tool, environment, protocol)
• key components (planner/executor, memory, verifier)
• evaluation terms (benchmark, metric, budget)

1. Choose canonical names and a synonym policy:

• one concept = one canonical term
• define allowed synonyms only when readers expect them (and use them sparingly)
• if outline/taxonomy.yml exists: prefer taxonomy node names as canonical labels (avoid inventing new names)
• if outline/outline.yml exists: keep section headings aligned with the same canonical terms

1. Apply replacements conservatively:

• do not alter paper names, model names, benchmark names
• keep terminology consistent across headings, prose, and table captions

1. Optional: write a small output/GLOSSARY.md:

• term | canonical | allowed synonyms | notes

Mini examples (what to do / what to avoid)

• Bad (term drift): tool API, tool interface, action schema used interchangeably without a rule.
• Better (canonical + light synonym policy): pick one canonical term (e.g., tool interface) and allow one synonym only when first introduced (e.g., tool interface (API contract)), then stick to canonical thereafter.
• Bad (over-normalization): replacing distinct terms so a contrast disappears.
• Better: keep distinct terms when they encode different mechanisms; normalize only spelling and naming consistency.

Guardrails (do not violate)

• Do not add/remove citation keys.
• Do not move citations across ### subsections.
• Do not introduce new claims while renaming.

Troubleshooting

Issue: normalization changes citation keys or moves citations

Fix:

• Revert; this skill must not add/remove keys or move citations across subsections.

Issue: synonyms policy is unclear

Fix:

• Define one canonical term per concept and list allowed synonyms; apply consistently across headings, tables, and prose.