Menรผ
Collects research sources with quality evaluation. Use when gathering or finding sources, building a source library, searching for academic papers, or performing RADAR assessment.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fรผgen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prรผfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Delegates a task to OpenAI Codex running as an interactive session in a herdr pane - uses the user's ChatGPT subscription, visible in herdr, steerable mid-session. Use when delegating work to codex, offloading a task to a GPT model (sol, terra, luna), or running a second opinion from a non-Claude frontier model. Not for Claude Code's own Agent tool or local-LLM delegation (delegating-to-local-llm).
Delegates a task to a local LLM running as a Pi coding-agent session in a herdr pane - the subagent is visible in herdr, can be steered mid-session, and costs no cloud tokens. Use when delegating work to a local model, spawning a Pi subagent, offloading a task to omlx (qwen, gemma), or running agent work offline. Not for Claude Code's own Agent tool or subagents inside Pi.
Loads the rageatc-code-oss operating rules directly from the plugin when the SessionStart hook has not delivered them. Use when the rageatc-code-oss failsafe line appears in session-start hook output, when the plugin's operating rules are absent from context, or when the injected rules look truncated.
Loads the rageatc-core-oss operating rules directly from the plugin when the SessionStart hook has not delivered them. Use when the rageatc-core-oss failsafe line appears in session-start hook output, when the plugin's operating rules are absent from context, or when the injected rules look truncated.
Raises an artefact's signal-to-noise ratio by removing text that earns nothing for its reader โ filler, repetition, over-explanation, ornamental phrasing โ and relocating conditional detail to references. Use to slim a skill, agent, or document, cut bloat, tighten prose, or shrink a file that has grown too long. Tuned for SKILL.md files.
Creates a design system for software with a UI. Use when a project has a user interface and architecture is confirmed โ whether creating from scratch or extracting patterns from existing code.
| name | collecting-sources |
| description | Collects research sources with quality evaluation. Use when gathering or finding sources, building a source library, searching for academic papers, or performing RADAR assessment. |
Phase 1 of two-phase research: discover, evaluate, and store sources with metadata. Synthesis (Phase 2) and fact-checking (Phase 3) are separate phases โ this skill ends at handoff.
Everything lands in the task workspace:
work/<task-id>/
โโโ source_index.md # human-readable catalogue
โโโ sources/
โโโ papers/ # academic papers (.txt from CORE/Unpaywall, .pdf from arXiv)
โโโ web/ # web pages (Markdown)
โโโ blogs/ # blog posts (Markdown)
โโโ docs/ # documentation (Markdown/HTML)
Every source file has a .meta.yaml companion conforming to schema v1.0 โ fields, missing-data conventions, and a complete example in references/source-metadata-schema.md.
mkdir -p work/<task-id>/sources/{papers,web,blogs,docs}src_001, src_002, โฆsource_index.md exists from a previous collection, load it for duplicate detection.Detect the domain from the research question:
Academic research โ discovery finds papers and extracts identifiers (DOI, arXiv ID, PubMed ID, exact title) for the retrieval chain; do not fetch full text during discovery. Run 3โ5 targeted WebSearch queries mixing plain and site-specific forms:
site:pubmed.ncbi.nlm.nih.gov [topic]site:arxiv.org [topic]site:scholar.google.com [topic], [topic] research paper peer-reviewedCollect candidate metadata: title, authors, publication date, venue, abstract excerpt, identifiers.
General web research โ run 3โ5 targeted WebSearch queries with alternative phrasings; target authoritative sources (official docs, expert blogs, reputable sites); use site-specific searches when the site is known (site:docs.python.org async). Collect URLs, titles, authors, dates.
Target: 15โ25 candidates from diverse source types, to be filtered to 8โ15 on quality.
Before fetching each candidate, check for duplicates: URL match against existing sources, DOI match against doi fields in existing .meta.yaml files. If already collected, skip and log.
