scdenney
UserAgentic skills for Claude Code and Codex, built from published social-science methods sources. Covers experimental design, computational text analysis, manuscript QA, and transparent reporting.
Categories
Indexed Skills (53)
deliverable-intake
Turns raw captures, dictated braindumps in a deliverable’s inbox/, or recent macwhspr dictation-log entries into proposed planning-wiki changes, shown as a unified diff for author acceptance or rejection. Classifies every capture line as a decision, evidence claim, question, task, or dropped item with a reason, and traces it to its source. Uncertain names and citations remain UNRESOLVED. Applies decisions only with explicit assent. Use when the user says "intake", "absorb my notes", "process the inbox", "I dictated something", or after notes from a talk, meeting, or reading session. Never edits source files, only planning/ and open questions.
deliverable-lint
Reviews a whole deliverable, such as a talk, course module, paper, chapter, or referee report, against its manifest without rewriting. Runs a deterministic gate for build, citations, numbers, facts, and leaks, then stops on failure. It assigns a cheap finder to each section, runs a strong-model whole-work pass, and adjudicates into checks/<date>/findings.json and report.md with anchors, severities, and honest NOT-CHECKED coverage within a dollar budget. Use when the user says "lint the deliverable", "review the whole talk/module/paper", "night shift", "pre-ship review", or before release. Revision remains serial and author-led.
deliverable-open
Opens a deliverable, such as a talk, course module, paper, book chapter, or referee report, for AI-assisted work across sessions. Interviews the author until the manifest is written, then scaffolds deliverable.yml, append-only HANDOFF.md, a numbered planning/ wiki with index and open questions, inbox/ for dictated braindumps, gitignored checks/, and the house deck template for talks. Use when the user says "open a deliverable", "set up this talk/paper/module for the pipeline", "make a manifest", or starts deadline-driven work that spans sessions. Not for one-off tasks.
referee-response
Organizes and formats an author's response to peer review — extracts every distinct point from the referee reports and the editor's letter, tags severity and type, maps the dependencies so the revision runs in the right order, flags which points the author may want to push back on, and builds the response-to-reviewers letter as a numbered comment → response → location table with the substantive answers left for the author to write. It formats and checks the response; it never writes the scientific content of an answer. Use when the user has referee or reviewer reports, a revise-and-resubmit decision, or an editor's letter and says "respond to reviewers", "plan the revision", "map the referee comments", "draft the response letter", or "check I addressed everything". The reviewer-side twin is journal-review; prose polish afterwards goes to sci-edit.
research-grill
Interviews a researcher in rounds until a research idea, design, or draft has no unexamined decision. Numbers every question, explains "why this matters", recommends an answer, fetches facts rather than asking for them, and records settled decisions. Three stages cover idea, a topic or hunch to a falsifiable question and contribution claim, design, a question to estimand, identification, sample and power, measurement, pre-registration, analysis plan, and venue, and defend, reviewer objections to a finished design or draft. Use when the user says "grill me", "grill this", "stress-test my idea/design/plan", "interview me about this project", "poke holes in this", "what am I assuming", or brings an idea that is not yet a design. Use `--plan` for software or process plans.
fable-orchestrate
Run a multi-model orchestration workflow led by Fable 5, the strongest model on the team. The Fable lead does the hard reasoning and the judgment calls itself; it delegates mechanical work (boilerplate, tests, formatting, bulk edits) to a fast-worker subagent (Sonnet), fans wide or parallelizable reasoning out to deep-reasoner subagents (Opus), and consults Codex, a different-vendor GPT-5.6 (Sol by default) peer, as a decorrelated cross-check on high-stakes, hard-to-verify calls. Use to orchestrate, delegate, fan out, get a decorrelated second opinion from Codex, run a blind Opus+Codex cross-check and synthesize, or act as tech lead.
