Research Lab

Research Lab

Use this for research-level work where the answer should be selected from evidence, not intuition. The workflow adapts to software, data analysis, academic synthesis, product strategy, policy, writing, or other domains.

Scope

• Work on one research item at a time.
• Keep research artifacts under research/<slug>/ unless the user or repo has an existing equivalent location.
• Use the local project, user-provided sources, web sources, datasets, benchmarks, interviews, or examples as evidence depending on the domain.
• Ask questions only when the research item, constraints, or acceptable output cannot be inferred.

Load References

• Load references/artifacts.md before creating or resuming a research folder.
• Load references/phases.md for the detailed phase-by-phase workflow.
• Load references/domain-adaptation.md when the theme is not software or the evidence and metrics need domain-specific translation.
• Load references/subagent-orchestration.md when the user asks to use subagents to run framework phases.

Startup

1. Identify the target research item, theme/domain, expected final output, user constraints, and any existing tracker or checklist item.
2. Create a stable slug from the research item.
3. Create or resume research/<slug>/; read state.md, decision.md, and prior runs/*/results.md if present.
4. If the user asks for one phase, complete only that phase and write the handoff. Otherwise, run phases in order while feasible and stop at blockers, missing evidence, or an explicit decision gate.
5. If subagent phase delegation is requested, keep the main agent as owner: delegate one phase at a time, review the result, then continue or fix.

Phase Flow

1. Prepare: create the workspace, capture scope, list hypotheses, and record open questions.
2. Specify: define inputs, outputs, success criteria, evidence sets, metrics, edge cases, and validation requirements.
3. Baseline: build or document the simplest credible baseline and measure it.
4. Explore: test one new or refined alternative per loop without repeating old approaches.
5. Compare: evaluate viable approaches under a shared matrix of criteria, workloads, sources, or cases.
6. Decide: select the approach or conclusion only when evidence gates pass; otherwise send the task back to Explore or Compare.
7. Apply: create the final implementation, recommendation, report, policy, design, or other artifact exactly from the decision scope.
8. Review: rerun targeted validation, compare against the saved decision, and reopen the loop if the result regresses or evidence no longer holds.

Non-Negotiable Rules

• Do not skip the spec, baseline, and decision phases unless the exception is documented in decision.md or state.md with a reason.
• Do not create final target-artifact changes or final user-facing conclusions before decision.md exists, unless the user explicitly asks for a lightweight answer.
• Ground claims in observed evidence and record enough provenance to reproduce or audit the conclusion.
• Use comparable methods across runs; record environment, sources, assumptions, and limits.
• Update state.md after every completed phase with status, evidence summary, open questions, and next phase.
• Preserve prior research artifacts on resume; append new runs instead of rewriting history.
• Do not mark a delegated phase complete until the main agent has reviewed the subagent output and required validation passes.

Deliverables

• research/<slug>/README.md
• research/<slug>/state.md
• research/<slug>/hypotheses.md
• research/<slug>/runs/<timestamp>-<phase>/plan.md
• research/<slug>/runs/<timestamp>-<phase>/results.md
• research/<slug>/runs/<timestamp>-<phase>/metadata.json when measurements, sources, or reproducibility details need structured capture
• research/<slug>/decision.md before final application

Handoff

When stopping, report the research path, completed phase, current best strategy or conclusion, missing evidence, and exact next phase.

References

• references/artifacts.md
• references/phases.md
• references/domain-adaptation.md
• references/subagent-orchestration.md