CommunityResearch & Data Analysisgithub.com

good-research-ideas

Use when the user asks for literature-backed research topics, thesis ideas, open problems, or near-term directions in a field or subfield, such as robotics, robot vision-language-action (VLA) models, or robotic action policies. Not for a standalone paper summary.

What is good-research-ideas?

good-research-ideas is a Claude Code agent skill that use when the user asks for literature-backed research topics, thesis ideas, open problems, or near-term directions in a field or subfield, such as robotics, robot vision-language-action (VLA) models, or robotic action policies. Not for a standalone paper summary.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/RedZadios/good-research-ideas/tree/HEAD/skills/good-research-ideas

Installed? Explore more Research & Data Analysis skills: affaan-m/uspto-database, affaan-m/scholar-evaluation, affaan-m/literature-review · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Good Research Ideas

Find recurring, unresolved gaps in recent top-tier research and turn them into testable ideas. Lead with the ideas in a first answer of 200 words or fewer. Keep the evidence traceable separately.

Search and select up to 20 papers

  • Turn the user's scope into a research question and identify what must match: the problem, task, system, inputs, outputs, or evaluation setting. Explore several relevant subareas for a broad topic. For a narrow request, keep its defining constraints rather than broadening it to fill the sample. Use the user's stated resource constraints without assuming access to hardware or private data.
  • Use Consensus and SciSpace for complementary discovery when both are available. Search with natural-language research questions, use official proceedings and journals to cover venues and awards, and expand using author terminology and citation links from strong matches. See Search and evidence methods for the search passes, relevance checks, and fallbacks.
  • Read the full abstract and check the paper's author-supplied keywords where available. Judge whether the core contribution and task or evaluation address the requested scope. Look past exact wording: a paper may qualify without the query's acronym, and a title, keyword, or award cannot rescue an off-topic paper. If no author keywords are published, record that and use terms expressly stated in the abstract.
  • Use a rolling three-calendar-year window ending on the search date. Verify peer-reviewed publication and the earliest known public date of the same study before reading full text. Deduplicate preprints, proceedings, and journal versions of the same work. Exclude papers published only as preprints.
  • Prefer the closest direct matches and use verified quality signals, such as an official Best Paper award or substantial citation growth within the last two years, to rank papers with comparable relevance. Set the selection criteria and reserve order before reading limitations. If needed, label directly relevant papers with verified secondary signals, such as an official finalist designation or independent external validation. Recency and venue prestige alone are insufficient. Replace papers without readable full text using eligible reserves. Aim for 20; report the actual number rather than padding.

Read, cross-check, and rank

  • Read the original Limitations and Future Work sections where present. Also inspect relevant discussion, results, conclusion, and supplementary material. Log each paper's author-stated limitations, author-proposed future work, and gaps inferred from results separately, with a version link and section/page/figure locator. Flag absent or unverified statements; abstracts and automated extractions do not count as full-text review.
  • Group claims only when they concern the same concrete bottleneck or proposed next step. Count author-stated limitations and future-work proposals separately as distinct papers out of the number reviewed. Record the supporting paper IDs and independent teams. Note shared authorship, data, or methods; keep inferences, partial fixes, and counterevidence separate.
  • Check newer studies within the same window for partial fixes or counterevidence. Outside-sample papers can change the conclusion without changing the sample counts. Rank testable directions by frequency, independence, remaining uncertainty, and feasibility. Explain why a direction looks promising over the next 1-2 years without equating repetition with a forecast or claiming novelty is proven.

Deliver the concise answer first

  • Keep the entire first chat answer to 200 words or fewer, including headings and notes. Match the user's language and use equivalent brevity in languages without space-delimited words. Lead with the strongest directions. For each, give a concrete question, the key evidence or support count, and a minimal experiment or metric. State the actual number reviewed and any limitation that changes the ranking.
  • Keep the full paper ledger, both support rankings, and detailed experimental designs out of the first chat answer. Put them in a linked evidence file if file output is available; otherwise retain the evidence for a follow-up. Keep key source citations in the short answer itself.

Related Skills