research-internet

name	research-internet
description	Conduct structured internet research and write `research/research_internet.md`.
model	claude-sonnet-4-6

Research Internet

Version: 4

Goal

Conduct comprehensive internet research to extract all available human knowledge relevant to the current task, discover new papers, add them to the project, and produce a structured research_internet.md.

Inputs

• $TASK_ID — the task folder name (e.g., t0003_baseline_wsd_with_bert)

Context

Read before starting:

• project/description.md — project goals and scope (defines search boundaries)

• tasks/$TASK_ID/task.json — understand the task objective

• tasks/$TASK_ID/research/research_papers.md — identify gaps and context

• arf/specifications/research_internet_specification.md — authoritative format specification for research_internet.md

• arf/styleguide/markdown_styleguide.md — formatting rules (100-char lines, * bullets, heading hierarchy)

Steps

Phase 1: Define Research Scope

1. Read tasks/$TASK_ID/task.json and extract the task objective. If it contains long_description_file, also read the referenced markdown file.

2. Read tasks/$TASK_ID/research/research_papers.md. Extract:

   • Gaps from the Gaps and Limitations section
   • Key topics, methods, and open questions from Key Findings
   • Terminology and author names for targeted searches

3. Formulate two lists:

   • Gap-filling questions — one per gap from research_papers.md
   • Broadening questions — what else exists about this topic that the paper corpus does not cover? Think: recent advances, alternative approaches, practical implementations, community best practices, available tools/libraries, datasets, benchmarks, blog posts with reproducibility insights, and any non-academic knowledge

Phase 2: Systematic Multi-Source Search

Execute a structured, multi-pass search strategy.

1. Define search parameters before searching:

   • Target sources: Google Scholar, Semantic Scholar, arXiv, ACL Anthology, GitHub, Hugging Face, Papers With Code, relevant blogs, documentation sites, Stack Overflow, Reddit, Twitter/X threads, conference workshops

   • Keywords with synonyms and Boolean operators

   • Date range and inclusion/exclusion criteria

   • Minimum 8 distinct search queries

2. Pass 1 — Gap-targeted queries using WebSearch:

   • One query per gap from Phase 1
   • Use specific terminology from the papers already reviewed

3. Pass 2 — Broadening queries using WebSearch:

   • Survey queries for the general topic (recent advances, state of the art)
   • Code/implementation queries (GitHub, model hubs, pip packages)
   • Practical experience queries (blog posts, tutorials, benchmarks)
   • Community discussion queries (forums, issue trackers)

4. Pass 3 — Snowball queries based on findings from passes 1-2:

   • Follow references mentioned in discovered sources
   • Search for specific authors, tools, or datasets uncovered
   • Check structured sources: Papers With Code leaderboards, Hugging Face model/dataset hubs, conference proceedings pages

5. Deep-read promising sources using WebFetch:

   • Read full content of the most relevant blog posts, documentation pages, and repository READMEs
   • Extract specific numbers, benchmarks, and implementation details
   • Note reliability of each source (peer-reviewed vs. not)

6. Record every query — exact text and source. Do not paraphrase.

Phase 3: Identify and Add New Papers

Deliverable-first rule: Write research/research_internet.md including the ## Discovered Papers section. If context is running low, prioritize completing the research document — the Discovered Papers section is enough for the orchestrator to add papers later.

List ALL relevant papers in the ## Discovered Papers section — not just papers about the narrow task topic, but any paper relevant to the broader project. A paper referenced in search results, cited by other papers, or appearing in related work sections counts. If research_papers.md cites a paper that is not yet in the corpus, list it. The goal is to grow the project's paper corpus comprehensively.

1. Compile all papers encountered during search into the ## Discovered Papers section of research_internet.md using the format from arf/specifications/research_internet_specification.md. Include:

   • Papers directly about the task topic
   • Foundational/seminal papers referenced by multiple sources
   • Papers cited in research_papers.md that are not yet in the corpus
   • Papers defining datasets, benchmarks, or evaluation frameworks used by the project
   • Recent papers that advance the state of the art in related areas

2. Run the paper aggregator to get existing papers with titles and DOIs:

   uv run python -u -m arf.scripts.aggregators.aggregate_papers \
     --format json --detail short
   

3. Cross-reference discovered papers against the existing corpus by comparing DOIs, normalized titles (lowercase, stripped punctuation), and author+year combinations. Remove duplicates.

4. Write all new papers into the ## Discovered Papers section of research_internet.md. Do NOT invoke /add-paper or /download_paper. The orchestrator handles paper addition after this skill completes.

Phase 4: Synthesize and Write research_internet.md

1. Organize all findings by topic, not by source. Each topic becomes a ### subsection under ## Key Findings.

2. For every gap from research_papers.md, determine resolution status: Resolved, Partially resolved, or Unresolved.

3. Write the full research_internet.md following the format in arf/specifications/research_internet_specification.md:

   • YAML frontmatter with all required fields

   • All 8 mandatory sections: Task Objective, Gaps Addressed, Search Strategy, Key Findings, Methodology Insights, Discovered Papers, Recommendations for This Task, Source Index

   • Additional sections wherever the task demands deeper treatment (e.g., Tool and Library Landscape, Benchmark Comparison, Implementation Patterns, Community Discussions, Dataset Availability)

4. Ensure every factual claim has an inline [SourceKey] citation.

5. Ensure every Source Index entry is cited at least once, and vice versa.

6. Distinguish peer-reviewed from non-peer-reviewed sources in the text.

7. Include specific numbers everywhere. Reject vague claims like "performs well" — replace with actual metrics.

8. Extract and explicitly state:

   • Hypotheses — testable claims emerging from the research
   • Best practices — community-converged approaches and pitfalls
   • Contradictions — where new findings disagree with existing papers

Phase 5: Verify

1. Run the verificator:

   uv run python -u -m arf.scripts.verificators.verify_research_internet $TASK_ID
   

2. Fix all errors. Address warnings unless there is a documented reason to skip them.

3. Re-run until zero errors.

Done When

• All discovered papers are listed in the ## Discovered Papers section of research_internet.md
• tasks/$TASK_ID/research/research_internet.md exists and follows the specification
• Verificator passes with zero errors
• Every gap from research_papers.md has a resolution status
• At least 8 distinct search queries are documented in Search Strategy
• Research covers both gap-filling AND broad topic exploration
• Every inline [SourceKey] has a Source Index entry and vice versa

Forbidden

• NEVER run prestep or poststep — the orchestrator handles the step lifecycle

• NEVER commit — the orchestrator handles all commits

• NEVER modify step_tracker.json — the orchestrator manages step state

• NEVER write step_log.md — the orchestrator writes it after this skill completes

• NEVER fabricate sources, URLs, or metrics. If a search yields nothing, say so.

• NEVER paraphrase search queries — record the exact text used.

• NEVER skip the gap analysis from research_papers.md.

• NEVER cite non-peer-reviewed sources without noting their reliability.

• NEVER add papers without checking the aggregator for duplicates first.

• NEVER invoke /add-paper or /download_paper — paper addition is the orchestrator's responsibility

• NEVER skip listing a paper just because it is not "about" the narrow task topic. If a paper is relevant to the project's research area, list it in Discovered Papers.

• NEVER claim "no new papers to list" without verifying that every paper cited in research_papers.md and every paper encountered in search results already exists in the corpus.