Menu
Research with rforge provenance. Use for literature search, OSS study, systematic review, evidence extraction, meta-analysis, or review packages.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Resolve short approval against the latest checkpoint. Use for "lgtm", "looks good", "approved", or "go ahead"; not risky or review feedback.
Re-check a prior assistant claim. Use when user says "nack", "not convinced", or "check again"; not for review feedback.
UI/UX routing front door. Use as the front door for broad or ambiguous UI/UX work that needs the right build, redesign, audit, design-system, accessibility, or browser specialist. Do not use when the user already names a narrower UI skill.
Proposes UI Vault resource upgrades for a selected webpage. Use when the user asks for libraries, assets, effects, or inspiration for an existing page. Do not use for implementation or greenfield design.
Upload, list, or delete files with s3upload and return expiring Azure links. Use for configured temporary Azure storage; not for S3 or cloud setup.
Ship local Git changes by inspecting, polishing, validating, split-committing, and pushing. Use for commit, push, ship, delivery audit, or delivery blockers; not deploys or releases.
| name | research-forge |
| description | Research with rforge provenance. Use for literature search, OSS study, systematic review, evidence extraction, meta-analysis, or review packages. |
Use rforge for retrieval-first, provenance-first research. It is a standalone CLI, not a project dependency.
Follow the shared skill contract for repo study, dirty-worktree hygiene, verification evidence, safe handoffs, and safety defaults.
command -v rforge && rforge version || true
rforge automation policy 2>/dev/null || true
queries.txt for multi-query work.rforge search batch ... --stats; prefer it over shell loops.rforge search stats --dir <topic-dir> and save the output if not already written.report.md and provenance.json before finishing.If rforge is missing, do not silently install for a plan-only request. For an execution request, ask before network install unless the user already authorized setup. Install options:
curl -fsSL https://raw.githubusercontent.com/TrebuchetDynamics/research-forge/main/install.sh | bash
# or, with Go available:
go install github.com/TrebuchetDynamics/research-forge/cmd/rforge@latest
Inspect before acting:
artifacts/ or research/ folder;rforge version and, for full reviews, rforge doctor when available;rforge automation policy for agent-allowed versus human-required actions;provenance.json, manifest.json, results.jsonl, reports, screening queues, or review package files in the target folder;references/deep-search.md for Comprehensive sweeps.| Context | Save to |
|---|---|
Project repo with artifacts/ | artifacts/research/<topic-slug>/ |
Project repo without artifacts/ | research/<topic-slug>/ |
| Known external project | that project's research/ or artifacts/research/ |
| No clear project home | ~/research/<topic-slug>/ |
Topic slug: lowercase, hyphen-separated, 3–6 words. Confirm the path when ownership is unclear.
| Request phrasing | Depth |
|---|---|
| "quick look", "any papers on", "what's out there" | Quick only if explicitly requested: 3 query variants, fast sources |
| "standard", "small pass", "fast pass" | Standard only if explicitly requested: 5–8 query variants, scholarly-fast or relevant preset, citation expansion |
| Any research request by default; "full", "systematic", "comprehensive", "mega", "thorough" | Comprehensive: 20–30 variants, broad/all relevant sources, citation expansion, evidence grid, gaps |
Default to Comprehensive.
Academic search:
printf '%s\n' \
"query variant one" \
"query variant two" > queries.txt
rforge search batch --out <topic-dir> --queries queries.txt \
--sources all --limit 20 --continue-on-error --stats
rforge search stats --dir <topic-dir> | tee <topic-dir>/coverage-stats.log
Useful presets: openalex,arxiv, scholarly-fast, openalex,arxiv,semantic-scholar, biomedical, preprints, open, all.
If batch is unavailable, save one file per source and still run stats:
rforge search --source openalex --query "QUERY" --limit 20 > <topic-dir>/search-openalex.txt 2>&1 || true
rforge search stats --dir <topic-dir> | tee <topic-dir>/coverage-stats.log
OSS study starts with a plan, not cloning or installing:
rforge oss search-plan --query "<project/functionality>" --ecosystem all > <topic-dir>/oss-search-plan.txt
Assess license, maintenance, security posture, release cadence, package metadata, Software Heritage/archive status, and domain fit. Stars/downloads are only weak signals.
Pick 3–5 seed papers from retrieved metadata: recent surveys, high-impact venues, high-citation works, method-defining preprints, or bridge papers.
rforge citations expand --source semantic-scholar --paper <id-or-doi> \
--direction both --depth 1 --out <topic-dir>/citation-graph.json
rforge citations report --graph <topic-dir>/citation-graph.json --out <topic-dir>/citation-report.md
Screen before downloading PDFs:
rforge screen queue --dir <topic-dir> --out <topic-dir>/queue.csv
# human fills decision and reason columns
rforge screen import --dir <topic-dir> --csv <topic-dir>/queue.csv --reviewer <name>
rforge screen progress --dir <topic-dir>
Only after inclusion screening and any required acquisition approval:
rforge oa fetch --dir <topic-dir>
For Comprehensive reviews:
rforge evidence grid --out <topic-dir>/evidence-grid.json
rforge evidence gaps --out <topic-dir>/evidence-gaps.json
For systematic/meta-analysis work, prefer a project and guided state machine:
rforge project create <project-dir> --title "<title>"
rforge forge init --project <project-dir> --question "<research question>"
rforge forge status --project <project-dir>
rforge forge next --project <project-dir>
Prepare but do not self-approve analysis or final claims:
rforge analysis prepare
rforge report trace --claims <claims.json> --analysis <run.json> --out <trace.json>
rforge report claim-panel --trace <trace.json> --out <claim-panel.json>
report.md sections:
Never assert a performance or scientific result without naming the paper/repo, venue/source, year, and conditions visible in retrieved evidence.
Always write valid provenance.json:
{
"question": "<exact research question>",
"rforge_version": "<version or not available>",
"timestamp": "<ISO 8601>",
"depth": "quick|standard|comprehensive",
"queries": ["<query 1>"],
"sources": ["openalex", "arxiv"],
"search_stats": {"openalex": 0, "arxiv": 0, "total_unique_dois": 0},
"citation_expand_attempted": ["<paper id or DOI>"],
"citation_expand_succeeded": ["<paper id or DOI>"],
"outputs": ["report.md", "provenance.json"],
"errors": ["<rate limit, API failure, empty output, or missing source notes>"]
}
For review packages, require human package/export approval and then verify:
rforge package create --out <dir> --created-by <name> --question "<text>"
rforge package audit <dir>
rforge package replay <dir>
Before live API calls or file writes, confirm save path, source preset, depth, privacy/licensing posture, and whether the request is academic literature, OSS study, or reproducible review package.
Before done:
rforge version printed or provenance records unavailable.report.md or requested artifact cites exact papers/repos/artifacts.provenance.json exists, is valid JSON, and names all outputs.rforge package audit and rforge package replay when package export was approved.Do not self-approve screening decisions, full-text acquisition, privacy/licensing review, extraction acceptance, analysis method selection, final scientific claims, LLM suggestion review commands, package export, copyrighted full text, or live scholarly APIs in tests/CI.
Final reply names the save path, depth, sources queried, output files, validation receipts, unresolved gates, and weakest evidence gap.
User: research recent OSS tools for citation graph exploration
Agent: choose research/citation-graph-oss-tools/, run rforge oss search-plan, save plan/report/provenance, and stop before cloning or recommending integration without license/security review.