Run end-to-end research workflows across hypothesis framing, experiment design/execution, decision synthesis, and diagnostics. Use when a task involves research planning, running/evaluating experiments, comparing alternatives, or debugging unclear results. Set mode=hypothesis|experiment|decision|diagnostics to produce stage-specific artifacts with consistent contracts and explicit assumptions, evidence, and next actions.
<when_to_use> Use for core lifecycle stages: hypothesis framing, experiment design, decision-focused evaluation, or diagnostics. </when_to_use>
<source_of_truth>
Follow .grd/workflows/research-pipeline.md stages 1/2/3/3.5 and .grd/templates/research-artifact-format.md.
</source_of_truth>
<bundled_references>
references/mode-contracts.md for stage-specific deliverables and required fields.references/artifact-paths-and-alias.md for run alias and artifact path rules.
</bundled_references><clarification_rule> Ask one high-leverage question only when missing mode/ids/decision thresholds materially change outputs. </clarification_rule>
<context_policy>
.grd/STATE.md before drafting:
snapshot, held-out action_matching loss, potential-function condition), mirror exact wording in downstream prompts and questions..grd/STATE.md read and state-relevant terms/constraints either applied or explicitly justified as irrelevant.<template_convention>
.grd/templates/.state.md, roadmap.md, research-notes.md, run-index.md, research-artifact-format.md, deep-question.md).assets/templates/ only for genuinely skill-specific variants or overrides.<intent_lock>
<questioning_loop> defines the ambiguity-resolution loop (prefer 1 focused question per turn, cap 2 if tightly coupled, stop once next action is clear); <source_of_truth> is the canonical file/path contract declared by each skill.<questioning_loop>
Only run this loop when missing information would materially change:
Protocol:
Do not force users into provided options; options are scaffolding, not constraints. </questioning_loop>
<precision_contract>
<anti_enterprise> NEVER include phases for:
If it sounds like corporate PM theater, delete it. </anti_enterprise>
<delivery_rule> Default to concise chat output.
<output_format> Always structure the response as:
If the profile adds extra numbered items, keep their order after item 5.
If the skill defines additional required sections (for example, evidence taxonomy or artifact tables), include them after the last numbered item in this profile.
For Markdown output containing math, use $...$ inline and $$...$$ for display math.
Keep language execution-centric: concrete file paths, exact commands, and explicit done criteria.
6) Execution record
If the skill defines additional required sections, include them after item 6. </output_format>
<action_policy> Default: propose actions; do not execute side-effecting steps unless user explicitly asks. Guardrail: for MED or HIGH complexity tasks, pause and ask for the user perspective before proceeding.
Risk tiers:
Execution confirmation rule:
Contract:
<execution_contract>
run_id, hypothesis_id, target criteria).references/mode-contracts.md.references/artifact-paths-and-alias.md when run context is active.