advisor
Consult an independent, read-only GPT-6 Astra advisor before committing to a substantive interpretation, approach, or final result. Not for routine work, implementation, or file edits. Defaults to gpt-6-astra at xhigh for demanding reviews.
citation-check
Audits the citation layer of a manuscript — in-text and reference-list parity, fabricated or nonexistent sources, DOIs that resolve to a different work, style and completeness against APA 7 or a named journal style, and whether each cited work actually supports the claim attached to it. Verifies against Crossref, OpenAlex, DataCite, and Semantic Scholar, drives LaTeX audits off the keys actually cited, and marks anything it cannot check as NOT CHECKED rather than guessing. Use when the user asks to check citations or references, suspects an AI-invented source, wants a .bib checked against the text, or asks whether the DOIs are right. Figures and tables go to figure-table-audit.
conjoint-cleaning
Turns a raw Qualtrics conjoint export into an analysis-ready long-format dataset — export settings and metadata rows, identifying which Qualtrics implementation produced the file, reshaping wide to long by hand or with cjoint read.qualtrics and cjdata reshape_conjoint, mapping the choice variable onto profiles, ratings and secondary DVs, attribute-level translation and factor ordering, pilot data-quality diagnostics, and subgroup merges. Use when the user has a conjoint CSV and asks how to clean, reshape, restructure, or debug it, cannot line the choice column up with the profiles, or asks what the data must look like before AMCE estimation. Design questions go to conjoint-design, validity review to conjoint-diagnostics.
conjoint-design
Designs conjoint and factorial-vignette experiments end to end — attribute architecture and randomization restrictions, effective-N power from the closed-form AMCE standard error, treatment realism, estimand choice among AMCE, marginal means, and AMIE, design variants such as forced choice versus rating, PAP tiers for conjoint flexibility, and the regression models and R packages that implement each. Use when the user is planning a conjoint, drafting or critiquing an attribute table, asking how many respondents or tasks are needed, asking whether to report AMCEs or marginal means, or asking how to test interactions. Reviewing an existing design goes to conjoint-diagnostics, cleaning the export to conjoint-cleaning.
conjoint-diagnostics
Reviews an existing conjoint study for threats to inference and returns prioritized findings across five areas — design integrity (attributes, profile restrictions, task count and satisficing, randomization, power), estimation (estimand clarity, reference levels, subgroups, clustered standard errors, multiple testing), measurement error, external validity and behavioral benchmarking, and interpretation, including the guard-rail against reading an AMCE as a majority preference. Use when the user asks whether a conjoint design or analysis holds up, has referee comments on a conjoint, or wants a second opinion on estimation and interpretation choices. Building a design from scratch goes to conjoint-design, reshaping the data to conjoint-cleaning.
cross-national-design
Designs multi-country survey experiments — case selection with a stated comparative logic, instrument localization and TRAPD translation, origin-country and composition choices for stimuli, per-country power and error management, and a pooled-versus-per-country analytical strategy with measurement-equivalence checks. Use when the user is fielding the same experiment in several countries, asks which countries to include and why, asks how to translate or adapt an instrument without breaking comparability, asks whether to pool or estimate per country, or asks how large each country sample must be. Question wording goes to survey-design, per-country CONSORT and DA-RT reporting to methods-reporting, and the study-level plan to pre-registration-writing.
diverge-codex
Delegate creative divergence to a fresh Codex subagent before implementation. Use when the user explicitly asks Codex to delegate brainstorming, obtain an independent Codex context, or keep implementation separate from approach generation. Generate 3–5 conceptually distinct approaches, pause for selection, then assign the selected approach for implementation.
diverge
Generate 3-5 conceptually distinct approaches labeled by creativity dimension (Novel, Surprising, Diverse, Conventional) and hold for selection instead of implementing the first idea; the --codex mode delegates the brainstorm to a fresh Codex subagent for a clean, unanchored context and hands it the selected approach to implement. Use when a task has more than one non-obvious solution — creative, architectural, or analytical work — and before committing to an approach; use --codex when the user explicitly asks for a subagent, an independent context, or a delegated brainstorm.