Try in order; record which step succeeded as retrieval_method:
1. CORE API โ always first; 46M full texts as plain text, all domains:
https://api.core.ac.uk/v3/search/works?q=doi:[DOI] โ or q=title:"[exact title]" if no DOIfullText field is non-empty, save it as .txt โ done. Otherwise continue.2. Unpaywall MCP โ open-access PDF discovery and extraction:
unpaywall_get_fulltext_links(doi=...) โ best_oa_location.url_for_pdf; without a DOI, find one via unpaywall_search_titles(query="exact title")unpaywall_fetch_pdf_text(doi=..., truncate_chars=50000) โ save as .txt (typical papers run 20,000โ30,000 chars; surveys 90,000โ100,000)3. Domain repository:
https://export.arxiv.org/pdf/[arxiv_id].pdf โ this stores the binary PDF only (WebFetch cannot extract PDF text). Max ~4 requests/second.https://pmc.ncbi.nlm.nih.gov/articles/PMC[PMCID]/ โ full-article HTML, store as Markdown or HTML.4. WebFetch the landing page โ occasionally yields full text as HTML; at minimum, abstract and bibliographic metadata.
5. Metadata-only (paywalled, last resort):
site:arxiv.org [title], site:biorxiv.org [title], [author] [title] pdf.file_path: "unavailable" and content_hash: "unavailable"; list the attempted strategies in provenance.notes; flag "Metadata only (paywalled)" in the index.Social sciences caveat: OA coverage is ~33% vs ~66% for STEM, and SSRN (1.74M preprints) has no API. Expect 50โ60% metadata-only rates and say so in the collection summary.
WebFetch (converts HTML to Markdown). Preserve HTML only when formatting is critical.
Classify source_type (academic_paper, web_page, blog, documentation) and store:
| Source | Path | Format |
|---|---|---|
| Paper via CORE/Unpaywall | sources/papers/src_NNN.txt | plaintext |
| Paper via arXiv direct | sources/papers/src_NNN.pdf | |
| Paper, metadata-only | no file โ file_path: "unavailable" | โ |
| Web page / blog / documentation | sources/{web,blogs,docs}/src_NNN.md | markdown (or html) |
Generate the content hash โ shasum -a 256 <file> โ content_hash: "sha256:HEX" โ and compare against existing hashes to catch duplicate content fetched from different URLs.
Write src_NNN.meta.yaml beside each source file, conforming to references/source-metadata-schema.md: core fields, source-type-specific fields, provenance (collected_by, collection_method, collected_at, task_id, notes), and retrieval_method for academic papers.
Assess every source on five RADAR dimensions, scored 1โ5 with brief notes:
reliability_score = (Relevance + Authority + Date + Accuracy + Rationale) / 25.0
Full scoring rubric: skills/verifying-claims/references/source-evaluation-radar.md. Paywalled sources are assessed on abstract, venue, and authors.
Minimum 8 high-quality sources per research question. If short: broaden queries, try alternative tools, lower the floor to 0.5 if necessary, and document the limitation. Retain everything above threshold โ more coverage helps synthesis; prioritise the highest scores there.
Write work/<task-id>/source_index.md:
# Source Index
**Task**: [task-id] ยท **Research question(s)**: [โฆ] ยท **Collection date**: [YYYY-MM-DD]
## Summary Statistics
- Total sources: [N] โ [breakdown by type]; papers: [X] full-text / [Y] metadata-only
- Retrieval methods (papers): CORE [n] ยท Unpaywall [n] ยท arXiv [n] ยท PMC [n] ยท WebFetch [n] ยท metadata-only [n]
- Average reliability: [score]; distribution: excellent โฅ0.8 [n] ยท good 0.7โ0.79 [n] ยท acceptable 0.6โ0.69 [n]
## Sources by Topic
### [Topic/theme]
**src_NNN** โ [Title]
- **URL**: [โฆ] ยท **Type**: [โฆ] ยท **Reliability**: [โฆ] ยท **Retrieval**: [โฆ]
- **Summary**: [one or two lines]
- **Metadata**: `sources/<subdir>/src_NNN.meta.yaml`
## Collection Notes
- Tools used; quality threshold applied
- Limitations: [paywalled count, coverage gaps, domain caveats]
Group sources by themes extracted from abstracts/descriptions and the research question; note sources relevant to multiple topics.
Report to the orchestrator: total sources (full-text vs metadata-only), breakdown by type and retrieval method, average reliability, paths to source_index.md and the sources directory, exclusions and limitations (including domain caveats), and whether the collection is ready for synthesis or needs additional searches.