Steven Denney

Runs the GPU part of a vlm-ocr pipeline on Leiden University's ALICE cluster. Decides whether a corpus needs ALICE, picks a GPU partition for the model, sets up storage, a pinned vLLM environment and offline weights, ships code and scans, runs a gate job and chained bulk jobs with the bundled SLURM template, and brings the page records home for post-processing. Use with vlm-ocr when OCRing scans on ALICE or deciding whether to.

The English summary is being prepared. Showing the original description.

Run the GPU part of a $vlm-ocr pipeline on Leiden University's ALICE cluster: decide whether a corpus needs ALICE, pick a GPU partition for the model, set up storage, a pinned vLLM environment and offline weights, ship code and scans, run a gate job and chained bulk jobs with the bundled SLURM template, and bring the page records home for post-processing. Use with $vlm-ocr when OCRing scans on ALICE or deciding whether to.

The English summary is being prepared. Showing the original description.

Turns raw captures (dictated braindumps in a deliverable's inbox/, or recent macwhspr dictation-log entries) into proposed planning-wiki changes, shown as one unified diff. Classifies each statement as a decision, evidence claim, question, task, or dropped item, traces it to its capture paragraph, tags unconfirmed names, numbers, and citations UNRESOLVED, and writes decisions only with the author's assent. Edits only the planning wiki, never source files. Use when asked to intake, absorb notes, or process the inbox, or after dictating notes from a talk, meeting, or reading session.

The English summary is being prepared. Showing the original description.

Interviews a researcher in rounds until an idea, design, or draft has no unexamined decision — numbered questions, each with why it matters and a recommended answer, facts looked up rather than asked, and settled decisions written to a file. Stages are idea (topic or hunch to a falsifiable question and contribution claim), design (estimand, identification, sample and power, measurement, analysis plan, pre-registration, venue), and defend (reviewer objections to a finished design or draft). Use when the user says "grill me", asks to stress-test or poke holes in an idea, design, or plan, asks what they are assuming, or brings an idea that is not yet a design. --plan grills software or process plans.

The English summary is being prepared. Showing the original description.

End-of-session closure for a research or code repository. Records what changed into the project's own handoff and log files, separating verified work from work merely attempted, and reports what is left uncommitted. Use when wrapping up a session so the next one can resume without re-deriving state. Its start-of-session twin is sitrep.

The English summary is being prepared. Showing the original description.

Reviews a whole deliverable (talk, course module, paper, chapter, or referee report) against its manifest without rewriting it. Runs a deterministic gate for build, citations, numbers, facts, and leaks and stops on failure, then runs a finder over each section and a strong-model pass over the whole work, and adjudicates into checks/<date>/findings.json and report.md with anchored findings, severities, and honest NOT CHECKED coverage within a dollar budget. Use when asked to lint or review the whole talk, module, or paper, for a pre-ship review, or for the night shift. Revision stays serial and author-led.

The English summary is being prepared. Showing the original description.

Start-of-session situational report for a research or code repository. Reads the handoff, log, or status files the project actually keeps, checks live git state, and reports the objective, verified baseline, open tasks, blockers, and next actions, flagging where the documents and the repository disagree. Use when resuming work after time away or before changing anything. Its end-of-session twin is finished.

The English summary is being prepared. Showing the original description.

Opens a deliverable (talk, course module, paper, book chapter, or referee report) for AI-assisted work across sessions. Interviews the author for the manifest, then scaffolds deliverable.yml, an append-only HANDOFF.md, a numbered planning/ wiki with open questions, an inbox/ for dictated braindumps, a gitignored checks/, and the house deck template for talks. Use when starting deadline-driven work that spans sessions, or when asked to open a deliverable, set up a talk, paper, or module for the pipeline, or make a manifest. Not for one-off tasks.

The English summary is being prepared. Showing the original description.