fact-check
Fact-check manuscript claims against cited sources in a per-source Markdown knowledge base. Use to audit claim support, overclaiming, direction, scope, and misattribution after source intake is complete.
fair-check
Audit manuscript and replication package against FAIR open-science principles.
doc-to-markdown
Read or convert any document a research workflow hands you — PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, or CSV. Use whenever a document has to be read, opened, quoted, summarized, searched, extracted, or added to a source library. Decides whether to read the file directly or convert it, picks the converter from the document's actual structure, and decides whether the resulting Markdown is a tracked artifact or a scratch file to delete.
figure-table-audit
End-stage QA on a finished figure and table set — an inventory with in-text callouts and producing scripts, cross-reference and numbering checks, text-to-evidence consistency between claims and plotted or tabulated values, captions and statistical notes that stand alone, accessibility and production quality, and SI and replication-package linkage. Marks anything requiring values read off an image as needing author verification rather than inventing them. Use when the user is preparing a submission and asks to check figures, tables, captions, cross-references, or table notes, or asks whether the numbers in the text match the exhibits. Design guidance during drafting lives in figures and tables.
figures
Design and format publication-quality figures. Use for chart choice, color, scales, legends, captions, accessibility, and reproducible figure workflows.
hypothesis-building
Turns a theory into falsifiable, pre-registerable hypotheses — DAGs and backdoor closure, how the design resolves the FPCI, SATE versus PATE, counterfactual and directional framing, a named theoretical and empirical estimand, a justified SESOI, the choice among NHST, interval, equivalence (TOST), and minimum-effect tests, scope conditions, and primary, secondary, and exploratory tiers across multi-experiment designs. Use when the user has a research question or theory and asks to turn it into hypotheses, asks whether a prediction is falsifiable or testable, asks what the estimand is, or asks how to predict a null. Framing the paper around it goes to narrative-building, the plan to pre-registration-writing.
journal-review
Drafts a referee report on someone else's manuscript for a journal editor — a recommendation, a summary of the claim and design, three to six major concerns that drive the decision, additional concerns, and a coherent revision plan, produced by five parallel adversarial finders (Breaker, Butcher, Shredder, Void, Situator), a Blue Team error filter, a Chief Reviewer synthesis, and a Tone Guard legal pass, with an optional confidential note to the editor. Use when the user has been asked to referee a manuscript for a journal. Self-audit of the user's own draft goes to paper-review-lite or presubmit, and writing an author-side response to reviewers goes to referee-response.
list-experiment
Designs and diagnoses list experiments (the item count technique) — whether indirect measurement is warranted at all, control-list construction against ceiling and floor effects, design variants such as double list and direct-question pairing, difference-in-means and maximum-likelihood estimators via ictreg, the no-design-effect and no-liar assumptions with ict.test and ict.hausman.test, corrections for common failures, and power planning that accounts for the large precision penalty. Use when the user is measuring a sensitive attitude or behavior and asks about list experiments, item count, unmatched count, or veiled questioning, or has one returning a negative or implausible prevalence estimate. Deciding whether sensitivity bias exists at all goes to survey-design.
llm-calibration-logprobs
Reads a model's own uncertainty off its token log-probabilities — collecting logprobs and aggregating multi-token labels, confidence tiers and margins for triage, calibration assessment with ECE, Brier scores, and reliability diagrams, using confidence downstream without laundering it into evidence, and what to archive for reproducibility. Use when the user asks how confident a classifier was on each item, asks about logprobs, top-k tokens, calibration, ECE, Brier, or reliability diagrams, or wants low-confidence cases routed to human review. This is within-model confidence — agreement across several independent model coders goes to model-council-voting, and codebook and validation design to text-classification.
