Explore literature-grounded research innovation ideas and paper framing in a host-neutral way. Use when an AI agent needs to collect a recent paper pool, run broad and deep literature search, decompose methods into reusable capabilities, generate A+B or module-combination candidates, shortlist feasible ideas, design a defensible unifying framework, and produce an elegant Markdown report with citations, analysis rationale, and visual summaries.
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.
A direct command skips the review prompt. Inspect the source before running it.
The command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
File Explorer
29 files
Showing SKILL.md
SKILL.md
Source instructions · Read-only preview
More from this repository
name
research-innvtn-explorer
description
Explore literature-grounded research innovation ideas and paper framing in a host-neutral way. Use when an AI agent needs to collect a recent paper pool, run broad and deep literature search, decompose methods into reusable capabilities, generate A+B or module-combination candidates, shortlist feasible ideas, design a defensible unifying framework, and produce an elegant Markdown report with citations, analysis rationale, and visual summaries.
Research Innovation Explorer
Overview
Use this skill to turn a vague "find a publishable idea" request into a disciplined workflow: search broadly, collect strong recent papers, decompose them into reusable capabilities, generate structured candidate combinations, and draft an honest framing plus experiment plan.
Treat this as efficient incremental research planning. Stand on strong prior work, keep claims proportional to evidence, and prefer explicit assumptions over inflated novelty language.
Use this skill as a host-neutral contract. If the current environment supports native Skills, load it with the host's own mechanism. If the environment does not support Skills, follow SKILL.md, references/, and scripts/ directly.
Quick Start
Clarify the domain, target venue, resource limits, and whether the user wants concept-only output or a code-ready shortlist.
Read references/host-neutral-usage.md and adapt invocation to the current host.
Read references/search-playbook.md, then copy assets/templates/search-log.csv to a working file.
Run python scripts/build_search_queries.py --topic "<topic>" --keywords "<comma-separated keywords>" to generate a search pack.
Search across multiple sources, log results, and fill assets/templates/paper-pool.csv with 20-50 strong papers.
Run python scripts/build_idea_matrix.py <paper_pool.csv> --output <idea_matrix.csv> to generate pairwise candidates.
Read references/scoring-rubric.md to shortlist roughly 10-20 promising combinations.
For finalists, read references/framing-and-theory.md and write idea briefs with assets/templates/idea-brief.md.
Read references/experiment-plan.md and draft the validation plan with assets/templates/experiment-plan.md.
Read references/reporting-and-visualization.md and run python scripts/build_markdown_report.py ... to produce the final Markdown report.
Workflow
1. Use Search Aggressively and Systematically
Treat search as mandatory during both collection and analysis, not as an optional helper.
Use the best search surfaces available in the current environment: web search, browse tools, academic search APIs, paper databases, official docs, code search, and repository search.
Search at multiple depths:
broad topic scan
targeted method scan
benchmark and dataset scan
citation chaining
negative evidence and failure-case scan
code and implementation scan
Read references/search-playbook.md before collecting papers or writing any analysis.
2. Build the Paper Pool
Prefer 20-50 recent, strong, code-accessible papers with real downstream impact.
Record task, modules, strengths, weaknesses, benchmarks, and open_source for each paper.
Mix academic papers with strong open-source or industry systems only when they change the practical frontier.
Read references/workflow.md for sourcing rules and intake guidance.
3. Decompose Papers into Capabilities
Rewrite each paper as reusable components, not paper titles.
Separate task from mechanism: objective, backbone, routing, memory, loss, training recipe, inference trick, evaluator, or data recipe.
Normalize fields so pairwise comparisons are meaningful.
4. Generate Candidate Combinations
Use scripts/build_idea_matrix.py for a first-pass matrix.
Keep combinations where task overlap is real, mechanisms are complementary, and implementation remains tractable.
Kill combinations that are purely cosmetic, redundant, or impossible to evaluate with available code and data.
Read references/scoring-rubric.md when the shortlist is noisy or overfull.
5. Write the Innovation Hypothesis
State the claim as a falsifiable hypothesis, not a slogan.
Prefer forms like:
"Method A supplies X that method B lacks under condition C."
"A unified controller over A-style and B-style mechanisms should dominate each extreme on regime R."
"A shared objective reveals A and B as limiting cases of a broader family."
Write why the combination should help, what assumption it needs, and where it can fail.
6. Frame the Theory Honestly
Use references/framing-and-theory.md before writing any framework section.
Build a higher-level abstraction only if you can define the latent variable, control knob, objective decomposition, or limiting cases clearly.
Claim "A and B are special cases" only when the algebra or algorithm really supports it.
If the work is primarily a strong engineering combination, say so and focus on mechanism plus evidence rather than pretending to have a theorem.
