菜单
Single human-facing command for all Citum style work. Use whenever someone asks to fix, improve, convert, create, or tune a citation style. Routes to upgrade, migrate, create, or tune. Always use this rather than calling internal skills directly.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
High-throughput migration waves converting priority parent CSL 1.0 styles to Citum with repeatable before/after metrics and migration-engine gap recommendations. Fidelity is the hard gate.
Standardized QA gate for style work. Verifies fidelity, SQI drift, formatting defects, and regression surface. Produces approve/reject verdict with numbered findings.
Iterative LLM hand-tuning loop for embedded-core styles. Drives a style to 100% fidelity and clean SQI. Both are hard gates. Seeded from migrate evidence, not from the converter output as a terminal deliverable.
Autonomous research loop for Citum migration fidelity gaps and converter-focused Rust fixes.
Public entrypoint for Citum style work. Activate on: 'upgrade', 'migrate', 'create', 'tune', any style authoring request, or any request to fix/improve/convert a Citum or CSL citation style. Route to the shared workflow docs and internal roles.
Repo-owned test soundness auditor and navigator for citum-core. Reasons from a spec (not the implementation) to classify each test as good, suspicious, broken, or redundant; simultaneously reviews the spec for ambiguity, contradiction, and silence, halting to prompt the user when a spec defect blocks an honest verdict. Records state in the living ledger docs/architecture/TEST_SOUNDNESS_STATUS.md — never ephemeral JSON. Trigger on: "/test-soundness-review", "review my tests against the spec", "audit test soundness", "are my tests vacuous", "find broken tests", "find redundant tests", "are these tests just for the sake of it", "test quality review", "review the spec while you're at it", "what's the state of our test quality", "what test audits are outstanding", "what's next for test soundness". Also trigger proactively after a disambiguation, sorting, algorithm, or engine feature is implemented when the user asks "do the tests actually cover the spec?". Takes up to two arguments: /test-soundness-review [spec-pa
| name | style-evolve |
| type | user-invocable, agent-invocable |
| description | Single human-facing command for all Citum style work. Use whenever someone asks to fix, improve, convert, create, or tune a citation style. Routes to upgrade, migrate, create, or tune. Always use this rather than calling internal skills directly. |
| model | sonnet |
| routes-to | style-maintain, style-migrate-enhance, style-tune, style-qa |
/style-evolve upgrade <style-path>
/style-evolve migrate <csl-path>
/style-evolve create
/style-evolve tune <embedded-style-id>
Do not ask users to call internal skills directly.
Before doing anything else, read:
docs/policies/STYLE_WORKFLOW_DECISION_RULES.md — failure classification and stop conditionsdocs/guides/STYLE_WORKFLOW_EXECUTION.md — decision flow, evidence ladder, shared gatesdocs/guides/JJ_AI_CHANGE_STACK.md when .jj is present and local change-stack
curation would help isolate a style passRun the full pipeline without stopping to ask questions between steps. Only interrupt for the explicit permission gates in CLAUDE.md.
Route to ../style-maintain/SKILL.md.
Route to ../style-migrate-enhance/SKILL.md.
Build a new Citum style from source evidence. Escalate to @dplanner for design.
Route to ../style-tune/SKILL.md. Use when the target is an embedded-core
style and the goal is 100% fidelity and clean SQI. Both are hard gates for
this mode. The migrate output is the seed, not the deliverable.
Use docs/policies/STYLE_WORKFLOW_DECISION_RULES.md for failure classification and stop conditions. Keep the router focused on dispatch and host entrypoint behavior.
embedded-core styles (fidelity AND SQI both required).
For dependent styles, SQI is optimization-only after fidelity is stable.
See docs/policies/STYLE_WORKFLOW_DECISION_RULES.md for the tier definition and quality bar.docs/specs/STYLE_TAXONOMY.md.create pass for a hidden family root
followed by upgrade reduction of the public handles.benchmark_runs, run node scripts/report-core.js --style <name> after the primary oracle pass and treat the rich-input results as official supplemental evidence.../style-qa/SKILL.md before completion../scripts/check-docs-beans-hygiene.sh must pass.Every completed task delivers:
noneInclude a routed sub-skill header line in the final output.