methods-reporting
Audits a methods section against a 45-item checklist synthesized from the APSA Experimental Section rubric, JARS-Quant, CONSORT, and DA-RT — pre-registration and design documentation, subjects and recruitment, randomization and treatment detail, CONSORT-style sample flow with attrition, statistical analysis with sample-size justification and three-tier results labeling, conjoint-specific reporting, the four validity types, and open-science infrastructure. Use when the user asks whether a methods section reports enough, is preparing for submission or a replication archive, asks what CONSORT, JARS, or DA-RT require, or asks how to document a deviation from the pre-analysis plan. Writing that plan beforehand goes to pre-registration-writing.
model-committee
Run one consequential, contestable decision through a deliberative committee of GPT-6 Astra and Claude Opus 5, chaired by fable (default), astra, opus, or sol. Not for factual lookups, brainstorming, routine implementation, independent-coder reliability, or final professional judgment. Use when several options are defensible and the two model families should propose, critique, revise, and cross-rank under a predeclared rubric.
model-council-voting
Runs several language models as independent coders on the same labeling or discovery task and reads their disagreement as data — panel assembly for model diversity, keeping votes independent, consensus rules, agreement statistics (Cohen's and Fleiss kappa, Krippendorff's alpha), the correlated-errors caveat that stops agreement being mistaken for validity, human validation beyond the panel, and reporting. Use when the user asks about a council, panel, ensemble, or jury of models, asks how to combine several models' labels, or asks what kappa or alpha to report for model coders. Single-model codebook and validation work goes to text-classification, per-item confidence to llm-calibration-logprobs.
paper-review-lite
Run a pre-submission manuscript audit covering argument, numerics, references, writing, figures, methods, preregistration, and replication readiness. Add --codex for a cross-model pass in which Codex and Claude review independently and verify each other's findings.
pre-registration-writing
Writes a pre-analysis plan before data collection — registry selection (OSF, AsPredicted, AEA, EGAP), PAP document structure, an analytical strategy specified down to the model and the decision rule, analysis code pre-registered against simulated data, contingency planning for attrition, failed manipulations, and exclusions, deviation documentation, and timeline. Operationalizes the pre-data-collection side of DA-RT. Use when the user asks to write or review a pre-registration or PAP, asks which registry to use, asks what to lock down versus leave exploratory, or asks how to handle a deviation later. Hypotheses and estimands come from hypothesis-building, post-hoc reporting from methods-reporting.
replication-package
Scaffold or audit a social-science replication package, and audit the manuscript and its archived research objects against the FAIR principles. Scaffold mode writes the folder structure, README, master.R, figure/table crosswalk, codebook template, LICENSE placeholder, .gitignore, and pre-release checklist. Audit mode grades an existing package against that checklist and runs the FAIR block over data, code, materials, prompts, preregistrations, DOIs, metadata, licenses, access restrictions, and availability statements. Use when setting up or repairing a replication package, checking one before submission, auditing research objects against FAIR (Findable, Accessible, Interoperable, Reusable), or drafting and verifying data-, code-, and materials-availability statements. Adapted from Yusaku Horiuchi's replication-package-guide; platform-neutral (Harvard Dataverse, OSF, Zenodo, GitHub releases, institutional archives).
research-wayfinder
Plan a research project as a durable decision map, resolving one design decision per session until the destination is a defensible, pre-registerable design. Not for executing analyses or writing the paper. Use at the start of a project, when a design has more open decisions than one conversation holds, or when planning keeps restarting.
spawn
Spawn full peer Codex or Claude sessions in their own panes and git worktrees — real sessions, not subagents — briefed by the lead and merged back. Not for bounded consults a subagent covers. Use when work must outlive or run beside this session, needs its own worktree, or must stay user-steerable.
