paper-writing-assistant

# Paper Writing Assistant Use this skill as the central controller whenever the user wants to draft, revise, evaluate, or distill an academic paper section in a specific author's writing style. ## Capabilities - **Write**: Draft a specific section using a selected author's style. - **Revise**: Rewrite an existing section to match a selected author's style. - **Evaluate**: Score a section against the author's style rubric. - **Distill**: Run the factory pipeline to create a new author style from PDFs. ## Project Base Path The PaperSkill project is located at: - WSL: `/mnt/g/project/PaperSkill/` - Windows: `G:\project\PaperSkill\` All operations must load `registry.json` from the project base first. ## Workflow ### Step 1: Determine Intent Parse the user's request into one of: - `write` - `revise` - `evaluate` - `distill` If unclear, ask: "你是想写/改/评某个 section，还是蒸馏新文风？" ### Step 2: Determine Author 1. Load `registry.json` from the project base. 2. Check if the user explicitly mentions an author name or ID. - Known aliases: "唐旭" → `xu-tang`, "Xu Tang" → `xu-tang`. 3. If no author is specified, use `registry.json["default_author"]`. 4. Validate that the author exists in `registry.json["authors"]`. - If not found, list available authors and ask the user to choose. ### Step 3: Determine Section Attempt to infer the section from the user's message. Supported mappings: | Section | Keywords | |---------|----------| | abstract | abstract, 摘要 | | introduction | introduction, 引言, 简介 | | related_work | related work, 相关工作, 文献综述 | | methods | methods, 方法, 方法论 | | results | results, 结果, 实验结果 | | discussion | discussion, 讨论 | | conclusion | conclusion, 结论, 总结 | If multiple sections are mentioned, ask which one to proceed with. If no section is detected and intent is `write`/`revise`/`evaluate`, ask: "请告诉我你要处理的 section（如 introduction / methods / results 等）。" ### Step 4: Load Assets For the resolved `(author, section)` pair, load: - Skill rules: `authors/<author_id>/skills/skill_<section>.json` - Prompt template: `authors/<author_id>/prompts/prompt_<section>.json` - Rubric (for evaluate): `authors/<author_id>/rubric.json` If files are missing, fall back to the author-style sub-skill references: - `hermes-skills/author-styles/<author_id>-paper-style/references/section-skills.md` - `hermes-skills/author-styles/<author_id>-paper-style/references/prompts-and-rubric.json` ### Step 5a: Write / Revise 1. Summarize the user's provided facts, figures, tables, and claims. 2. If facts are insufficient, mark missing items instead of inventing them. 3. Build a short outline following the loaded skill's `structure` and `logic_pattern`. 4. Output the outline first, then the draft. 5. For revise tasks, highlight what was changed to match the author's style. ### Step 5b: Evaluate 1. Load the rubric JSON. 2. Score the provided text across the 6 dimensions (1-5 scale). 3. Compute the average score. 4. Provide specific improvement suggestions for any dimension scoring ≤3. ### Step 5c: Distill When the user asks to distill a new author style, use the **two-phase workflow** below. Hermes acts as the LLM for the enrichment phase; no external LLM API configuration is required. #### Phase 1: Skeleton Generation (terminal) 1. Ask the user for the new author's ID, display name, and domain. 2. Verify that `authors/<new_id>/papers/` contains PDFs or extracted text files. 3. If the aggregation report does not exist yet, instruct the user to run their existing analysis scripts first (e.g., place extracted texts in `authors/<new_id>/extracted_texts/` and run the analyzer). 4. Run the skeleton pipeline: ```bash python factory/run_pipeline.py --author <new_id> --no-llm ``` This generates stable skill skeletons, prompts, rubric, and Hermes references without LLM enrichment. #### Phase 2: Hermes Enrichment (Hermes direct) 5. For each of the 7 sections, load: - `authors/<new_id>/analysis/aggregation_report.json` - `authors/<new_id>/skills/skill_<section>.json` 6. Read the relevant parts of the aggregation report (surface style, syntactic style, rhetorical structure for the section, academic stance). 7. **Use your own reasoning to generate**: - `preferred_phrases`: 5-8 complete sentence templates or starter phrases reflecting the author's style - `logic_pattern`: 1-3 strings describing the rhetorical flow 8. Write the enriched fields back into each `skill_<section>.json`. #### Phase 3: Rebuild (terminal) 9. Run: ```bash python factory/rebuild_after_enrich.py --author <new_id> ``` This rebuilds prompts, rubric, and Hermes references from the enriched skills. #### Phase 4: Registry Update 10. Append the author metadata to `registry.json["authors"]`. 11. If the user says "设为默认" or "set as default", update `registry.json["default_author"]`. **Example request:** "我想蒸馏一个新导师的风格，论文已经放在 authors/prof-li/papers/ 里了。" - Response: "好的。请告诉我这位导师的显示名称和研究领域（例如 prof-li / 计算机视觉）。接下来我会先跑骨架脚本，然后由我直接为每个 section 填充文风特征，最后同步生成 prompts 和 references。" ## Output Discipline for Drafting 1. **Outline first**: Always present the section outline before the body text. 2. **No hallucination**: Do not invent citations, datasets, numbers, or code links. 3. **Section fidelity**: Do not write content belonging to other sections. 4. **Style proximity**: Match rhetoric, progression, and claim strength; do not copy source wording verbatim. 5. **Missing info**: If critical inputs are absent, explicitly list them as `[待补充: ...]` instead of filling them in. ## Registry Update Rule Whenever a new author style is successfully distilled: - Append the author metadata to `registry.json["authors"]`. - If the user says "设为默认" or "set as default", update `registry.json["default_author"]`. ## Example Interactions **User**: "用唐旭老师的风格帮我写 Introduction，关于遥感变化检测的。" → Intent: write, Author: xu-tang, Section: introduction. Load skill, ask for facts if missing, then output outline + draft. **User**: "把这段 results 改成唐旭风格。" → Intent: revise, Author: xu-tang, Section: results. Load skill, rewrite, highlight changes. **User**: "评估一下这段 abstract 符不符合唐旭的风格。" → Intent: evaluate, Author: xu-tang, Section: abstract. Load rubric, score, give feedback. **User**: "我想蒸馏一个新导师的风格，论文已经放在 authors/prof-li/papers/ 里了。" → Intent: distill, Author ID: prof-li. Run pipeline, update registry.