This skill executes a research plan phase by phase. It starts by invoking academic-baseline as the standing policy layer, dispatches subagents for independent tasks, invokes replication-driven-research as a guardrail, and runs two-stage review at phase boundaries: correctness first, reproducibility second. It must honor each task's declared Skills involved field during execution rather than improvising generic behavior. Stops and asks the user at any scaffolding or destructive action. The skill never proceeds past a phase until both reviews pass.
-
Invoke academic-baseline first. This resolves CLAUDE.superpapers.md via the walk-up Read (current working directory, then parent directories) and carries its settings into the session. Keep academic-baseline active through every phase.
-
Load and parse the plan. Read the plan file in full. Extract tasks, phases, dependencies, verification criteria, and Skills involved.
-
Invoke replication-driven-research to ensure project structure. If directories are missing, propose scaffolding. Wait for user confirmation before creating directories or files at the project root.
-
Pre-flight declaration before every task. Before executing any task, output a structured block:
## Task: <task title>
Phase: <phase name>
Skills (from routing table): <list from CLAUDE.superpapers.md table>
Skills (from plan): <list from plan's Skills involved field>
Will invoke: <union of both lists, no duplicates>
Invoke every skill in the union list before doing any task work. If paper-writing appears, confirm you are in the main session — never dispatch it to a subagent. If journal-guidelines appears, confirm the target journal is resolved — if not, invoke journal-selection first. If a task is missing a clearly necessary skill, stop and repair the plan before proceeding.
-
Enforce journal routing. Any task or review involving a target journal, author instructions, formatting, templates, blinding, cover letters, checklists, or submission portals must invoke journal-guidelines before work begins. If the outlet is not fixed yet, invoke journal-selection first, then journal-guidelines. Never declare journal compliance or submission readiness without this step.
-
Execute tasks phase by phase. Within a phase, independent tasks can be dispatched to subagents in parallel. Sequential tasks run in order.
-
Verify after every task. Run the task's verification command. If it fails, stop and diagnose — do not proceed silently.
-
Run end-to-end integration at each phase boundary. Execute the full pipeline from data/raw/ to the farthest artifact produced so far. Confirm exit code 0 and that all expected outputs exist.
-
Two-stage phase review:
- Stage 1 — correctness. Are the results right? Does the analysis make sense? Narrative review of the outputs, specs, and diagnostics.
- Stage 2 — reproducibility. Is the pipeline clean end-to-end? Is the seed fixed? Is the manifest updated? Use the
replication-driven-research verification checklist.
-
Only proceed to the next phase after both stages pass.
-
On failure: stop, diagnose root cause, fix, re-run from the failing task. Never skip failed tasks. Never re-run from scratch without understanding what broke.
-
Final full-pipeline run. Before declaring the plan complete, run everything from raw data to final PDF. All verifications must pass.
-
Report status to the user at phase boundaries. At the end, summarize what was built, what passed review, and where the artifacts live.
-
Suggest the pre-submission audit. After the final summary, recommend that the user run /superpapers:paper-review before submission. The paper-review skill performs a cross-cutting audit of prose, code, tables, figures, citations, and reproducibility and writes a consolidated report. This suggestion is non-blocking — do not auto-invoke and do not gate the plan's completion on it.