survey-design
Designs survey instruments — item-specific question wording that avoids acquiescence, double-barreling, and leading language, scale construction (points, labeling, polarity, feeling thermometers, index reliability with Cronbach's alpha and McDonald's omega), flow and ordering with treatment placement and buffer items, pretesting through cognitive interviews and soft launches, respondent burden with attention checks and speeding rules, sensitive questions and social-desirability bias, and treatment delivery with comprehension gates. Use when the user is writing or revising survey questions, choosing a scale, ordering blocks, planning a pilot, or worried respondents will not answer honestly. Indirect measurement goes to list-experiment, multi-country instruments to cross-national-design, and operating a live survey to qualtrics-ops.
tables
Design and format publication-quality tables. Use for column order, row grouping, notes, statistical precision, accessibility, and reproducibility.
text-classification
Designs and validates LLM-based text classification for research data — codebook construction, choice of learning regime, model selection and reproducibility, prompt construction, pilot validation against human coding with agreement statistics (kappa, F1), hybrid human-LLM workflows, and reporting model-coded data. Also carries a resumable batch pipeline with a rule-based baseline for coding large sets of repeated free-text values against a closed codebook — occupation, institution, and registry text, and open-text survey responses — with a frozen input set, incremental output, residual buckets, stratified hand validation, a regex or dictionary comparison, and a published lookup table. Use when the user asks to classify, code, or label text at scale. Discovering categories rather than applying them goes to topic-modeling.
topic-modeling
Specifies and diagnoses structural topic models for survey and experimental text — choosing among STM, LDA, and BERTopic, preprocessing decisions and their consequences, prevalence and content formulas with spectral initialization and a recorded seed, selecting the topic count across semantic coherence, exclusivity and FREX, held-out likelihood, and residuals rather than one metric, interpretation and validation against representative documents, robustness checks, and DA-RT-compliant reporting. Use when the user has open-ended responses or another corpus and asks what topics are in it, asks how many topics to use, asks about STM, searchK, coherence, or FREX, or wants topic prevalence compared across treatment arms or countries. Coding against a fixed codebook goes to text-classification.
qualtrics-ops
Operate or audit a live Qualtrics survey via the v3 APIs without breaking fielding — publish gating, quotas, flow routing, embedded data, panel-vendor redirects, read-back verification, and a read-only pre-fielding audit. Use when publishing or patching a fielding instrument, when a quota counts but never blocks, when wiring panel-vendor redirects or flow gates, or when auditing a survey before launch (consent-before-anything gates, force-response completeness, quota and redirect checks, anti-bot instrumentation, and language-arm symmetry).
finished
End-of-session closure for a research or code repository. Records what actually changed into the project's own handoff and log files, distinguishing verified work from work merely attempted, and reports what was left uncommitted. Use when wrapping up a work session so the next one can resume without re-deriving state. Its start-of-session twin is sitrep.
orchestrate
Orchestrate complex work from an active gpt-6-astra or gpt-5.6-sol Codex session. The detected lead owns decomposition, integration, and verification; Astra keeps the hardest reasoning in the lead, while Sol escalates unusually difficult units to Astra. Both route bounded work to cheaper GPT-5.6 tiers and can use a cross-vendor Claude peer. Not for routine single-context work. Use when the user explicitly asks to orchestrate, delegate, fan out, parallelize, assign subagents, obtain independent checks, or have Codex act as tech lead.
sitrep
Start-of-session situational report for a research or code repository. Reads whatever handoff, log, or status artifacts the project actually keeps, checks live git state, and reports the current objective, verified baseline, open tasks, blockers, and next actions. Use when resuming work in a repository after time away, or when you need to know where a project stands before changing anything. Its end-of-session twin is finished.
vlm-ocr
OCR scanned or image-only corpora with vision-language models, in three phases. Use `evaluate` to compare candidate OCR systems against stratified human ground truth and pick one on measured CER/WER, `run` to build the production pipeline (model selection, image handling, prompts, architecture, batching, accuracy evaluation, reproducibility), and `clean` to correct raw OCR text with LLM and rule-based passes, quality diagnostics, multilingual handling, and span-level provenance. Not for born-digital documents with a text layer — those go to $doc-to-markdown.
