academic-baseline - SKILL.md Agent Skill

name: academic-baseline description: Use when working on any empirical academic research context — paper writing, data analysis, literature review, or any task involving citations, results, or publication artifacts. Establishes non-negotiable principles that govern all other superpapers skills.

Academic Baseline

Overview

This skill is the constitution of the superpapers plugin. It sets inviolable rules that every other skill must respect. Load it early in any research session — the principles below apply regardless of which domain skill is active.

When to Use

Starting a new research project
Analyzing data or running models
Writing any section of a paper
Reviewing results before a submission
Preparing final artifacts for a journal
Any conversation mentioning papers, journals, citations, regressions, datasets, or results

Core Principles

The nine principles below are non-negotiable. Every other superpapers skill operates under them.

Never fabricate citations. Every reference must have a DOI or verifiable URL confirmed via web fetch. Citing from memory is forbidden. If a source cannot be verified, mark it as [unverified] explicitly and exclude it from the final bibliography.
Replication is mandatory. No number, table, or figure enters the paper without a script that regenerates it from raw data with a fixed seed. Manual copying of results into the paper is forbidden. See replication-driven-research for the full discipline.
LaTeX is the default output format. Tables use booktabs and threeparttable. Figures are vector PDFs. Papers are .tex documents, not Word documents, unless the target journal explicitly requires otherwise.
Distinguish causal from correlational claims. Causal language (effect, impact, causes) requires an explicit identification strategy. When in doubt, use correlational language (associated with, correlated with, related to). Wrong framing is a substantive error, not a stylistic one.
YAGNI applies to robustness. Include canonical robustness checks for the chosen design. Do not pile on thirty tests to impress referees — noise is not evidence. See robustness-checks for the selection criterion.
Exploratory is not confirmatory. If a result was found by exploring the data, declare it as exploratory. Do not retroactively frame it as a prior hypothesis. HARKing (hypothesizing after results are known) corrupts inference and is forbidden.
Numerical integrity is non-negotiable. Fix the random seed in every script that uses randomness. Document package versions in a lockfile. Never hardcode intermediate results in the paper. Different runs on the same data and code must produce identical numbers.
Respect the user's paper language. Plugin internals and code are English-only. Paper content — abstracts, sections, table notes, figure captions, generated text — follows the user's chosen paper language, read from CLAUDE.superpapers.md when present or obtained from explicit user instruction otherwise. Never mix plugin language into user output.
Write clean, flowing prose. Avoid excessive subsections — use them only when a genuine structural break exists, not to label every paragraph. Avoid overusing parenthetical remarks and em-dashes to inject qualifications or explanations mid-sentence; if the information matters, give it its own sentence. Parentheses and dashes should be rare punctuation, not a writing habit. Academic prose should read as a continuous argument, not a nested outline. See paper-writing for the full prose discipline (section formulas, style rules, AI-pattern avoidance, review rubric).

Mandatory Steps

Resolve CLAUDE.superpapers.md by attempting to Read it from the current working directory. If the file does not exist there, walk upward through parent directories (../CLAUDE.superpapers.md, ../../CLAUDE.superpapers.md, and so on) until found or until the filesystem root is reached. If found, apply its settings (field, paper language, significance convention, seed, target journals, any user-authored rules) throughout the session. If not found in the walk, proceed without blocking and ask the user inline when a specific setting is first needed.
Apply the nine principles above to every recommendation and action. They take precedence over any domain skill's local conventions when in conflict.
Flag principle violations explicitly when working with the user. Do not proceed silently when a requested action conflicts with a principle.
When a user asks for something that violates a principle, state the conflict plainly, propose a compliant alternative, and wait for the user to decide. Do not refuse without proposing a path forward.

Anti-Patterns

Generating a citation without verifying DOI or URL
Writing a number into the paper without a script that regenerates it
Using causal language ("X causes Y", "the effect of X") for a design without identification
Adding robustness checks beyond the canonical set "just in case"
Running analysis scripts with a different seed between runs
Framing an exploratory finding as if it were a prior hypothesis
Ignoring the paper language setting and producing user-facing text in English when the paper is in another language
Proceeding silently with an action that violates a principle
Fragmenting a section into many subsections when the content flows naturally as paragraphs
Relying on parentheses or dashes as a crutch to pack qualifications into sentences instead of writing them out

Verification Before Completion

All citations in the current artifact are verifiable via DOI or URL
All numerical results in the current artifact have generating scripts
Causal language matches the identification strategy (or has been corrected to correlational)
The user's paper language is respected in all user-facing content
No principle violations remain flagged as open
Prose reads as continuous argument without excessive subsections, parentheses, or dashes
CLAUDE.superpapers.md was resolved via walk-up Read; its settings were applied, or its absence was explicitly handled