| name | Build Great Research |
| description | Comprehensive playbook for building rigorous, publication-quality research papers. Covers the full research output sequence from paper structure and statistical reporting through table formatting, figure design, LaTeX conventions, academic writing quality, and pre-submission review. |
| phase | work |
How to Build Great Research
You are producing a research paper that will be read, reviewed, and cited by academics. Your default instinct is to get the analysis working and move on. That instinct produces mediocre output. This skill exists to override that instinct.
The north star is rigor and clarity. Publication-quality papers are not just correct -- they are intentional in every detail: tables that communicate findings precisely, figures that reveal patterns, prose that argues cleanly, and methods sections that enable replication. The difference between "technically correct" and "publishable" is the discipline to never skip the steps that feel optional but are not.
You have unlimited time. You are a long-running agent. There is no deadline pressure. There is no excuse for cutting corners. Every phase below exists because skipping it produces output that a reviewer will immediately flag.
Research Quality Philosophy
Two intellectual foundations underpin this skill -- one for methodological rigor, one for presentation clarity.
Reproducibility is non-negotiable
Every result in the paper must be traceable from raw data to final output. This means:
- Pipeline over manual steps. Every number in the paper is generated by code, never manually entered. A reviewer should be able to run the pipeline and get identical results.
- Decisions are logged. Every methodological choice (estimation method, variable operationalization, sample restriction) is documented in
docs/decisions/ with rationale.
- Seeds are pinned. Any stochastic process (bootstrap, simulation, imputation) uses documented random seeds for exact reproducibility.
- Dependencies are locked. The Python environment is fully reproducible via
uv.lock.
Clarity compounds (Strunk & White for academics)
- Every sentence earns its place. Cut words that do not add information. "It is important to note that" adds nothing. Say the thing directly.
- Tables and figures stand alone. A reader should understand any table or figure without reading the surrounding text. Titles, notes, and labels must be self-contained.
- Methods enable replication. Another researcher with the same data should reproduce your results using only the methods section.
- Structure guides the reader. Section flow follows the logic of the argument, not the chronology of the analysis.
The Build Sequence
Every great paper follows this sequence. Do not skip phases. Do not jump to results before the methodology is solid. Do not submit without completing the polish phase.
| Phase | Focus |
|---|
| 01 -- Foundation | Research question, hypotheses, variable operationalization, data source selection |
| 02 -- Structure | Paper outline, section dependencies, argument flow, LaTeX skeleton |
| 03 -- Analysis | Estimation, robustness checks, sensitivity analysis, all results generated via pipeline |
| 04 -- Tables and Figures | Formatted outputs in paper/outputs/tables/ and paper/outputs/figures/ |
| 05 -- Writing | Methods section, results narration, introduction, conclusion, abstract (in that order) |
| 06 -- Polish | Internal consistency check, reference verification, formatting review, pre-submission audit |
Phase 01: Foundation
- Define the research question precisely
- State hypotheses with expected direction and mechanism
- Operationalize all variables (measurement, source, transformation) -- document in variable codebook
- Select data sources and document in
docs/decisions/
- Define the sample and exclusion criteria with justification
Phase 02: Structure
- Create
paper/main.tex with section skeleton
- Organize sections in
paper/sections/ (introduction.tex, literature.tex, methodology.tex, results.tex, discussion.tex, conclusion.tex)
- Map the argument flow: each section should have a clear purpose statement
- Set up
paper/outputs/tables/ and paper/outputs/figures/ directories
- Configure bibliography in
paper/references.bib
Phase 03: Analysis
- Implement analysis pipeline in
src/analysis/ with modular scripts
- Run main estimation and store results programmatically
- Execute robustness checks (alternative specifications, subsamples, different estimators)
- Run sensitivity analysis for key assumptions
- All outputs written to
paper/outputs/ -- never manually entered into LaTeX
Phase 04: Tables and Figures
- Tables use consistent formatting: clear headers, aligned decimals, standard error notation, significance stars with footnote legend
- Figures use readable labels, appropriate scales, colorblind-safe palettes
- Every table and figure has a descriptive title and explanatory notes
- Outputs are generated by pipeline scripts, not manually created
- File naming convention:
table_01_descriptive_stats.tex, figure_01_main_effect.pdf
Phase 05: Writing
Write sections in this order (not document order):
- Methodology: Write first while analysis is fresh. Enable replication.
- Results: Narrate findings. Reference tables and figures. State magnitudes, not just significance.
- Discussion: Interpret results. Address limitations honestly. Compare with prior literature.
- Literature review: Position the contribution. Only include papers that inform the argument.
- Introduction: Write last among body sections. Motivate, preview, and hook.
- Conclusion: Summarize contribution, limitations, and future work.
- Abstract: Write dead last. 150-250 words covering question, method, key finding, implication.
Phase 06: Polish
- Verify all cross-references resolve (
\ref, \cite, \label)
- Check internal consistency: numbers in text match tables
- Verify all tables and figures are referenced in the text
- Run
drl:compile to verify LaTeX builds cleanly
- Check formatting against target journal guidelines
- Proofread for clarity, concision, and academic tone
Statistical Reporting Standards
Regression Tables
- Report coefficients with standard errors (in parentheses) or confidence intervals
- Include number of observations, R-squared (or pseudo-R-squared), and F-statistic
- Note clustering level for standard errors
- Use stars sparingly: *** p<0.01, ** p<0.05, * p<0.10 (or state chosen threshold)
- Always include a table note explaining the specification
Descriptive Statistics
- Report N, mean, SD, min, max for continuous variables
- Report N, proportion for categorical variables
- Note missing data counts and handling strategy
Figures
- Label axes with variable name and units
- Include data source in figure notes
- Use consistent styling across all figures
- Export as PDF (vector) for print quality, not PNG
Mandatory Quality Checklist
Run this before declaring ANY paper build complete.
Reproducibility
Tables and Figures
Writing
Internal Consistency
Robustness
Common Pitfalls
- Manual number entry: Copying results from console output into LaTeX. Pipeline must generate all outputs programmatically.
- Significance-only reporting: Reporting p-values without effect sizes and confidence intervals. Always report magnitudes.
- Missing robustness: Reporting only the specification that "works." Reviewers will ask for alternatives.
- Opaque methods: Writing methods that describe what was done but not how to replicate it. Include estimation equations, software versions, and exact variable definitions.
- Orphaned outputs: Tables or figures in
paper/outputs/ that are not referenced in the text, or references to outputs that do not exist.
- Inconsistent notation: Switching between variable names, abbreviations, or symbols across sections. Define once, use consistently.
- Abstract first: Writing the abstract before the paper is complete. The abstract should be the last thing written.
- Monolithic analysis scripts: Single files mixing data cleaning, estimation, and output generation. Separate into modular pipeline stages.
Memory Integration
- Before starting:
drl search "research paper" for relevant lessons
- After completing: offer
drl learn to capture insights about what worked
Quality Criteria
- All 6 phases completed in order
- Pipeline generates all paper outputs programmatically
- Every methodological decision logged in
docs/decisions/
- Tables and figures follow statistical reporting standards
- LaTeX compiles cleanly via
drl:compile
- Internal consistency verified (prose matches tables)
- At least one robustness check reported
- Quality checklist completed with all items checked
