ucl-erl/skills

What This Is

ERL Research Skills is a public agent capability system for AI coding agents and research assistants. It connects literature discovery, evidence synthesis, research planning, remote experiments, paper writing, scientific review, submission, and Robotics & AI engineering into inspectable workflows.

This is not a prompt dump. The repository separates task-specific skills from persistent agent instruction contracts so users can choose both what an agent can do and how it should behave.

Agent Instruction Contracts

Skills define what workflow to run. Instruction contracts define how the agent behaves throughout the work.

These are persistent instruction policies, not invokable skills and not entries in catalog.json. Read the selection and installation guide before copying or merging one into a project.

For example, a survey workspace can load the Survey Writing and Audit Guide as its persistent contract, then invoke task workflows only when needed:

persistent behavior: survey-writing-audit/CODEX.md
task workflow:       rai-research-flow -> survey-synthesis-builder
focused audits:      citation-integrity-auditor / manuscript-structure-auditor

Flagship Workflows

Literature to Research Direction

rai-research-flow
  -> paper-search-protocol
  -> paper-reading-card
  -> evidence-matrix-builder
  -> research-idea-rubric / survey-synthesis-builder

Build a reproducible literature base, compare evidence, identify defensible gaps, and turn the result into a research direction or survey structure.

Experiments to Paper

experiment-dossier-builder
  -> paper-code-consistency-auditor
  -> paper-draft-builder
  -> scientific-figure-director

Package remote code and experiment results into a portable dossier, verify manuscript-code consistency, and construct a source-grounded paper narrative.

Paper to Submission

rai-paper-flow
  -> manuscript-structure-auditor
  -> citation-integrity-auditor
  -> paper-red-team-review
  -> reviewer-response-builder / latex-submission-checker

Audit logic, citations, evidence, scientific writing, reviewer risk, rebuttal commitments, and submission readiness without changing technical claims silently.

Robotics and AI code work is routed through rai-coding-flow and robotics-ai-coding-flow, with explicit checks for debugging, experiments, reproducibility, and deployment risk.

Research Lifecycle

flowchart LR
    A["Discover"] --> B["Map Evidence"]
    B --> C["Ideate"]
    C --> D["Design"]
    D --> E["Write"]
    E --> F["Audit"]
    F --> G["Package"]
    D --> H["Build & Experiment"]
    H --> F

Router skills select the minimum useful path. Atomic skills perform focused tasks. Reference skills supply shared standards. Tool skills add output-specific protocols and validation.

Get Started

Clone the repository and validate the system:

git clone https://github.com/UCL-ERL/skills.git
cd skills
python scripts/validate_catalog.py
python scripts/validate_forward_tests.py

Install the skill directories using the mechanism supported by your agent client, then invoke a router or atomic skill by name. Start with:

See docs/usage.md for installation patterns, invocation examples, paper modes, and forward-testing guidance.

Daily Paper Modes

rai-paper-flow supports three levels of effort:

The default is quick. Heavier modes are opt-in so routine research work stays efficient.

Capability Map

Browse all 28 skills in the Skill Catalog.

System Design

The repository has two complementary product surfaces:

Skills use four internal layers:

The rai-* prefix means Robotics & AI. These stable skill IDs describe the domain and are independent of repository ownership.

Read docs/architecture.md, docs/curation-policy.md, and docs/quality-rubric.md for the full design and acceptance rules.

Quality and Status

• Source-grounded claims: do not invent citations, benchmarks, APIs, or experimental results.
• Checkable completion: every skill must define what finished work looks like.
• Narrow boundaries: prefer focused skills and thin routers over mega-skills.
• Public reuse: no credentials, private paths, restricted data, or undocumented services.
• Inspectable provenance: record external inspiration without copying third-party prose.

The repository currently contains 28 draft skills, 3 agent instruction contracts, and 20 forward-test fixtures. draft is an explicit skill maturity label, not a claim of production stability. Skills move to beta or stable only when supported by realistic use evidence under the quality rubric.

Repository Map

.
|-- catalog.json                 # Machine-readable registry
|-- catalog.schema.json          # Catalog schema
|-- agent-instructions/          # Persistent agent behavior contracts
|-- docs/                        # Architecture, catalog, policies, and usage
|-- examples/                    # Inspectable examples and artifact shapes
|-- forward-tests/               # Manual forward-test prompts and pass criteria
|-- skills/                      # Published skills
|-- templates/SKILL.md           # Skill authoring template
|-- scripts/                     # Catalog and forward-test validators
`-- .github/                     # CI, ownership, and contribution workflows

About ERL Lab

ERL Lab (Embodied Reinforcement Learning Lab) is a research group within UCL Robotics & AI at University College London. Its research focuses on:

• Embodied Reinforcement Learning: reinforcement learning for robotic skill acquisition, decision-making, and adaptation under real-world uncertainty.
• Generalizable Robot Learning: transferable, composable, and continually improving robot skills.
• Foundation Models for Robotics: VLA models, multimodal models, and embodied agents for robotic learning and evaluation.

Rigorous evaluation, reproducibility, and open-source research are shared principles across these directions.

Visit the ERL Lab homepage for research themes, people, projects, and updates.

Contribute

ERL Research Skills accepts public issues, discussions, and pull requests. External contribution is open; roadmap, scope, quality standards, releases, and maintainer appointments remain governed by ERL Lab.

Read CONTRIBUTING.md, GOVERNANCE.md, and CODE_OF_CONDUCT.md before contributing.

Contributors

Current contributors are recorded in CONTRIBUTORS.md according to their actual work.

Star History

License

Released under the MIT License.