Publishable Research Orchestrator

Turn a vague research idea into a publication-grade execution plan.

publishable-research-orchestrator is a Codex skill for multi-agent technical research. It helps Codex discover promising research directions, score them for publishability, reject weak options, and translate the winning idea into a concrete execution program with milestones, exit criteria, and artifacts.

It is designed for AI-heavy and technical domains such as:

• LLM systems
• agent workflows
• software engineering research
• benchmarking and evaluation
• ML infrastructure and tooling
• adjacent computer science topics

Why This Exists

Most research assistants can collect papers. Far fewer can answer the harder questions:

• Which topic is actually publishable?
• Which direction is new enough without being unrealistic?
• What is the minimum publishable unit?
• What should happen in the next 72 hours, 2 weeks, and 6 weeks?

This skill is built to make those decisions explicit.

What The Skill Produces

By default, the skill aims to leave behind a complete decision package:

1. A candidate landscape with 3-7 plausible directions.
2. A publishability scorecard for each candidate.
3. One selected direction with a clear thesis.
4. A step-by-step execution plan with milestones and fallback paths.
5. A near-term action plan for 72 hours, 2 weeks, and 6 weeks.

Core Capabilities

• Multi-agent research orchestration when the user explicitly authorizes subagents.
• Research-gap discovery grounded in recent literature and primary sources.
• Publishability scoring using a structured rubric.
• Candidate down-selection with explicit rejection logic.
• Milestone design for execution, experiments, and write-up.
• Workspace scaffolding for repeatable research packages.

Research Modes

Mode	Use When
discovery	You only have a broad area and need candidate directions.
down-select	You already have several ideas and need ranking and rejection logic.
program-design	You know the likely direction and need a precise execution plan.
execution	You want the work to begin: artifacts, experiments, or writing support.

How It Works

The skill follows a compact but disciplined workflow:

1. Frame the research space.
2. Shape the problem into a researchable question.
3. Run parallel reconnaissance when multi-agent work is explicitly allowed.
4. Build a candidate matrix.
5. Score candidates with a publishability rubric.
6. Select one direction and lock the thesis.
7. Break the direction into milestones and artifacts.
8. Scaffold a working research pack if execution is requested.

Detailed playbooks live here:

• references/research-playbook.md
• references/publishability-rubric.md
• references/output-contract.md

Quick Start

Place this repository inside your Codex skills directory, or symlink it there:

git clone https://github.com/Nigmat-future/publishable-research-orchestrator.git

Then invoke it from Codex with prompts like:

Use $publishable-research-orchestrator to find 5 publishable research directions in AI agents and select the strongest one.

Use $publishable-research-orchestrator to compare three benchmark ideas for code review agents and produce a 6-week execution plan.

Use $publishable-research-orchestrator to identify a publication-grade topic in LLM evaluation and scaffold the research workspace.

Workspace Scaffold

The repository includes a helper script to generate a standard research pack:

python .\scripts\scaffold_research_pack.py --topic "agentic code review benchmarks" --output .

That produces a dated workspace with:

• 00-project-brief.md
• 01-landscape-map.md
• 02-topic-scorecard.md
• 03-selected-direction.md
• 04-execution-plan.md
• 05-experiment-matrix.md
• 06-writing-outline.md
• notes/source-log.md
• notes/decision-log.md

Repository Structure

publishable-research-orchestrator/
├── SKILL.md
├── README.md
├── LICENSE
├── assets/
│   └── banner.svg
├── agents/
│   └── openai.yaml
├── references/
│   ├── output-contract.md
│   ├── publishability-rubric.md
│   └── research-playbook.md
└── scripts/
    └── scaffold_research_pack.py

Design Principles

• Prefer decision-quality over literature dumping.
• Prefer primary sources for technical claims.
• Prefer falsifiable ideas over vague novelty.
• Prefer realistic publishability over fashionable wording.
• Keep synthesis local even when research is parallelized.

Best Fit

This skill works best when:

• the topic is technical and fast-moving
• recency matters
• you want explicit selection logic
• you need a plan that can actually be executed

This skill is not ideal for:

• simple factual lookups
• one-off debugging questions
• casual brainstorming without evidence
• non-technical domains that require field-specific publication norms

License

Released under the MIT License.