Paper Expert Generator

Paper Expert Generator

Generate a complete, ready-to-use domain-specific paper expert agent by adapting the PaperClaw architecture for any research field.

Workflow

Step 1: Domain Interview

Collect these details from the user before generating anything. Ask conversationally – do not dump all questions at once. Start with the most critical ones:

Critical (ask first):

1. Research domain – Primary field (e.g., "bioinformatics", "quantum computing", "computer vision")
2. Core topics – Specific sub-areas or problems (e.g., "protein folding, drug discovery, single-cell sequencing")
3. Key methods/techniques – Central methodologies (e.g., "transformers, GNN, diffusion models, RL")

Important (ask second): 4. Evaluation priorities – What dimensions matter most for paper quality in this domain? 5. Exclusion topics – What should be filtered out? (e.g., "finance, social media, NLP") 6. Output location – Where to create the agent? (default: ~/agents/<domain-slug>/)

Optional (ask only if needed): 7. Notification channel – Feishu/Lark webhook URL for push notifications 8. LLM config – API base URL, model name, API key (default: same as PaperClaw models.json) 9. Schedule timezone – Default is Asia/Singapore

Infer reasonable defaults for anything not provided and confirm before proceeding.

Step 2: Build Keyword Library

Construct a structured keyword library from the domain interview. Aim for:

• Core queries (3–5): Direct topic+method combinations for arXiv ti: searches
• Method queries (3–5): Method+application combinations
• Application queries (2–3): Use-case-specific terms
• Exclusion keywords (3–6): Out-of-scope terms to filter

See references/domain-adaptation-guide.md Section 1 for keyword examples across 8 common domains.

Step 3: Design Evaluation Rubric

Design 4 domain-specific scoring dimensions (each scored 1–10) that replace PaperClaw's SciML dimensions (engineering_value, architecture_innovation, theoretical_contribution, result_reliability).

The scoring formula is unchanged:

final_score = base_score × 0.9 + impact_score × 0.1
base_score = (dim1 + dim2 + dim3 + dim4) / 4
impact_score = date_citation_adjustment(citations, age_months)

See references/domain-adaptation-guide.md Section 2 for rubric examples by domain.

Step 4: Generate Agent Files

Run the scaffolding script to create the directory structure:

python ~/.comate/skills/paper-expert-generator/scripts/init_domain_agent.py \
  --domain "<domain_slug>" \
  --output "<output_dir>" \
  --paperclaw-skills "<paperclaw_skills_path>"

Example:

python ~/.comate/skills/paper-expert-generator/scripts/init_domain_agent.py \
  --domain "bioinfo-ml" \
  --output ~/agents/bioinfo-ml \
  --paperclaw-skills /work/work/PaperClaw/skills

Generated structure:

<output_dir>/
├── agent/
│   ├── AGENT.md          ← write domain content here
│   ├── models.json       ← pre-filled from template
│   └── schedules.json    ← pre-filled from template
├── skills/
│   ├── arxiv-search/     ← copy from PaperClaw (needs keyword update)
│   ├── semantic-scholar/ ← copy from PaperClaw (no changes needed)
│   ├── paper-review/     ← copy from PaperClaw (needs rubric update)
│   ├── daily-search/     ← copy from PaperClaw (minor text update)
│   └── weekly-report/    ← copy from PaperClaw (minor text update)
└── workspace/
    └── evaluated_papers.json  ← initialized empty

Step 5: Write AGENT.md

Use assets/templates/AGENT.md.template as the base. The AGENT.md must include:

1. Role Definition – Domain expert persona with specific depth. Replace SciML expertise with domain-specific expertise (key algorithms, theoretical foundations, benchmark datasets, top venues/conferences).

2. Keyword Library – Paste structured keywords from Step 2.

3. Four Core Tasks (preserve exact structure from PaperClaw):

   • Task 1 (Paper Research): Download PDF → write summary.md answering 10 domain-adapted questions
   • Task 2 (Paper Evaluation): 4-dimension scoring → write scores.md → update metadata.json → update registry
   • Task 3 (Daily Search): Cron trigger → daily_paper_search.py --top 3 → dedup → trigger Task 1+2
   • Task 4 (Weekly Report): Cron trigger → generate_weekly_report_v2.py → push notification

4. Mandatory <think> Reasoning – Required in Task 2 evaluation.

5. Dedup Gate – Always check evaluated_papers.json before starting paper review.

See references/agent-template-guide.md for the full AGENT.md authoring guide.

Step 6: Adapt Skill SKILL.md Files

Minimal adaptation needed – Python scripts are domain-agnostic:

Step 7: Configure models.json and schedules.json

models.json: Edit agent/models.json, fill in:

• baseUrl: LLM API endpoint
• apiKey: API key placeholder
• id and name: Model identifier

schedules.json: Default schedule is pre-filled. Adjust tz field if not in Singapore timezone.

Step 8: Validate and Deliver

Checklist before presenting results:

• AGENT.md has domain role, keywords, 4 tasks, rubric
• paper-review/SKILL.md has domain scoring dimensions + 10 adapted questions
• arxiv-search/SKILL.md has domain keyword list
• models.json has correct structure (API key placeholder)
• workspace/evaluated_papers.json initialized as []
• All 5 skill directories exist

Then present the output summary (see next section).

Output Summary Format

Always deliver this summary after generation:

## Generated Agent: <Domain Name> Paper Expert

**Domain**: <domain>
**Location**: `<output_dir>`
**Model**: <model_name>

### Keyword Library (<N> total queries)
**Core**: <query1>, <query2>, <query3>
**Methods**: <query1>, <query2>
**Exclusions**: <term1>, <term2>, ...

### Evaluation Rubric
| Dimension | Score Weight | Measures |
|-----------|-------------|---------|
| <dim1>    | 25%         | ...     |
| <dim2>    | 25%         | ...     |
| <dim3>    | 25%         | ...     |
| <dim4>    | 25%         | ...     |

### Schedule
- Daily search: `0 20 * * *` (<timezone>)
- Weekly report: `0 10 * * 0` (<timezone>)

### Quick Start
1. Open OpenClaw → select agent from `<output_dir>/agent/`
2. Set API key in `agent/models.json`
3. Test: "Search for recent papers on <core_topic>"
4. Or wait for first daily trigger at 20:00

References

• references/domain-adaptation-guide.md – Keyword and rubric examples for 8 common domains
• references/agent-template-guide.md – Full AGENT.md authoring guide with annotated sections
• assets/templates/AGENT.md.template – Base template for the generated AGENT.md
• assets/templates/models.json – Base models config template
• assets/templates/schedules.json – Base schedules config template