Research Workflow

Seven-stage research workflow scaffolding for AI-assisted research processes. Each stage produces structured prompts and acceptance criteria for agent loops.

Stages

| # | Stage | Purpose | | --- | --- | --- | | 1 | SCOPE | Define research question, scope, and success criteria | | 2 | LITERATURE | Survey prior work; produce annotated bibliography | | 3 | REASON | Identify gaps, contradictions, and open questions | | 4 | DESIGN | Formulate methodology and experimental plan | | 5 | COMPUTE | Execute analysis; collect and validate results | | 6 | SYNTHESIZE | Interpret results relative to prior work | | 7 | WRITE | Draft manuscript sections and iterative revision |

Quick Start

python
from infrastructure.research import ResearchWorkflow

# Print all stage descriptions
ResearchWorkflow.describe()

# Get a structured prompt for a specific stage
prompt = ResearchWorkflow.prompt("SCOPE", question="What drives protein misfolding?")
print(prompt)

# Iterate stages in order
for stage in ResearchWorkflow.stages():
    print(f"{stage.index}. {stage.name}: {stage.description}")

Pipeline Orchestrator

bash
# Run research workflow scaffolding for a named project
uv run python scripts/pipeline/stage_10_research_workflow.py --project my_project

# Run only specific stages
uv run python scripts/pipeline/stage_10_research_workflow.py --project my_project --stages SCOPE,LITERATURE

# Generate stage prompts without executing
uv run python scripts/pipeline/stage_10_research_workflow.py --project my_project --dry-run

Config Integration

Set stage overrides in projects/{name}/manuscript/config.yaml:

yaml
research_workflow:
  question: "How do transformer attention patterns relate to syntactic structure?"
  stages:
    SCOPE:
      depth: detailed
    LITERATURE:
      max_papers: 50
      sources:
        - arxiv
        - semantic_scholar
    WRITE:
      target_venue: NeurIPS

Key Types

python
from infrastructure.research import (
    ResearchWorkflow,        # instance API — workflow.stage(name) / all_stages() / describe()
    ResearchStage,           # Stage descriptor: name, label, description, inputs, outputs, gate, status, order
    ResearchWorkflowConfig,  # Parsed config.yaml `research_workflow:` block
)

> Doc/code drift (tracked for follow-up): the "Stages" table, .prompt(), > .stages(), .render(), and .record_output() calls elsewhere in this > file describe an OpenScience-ported design (uppercase SCOPE/LITERATURE/... > stage names, per-stage sub-agent fan-out and prompt templates) that > ResearchWorkflow/ResearchStage do not implement — the real stage names > are lowercase (scope, survey, hypothesise, experiment, validate, review, > write) and there is no .prompt()/.stages()/.render() API. Treat > those sections as design intent, not a verified API surface.

Stage Prompt Generation

python
from infrastructure.research import ResearchWorkflow

# Generate LITERATURE stage prompt for a specific project
prompt = ResearchWorkflow.prompt(
    "LITERATURE",
    question="Neural scaling laws",
    context={"prior_papers": 12, "target_venue": "ICML"},
)

Agent Loop Integration

python
from infrastructure.research import ResearchWorkflow

# Use in an agent loop
for stage in ResearchWorkflow.stages():
    prompt = stage.render(project_context)
    response = llm_call(prompt)
    stage.record_output(response, output_dir="output/research/")

Testing

bash
uv run pytest tests/infra_tests/research/ -v

See Also

• AGENTS.md — operating contract and architecture
• ../search/SKILL.md — literature search integration
• ../llm/SKILL.md — LLM review integration