46-orchestrate
Orchestrate complex work as the Codex lead, with a gpt-5.6-sol lead at xhigh effort owning decomposition, integration, and verification while delegating bounded work to cheaper GPT-5.6 tiers and a cross-vendor Claude peer. Not for routine single-context work. Use when the user explicitly asks to orchestrate, delegate, fan out, parallelize, assign subagents, obtain independent checks, or have Codex act as tech lead.
literature-review
Build or audit a literature review. Use for evidence maps, gap analysis, contribution checks, source verification, and synthesis planning.
model-committee-fable
Run a deliberative two-model committee between GPT-5.6 "Sol" and Claude Opus 5, chaired by Fable 5. Same two deliberating members as model-committee; the difference is the chair — Fable 5 aggregates the scores, applies the tie rule, and synthesizes the decision, so a lightweight Claude-family chair handles the tally and synthesis. Use when the user needs one consequential decision from multiple defensible options and wants a Fable-chaired deliberation. Suitable for architecture, research design and interpretation, manuscript strategy, ambiguous diagnosis, evaluation design, and policy or standards tradeoffs. Not for factual lookups, independent-coder reliability, open-ended brainstorming, routine implementation, or final high-stakes professional judgment.
model-committee-sol
Run a deliberative two-model committee between GPT-5.6 "Terra" and Claude Opus 5, chaired by GPT-5.6 "Sol." Same two deliberating members as model-committee; the difference is the chair — Sol aggregates the scores, applies the tie rule, and synthesizes the decision. Use when the user needs one consequential decision from multiple defensible options and wants a Sol-chaired deliberation. Suitable for architecture, research design and interpretation, manuscript strategy, ambiguous diagnosis, evaluation design, and policy or standards tradeoffs. Not for factual lookups, independent-coder reliability, open-ended brainstorming, routine implementation, or final high-stakes professional judgment.
narrative-building
Draft or audit scientific introductions. Use for argument logic, framing, contribution structure, and coherence across multiple studies or experiments.
paper-review-lite-codex
Run a cross-model adversarial pre-submission audit from Codex. Use when a user explicitly requests the heavier paper-review-lite-codex workflow, independent Codex and Claude review passes, or cross-model verification of a manuscript. Apply the paper-review-lite protocol, cross-check findings against the manuscript, and retain only evidence-grounded issues.
paper-tex
Typeset a working paper or journal submission in house-style LaTeX from Markdown, Word, TeX, ODT, RTF, or HTML. Use to convert drafts, build PDFs, or prepare journal-specific spacing, limits, anonymization, disclosures, and citations.
post-ocr-cleanup
Clean post-OCR research text. Use for correction, quality assurance, multilingual handling, uncertainty preservation, and provenance-aware cleanup.
research-repo
Scaffold or audit an entire research project repository organized around its source library. Use when starting, structuring, organizing, or reviewing a research repo. Build sources/{og,md,unprocessed}, references.bib, a PDF-to-Markdown converter, a repo-local process-source skill, AGENTS.md, .gitignore, .venv, and the relevant analysis/manuscript/review folders; or audit an existing layout. Not for one-PDF intake or publication replication packages.
vlm-ocr-evaluation
Compare OCR systems before a bulk run: candidate set, stratified ground truth, CER/WER, normalization, per-language and per-stratum accuracy.
vlm-ocr-pipeline
Design VLM-based OCR pipelines. Use for model selection, prompting, architecture, multilingual documents, structured output, and evaluation planning.
survey-data-audit
Audit fielded survey response data for registered elements, data quality, bot and AI-automation screening, and sample integrity. Emits an appendix-ready quality report.
survey-flow-audit
Pre-fielding audit of a live survey over the platform API, with an optional browser walk. Consent-before-anything gates, publish state, force-response completeness, quotas, vendor redirects, anti-bot instrumentation, language-arm symmetry.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.