John Barrios · empirical workflow
Skills teach how — Compustat filters, CIK discipline, attrition rows. MCP servers run live access — authenticated WRDS queries, EDGAR filings, macro series, with credentials in .env and never in chat. Neither replaces verification: agents make pipelines cheap to produce, and the costly part stays yours.
The small set for a first Compustat firm–year paper, with optional EDGAR text and a Stata or Python estimator. Everything else in the 46-skill catalog (Bayesian, geospatial, posters) can wait. Browse the full catalog when you need it.
| Role | Skills |
|---|---|
| Data access conventions | wrds; add sec-edgar if any filing work |
| Panel & estimate | Stata lane: stata-data-cleaning + stata-regression; Python lane: pyfixest + python-panel-data |
| Output | latex-tables, econ-visualization |
| Write & check | econ-write, econ-humanizer, econ-referee |
| Priority | Server |
|---|---|
| Required for WRDS papers | wrds-mcp setup |
| Optional | SEC EDGAR MCP · OpenEcon Data / FRED · Stata-MCP (install guide) |
These are habits, not packages. No install fixes their absence.
If the bottleneck is knowing the convention, use a skill; if it is talking to a remote system with secrets or rate limits, use MCP.
| Research task | Skill | MCP | Use |
|---|---|---|---|
| Compustat / CRSP / CCM pull | wrds | wrds-mcp | Both: skill for filters and query patterns; MCP for auth and live SQL |
| EDGAR filings, Form 4, XBRL facts | sec-edgar | SEC EDGAR MCP | Both: CIK and rate-limit conventions plus live fetch |
| FinBERT / 10-K text features | financial-text-nlp | EDGAR MCP only to obtain text | Skill (+ MCP upstream) |
| FRED / World Bank / IMF series | api-data-fetcher | OpenEcon Data or FRED MCP | Either / both: MCP to discover; skill for pinned reproducible scripts |
| Stata clean / reghdfe / esttab | stata-data-cleaning, stata-regression | Stata-MCP | Both if the agent drives Stata; skill only if you run .do files yourself |
| Panel FE in Python | pyfixest, python-panel-data | none | Skill only |
| Lit search / BibTeX | econ-lit-search, citation-management | none | Skill only |
| Figures / LaTeX tables | econ-visualization, latex-tables | none | Skill only |
| Paper / slides / referee prep | econ-write, econ-slides, econ-referee | none | Skill only |
| Choosing the tool rung | — | — | Chat for thinking; editor for edits; agent for pipelines; agent + MCP only when live licensed or public APIs are required |
Each stage lists its default skills, the MCP server if one applies, and the verification gate you pass before moving on. Chip text opens the skill in the catalog; the arrow downloads the zip.
Sharpen a question that Compustat, CRSP, EDGAR, or a public panel can answer; map prior designs and measurement conventions before writing any code.
Name the identifying variation and the unit of observation (firm–year, filing, event) in one sentence. If you cannot, do not open a data session yet.
Start in a clean directory and write the six-part prompt before any code runs. Require named scripts, DECISIONS.md, LOG.md, and attrition rows, not chat-only results.
qtrs, form type, Compustat iron laws)DECISIONS.md, LOG.md, attrition printed on every filterwrds / sec-edgar here so the prompt carries the iron laws and CIK discipline from the start.
Paste the six headings. Fill in every blank an agent would have to guess: years, tags, filters, join key, filename.
Pull WRDS extracts, EDGAR filings and XBRL facts, and macro controls into data/raw/. Every pull, authenticated or bulk, ends as files on disk.
.env, never in chat. The MCP server reads WRDS_USERNAME and related variables from an env file with chmod 600. If a password ever appears in a prompt or a log, rotate it.
Four inspections before trusting any extract: (1) row count vs an expectation written down beforehand, (2) a sample of rows, (3) null check on the analysis columns, (4) date-range min/max. Compustat SQL must include indfmt='INDL', datafmt='STD', popsrc='D', consol='C'. CRSP share-type and exchange filters are manual. Never assume a helper applied them.
Turn raw extracts into an analysis-ready firm–year panel. Join on the correct key: gvkey–fyear, adsh, the CCM link table. Never join on hand-typed tickers. Print attrition at every filter.
One join check: n before, n after, and the reason for every drop. Flag unmatched keys. No silent row inflation from many-to-many joins.
CIK resolution → filing retrieval (cached HTML) → section extraction with fallback patterns → quality report → text feature or NLP score. Construct validity is the accounting contribution; automation is not the finding. Resolve and pin CIKs: ticker strings are not stable firm ids across EDGAR joins.
Read the quality report: the failures and the suspiciously short sections, not the headline success rate. Before a text measure enters a main table, run a stratified human audit (~50 filings across measure level, size, and industry) against a written rubric.
Build ratios, treatments, instruments, and controls; winsorize or trim with documented rules; plot distributions and known-event windows before any regression.
Name two checks before looking at the figure: (1) benchmark industries or firms where you hold a strong prior, (2) plausibility around a known shock. Write both into the README.
Run the main specs and pre-committed robustness: FE OLS, Poisson, IV, DiD including modern estimators, clustering. Prefer high-dimensional FE tools that match accounting panel practice: pyfixest if Python, stata-regression if Stata.
Check one coefficient or N from the log file against an independent re-run or a hand calculation on a tiny subsample. Confirm FE and cluster choices match DECISIONS.md.
Export publication tables and figures from scripts. Every manuscript number comes from a generated file, \input{} or a read CSV, never hand-typed from a terminal. Finish the README so a stranger can rerun the project from empty folder to figure.
Pick one table cell. Check it against the script output file and confirm the manuscript inputs that file. No copied digits.
Draft and revise, strip AI tells, build the seminar deck, and run an adversarial pre-submission pass. The default writing sequence: econ-write → econ-writing-plus for identification prose → econ-humanizer (plus econ-humanizer-plus) for voice.
Cross-check one DECISIONS.md claim against the git commit it cites. Run a credential audit: grep the history for secrets; it should come back empty. On identification, the agent is an editor, not a rewriter: comments only, no ghostwritten argument.
Agents break the old correlation between effort and checking: output that once took a week of RA time now appears in minutes, and the checking cost did not fall with it. The danger sits exactly where this catalog is strongest: text measures from financial-text-nlp plus EDGAR, dense pyfixest / stata-regression batteries, and fluent econ-write drafts. All of it is cheap to produce and expensive to validate.
Pay the debt with the stage gates above, not with more autonomy. Row counts, iron-law filter checks, and date ranges are cheap; run them every time. Human audits of text measures and independent re-runs of headline coefficients are expensive; budget for them before the result enters a main table.