---
name: research-pipeline
description: Coordinate research from a defined question to reproducible analyses and communication using explicit stage inputs, decisions, handoffs, and completion evidence.
---

# Research pipeline

Use this method when work spans multiple research stages or needs a coherent route from a question to an artifact. Apply the [research contract](../../docs/RESEARCH-CONTRACT.md). Choose stages for the actual research design; a pipeline is an accountable coordination structure, not a claim that every discipline follows the same linear sequence.

## Establish scope and stage contracts

Record the research question, intended outcome, audience, available material, constraints, collaborators, and authorized actions. Distinguish exploratory work, confirmatory work, replication, synthesis, and artifact preparation. Identify dependencies that require human decisions, data access, ethical review, specialist judgment, or external services.

Define a contract for each needed stage: inputs with provenance; question or decision; method; output format; acceptance evidence; owner if known; and unresolved issues passed forward. An output becomes the next stage's input only when its limitations travel with it. Preserve versions or stable identifiers where useful so a later figure can be traced to an analysis and dataset.

A practical initial map may include question framing, evidence discovery, design, acquisition, analysis or derivation, validation, interpretation, drafting, artifact production, and release checks. Omit irrelevant stages and allow feedback. A failed validation can return work to design or modeling; it should not be treated as a formatting problem.

## Execute with discipline-specific handoffs

For empirical quantitative work, connect the question to population, measures, design, data handling, and analysis decisions. Carry inclusion rules, missingness, units, uncertainty, and exploratory choices into reporting. For qualitative work, connect question and methodological orientation to recruitment, context, material collection, reflexivity, analytic decisions, and interpretation. Preserve an audit trail without reducing every design to the same checklist.

For theory, hand off definitions and assumptions into lemmas, propositions, proofs, counterexamples, and regime analysis. Keep conjectures and heuristic steps distinct. For computational work, preserve model equations, conventions, inputs, algorithm choices, code versions, environment details, tolerances, and generated results. Connect each reported quantity to an identifiable computation or explicitly supplied value.

Use [scientific reasoning](../scientific-reasoning/SKILL.md) to clarify inferential gaps, [mathematical modeling](../mathematical-modeling/SKILL.md) to construct formal models, and [research validation](../research-validation/SKILL.md) to test claims and artifacts. No method is presumed superior across every task; choose it for the question, evidence, and relevant failure modes.

## Manage evidence and progress

Maintain a compact project ledger: artifact; source or parent artifact; current status; check performed; next dependency. Distinguish planned, drafted, executed, inspected, and accepted states. A generated script is not an executed analysis, a rendered document is not a visually inspected document, and a citation string is not a verified reference.

Batch independent read-only work when it helps, but sequence decisions that depend on previous findings. Parallel agent work, external actions, publication, or messages must follow the user's authorization and available tools. When evidence is missing, continue independent work and mark the dependent branch. Do not silently substitute guessed data or a different research question.

At each handoff, summarize what changed, why, what evidence supports it, and what remains uncertain. Avoid large status reports that obscure the next consequential decision. Define stopping conditions in terms of deliverables and checks, with resource bounds as separate limits. Iteration counts and elapsed time alone do not establish scientific success.

## Deliver and verify

Deliver the requested research artifact plus a concise provenance and status record. For manuscripts, use [manuscript writing](../manuscript-writing/SKILL.md); for final files, use [document production](../document-production/SKILL.md). Include editable sources and essential supporting material when authorized and expected.

Before calling the pipeline complete, trace principal conclusions backward to evidence and forward to every public-facing summary. Reconcile unresolved issues across stages, verify links between code, data, tables, and text where accessible, and identify checks not performed. State whether the result is a plan, a partial execution, a validated local result, or a prepared release package. Completion of the coordination task must not inflate the scientific status of its contents.