Define a research problem. Find a way to solve it.
12 composable agent skills for collecting papers, understanding a field, developing research ideas, checking novelty, planning informative experiments, and explaining concepts in textbook-style prose.
- Open Plugins in the Codex desktop app.
- Choose Add marketplace and enter
HytroZero/k-research. - Confirm to add the marketplace.
Download the full repository and keep its directory structure intact. Skills use shared files in docs/, templates/, and schemas/.
Ask your agent to read and follow skills/<skill-name>/SKILL.md, using this repository as the root for shared file paths. No separate installation is needed for this workflow.
Replace <repo> with the local repository path:
Read <repo>/skills/ask-k/SKILL.md and help me choose where to start
with research on efficient long-context inference.
Read <repo>/skills/discover-research-direction/SKILL.md and follow it
to explore research directions in efficient long-context inference.
You can also use any individual skill directly. Direction discovery centers on the problem, proposed method, and difference from prior work. Detailed experiment design follows when experiments or a full research plan are requested, or a developed method is ready for validation. Outputs include paper notes, literature maps, research proposals, and separate experiment plans.
Methods selected for recommendation receive parallel red-team and novelty-audit subagent reviews, followed by revision by the main agent. This applies to standalone method development and direction discovery. Review normally runs once, with at most one targeted follow-up after a substantial change. Users can request a skip; hosts without subagents perform sequential reviews and disclose that fallback.
Requests to explain or understand a theory, hypothesis, method, concept, or principle automatically trigger textbook-style-explanation, for example: “Explain why the chain rule works, assuming basic knowledge of derivatives.” No explicit writing-style request or research project is required. It provides a textbook-style explanation in the conversation by default and saves a file only when requested.
Research outputs use one standard layout: README.md is the human entry point, research/ holds the overview and direction proposals, and papers/ holds the complete index, bibliography, notes, and PDFs. Agent-only state, cache, and temporary work stay in .research/.
Skills update current documents rather than saving a report per run. Files are created only when needed; experiment plans remain a follow-on deliverable. Existing K Research outputs are consolidated into this layout. Meaningful decision changes stay in research/decisions.md; no file snapshots or backup history are generated, including without Git.
| Skill | Description |
|---|---|
| ask-k | Choose the right skill or short workflow for your current research stage, with a ready-to-use invocation. |
| textbook-style-explanation | Automatically respond to requests to explain a theory, hypothesis, method, concept, or principle with approachable textbook-style prose; no research project required. |
| discover-research-direction | Develop worthwhile problems and concrete methods, verify their differences from closest work, and recommend a research direction. |
| collect-paper | Build a screened paper collection for a topic, verify sources, download key PDFs, and save an index and BibTeX bibliography. |
| paper-analysis | Read an individual paper to explain its question, method, claims, evidence, and limitations, with source locators. |
| field-synthesis | Connect a body of papers into a field map covering research lines, method evolution, agreements, disagreements, and open questions. |
| opportunity-discovery | Find worthwhile research problems and possible approaches from practical bottlenecks, desired capabilities, observations, and unresolved questions. |
| hypothesis-development | Develop a concrete method with a problem-grounded rationale, worked example, closest-work comparison, and expected contribution. |
| novelty-audit | Search and inspect the closest prior work to assess overlap, identify substantive differences, and support a novelty judgment. |
| decisive-test-design | Plan the smallest informative experiment or analysis, specifying comparisons, measurements, and how results guide the next decision; hand the plan to an execution agent. |
| belief-update | Incorporate new papers or results into hypotheses and research decisions, recording the evidence and preserving decision history. |
| red-team | Conduct an adversarial review of a developed idea, prioritize consequential objections, and propose concrete probes to resolve them. |
Architecture · Research state · Examples
MIT © 2026 HytroZero.