Spawns full Claude Code peer sessions in separate terminal panes and git worktrees, real sessions rather than subagents, each with a directed task and contract brief, monitored and merged by the spawning lead. Detects the environment and uses herdr first, then tmux, then a native claude background agent. Use when work must persist or run beside the current session, needs a separate worktree or permission settings, should remain user-steerable in a visible pane, or when the user asks to spawn, hand off, or parallelize full sessions. Also spawns Codex peers into those panes. Excludes bounded consults and subagent-scale work.

The English summary is being prepared. Showing the original description.

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.

The English summary is being prepared. Showing the original description.

Runs several language models as independent coders on one labeling or discovery task and treats disagreement as data. Covers diverse panel assembly, independent voting, consensus rules, Cohen’s and Fleiss kappa, Krippendorff’s alpha, correlated-error limits on validity claims, human validation, and reporting. Use when the user asks about a council, panel, ensemble, or jury of models, how to combine model labels, or 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.

The English summary is being prepared. Showing the original description.

House rules and craft for LaTeX Beamer slide decks (lectures, seminars, talks), drawn from the author's own Overleaf edits and published slide-design guidance. Covers storyboarding, frame titles, slide prose, word and frame budgets, layout without overfull boxes, XeLaTeX fonts and Korean (xeCJK), color, figures and screenshots, Overleaf source hygiene, speaker notes, and a build-and-QA loop with a contact sheet. Use when writing, editing, reviewing, or building a Beamer deck, theme, or storyboard. A course skill supplies course facts and the frame skeleton; this skill supplies the craft. Not for HTML or PowerPoint decks, posters, or papers.

The English summary is being prepared. Showing the original description.

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.

The English summary is being prepared. Showing the original description.

Advises on setting up and validating topic models for survey and experimental text. Covers the discovery-versus-measurement question, choosing STM, LDA, or BERTopic, preprocessing consequences, prevalence and content formulas, topic-count selection across coherence, exclusivity and FREX, held-out likelihood, and residuals, validation of topics against human judgment, robustness, and DA-RT-compliant reporting. Use when the user has open-ended responses or another corpus and asks what topics are present, how many topics to use, about STM, searchK, coherence, or FREX, or wants topic prevalence compared across treatment arms or countries. Fixed-codebook coding goes to text-classification.

The English summary is being prepared. Showing the original description.

Audits fielded survey-response data for registered elements, data quality, bot and AI-automation screening, and sample integrity, and produces an appendix-ready quality report. Use at soft launch before full field, after fielding before outcomes are analyzed, or when asked what a study did about bots and AI respondents.

The English summary is being prepared. Showing the original description.

Plans a research project as a decision map that persists across sessions, resolving one ticket per session until the design is defensible and pre-registerable. The map lives as Markdown in the repository, literature sweeps run as parallel research tickets, and the resolved map compiles into a pre-analysis plan. Use at project start, when a design has more open decisions than one conversation can hold, or when planning restarts each session. Not for executing analyses or writing the paper.

The English summary is being prepared. Showing the original description.

Designs publication tables for social-science manuscripts. Covers whether a result belongs in a table at all, column order, row grouping and panels, decimal precision, standard-error and confidence-interval conventions, self-contained titles and notes, and code-generated output. Use when building or revising a regression, balance, descriptive, or robustness table, or when working with modelsummary, stargazer, gt, esttab, or booktabs. Finished tables go to figure-table-audit for end-stage QA.

The English summary is being prepared. Showing the original description.

Advises on setting up LLM-based text classification as a validated measurement design. Covers codebook construction, learning-regime and model choice, prompt construction, validation against human coding with kappa and F1, hybrid human-LLM review, downstream measurement error, and reporting, plus a short resumable batch pattern with a rule-based baseline for repeated free text against a closed codebook. Use when the user wants to classify, code, or label text at scale. Category discovery goes to topic-modeling.

The English summary is being prepared. Showing the original description.

Audits a methods section against a 45-item reporting checklist synthesized from the APSA Experimental Section rubric, JARS-Quant, CONSORT, and DA-RT, covering pre-registration, recruitment, randomization and treatments, sample flow and attrition, sample-size justification, three-tier results labeling, conjoint reporting, validity, and open-science infrastructure. Use when checking whether a methods section reports enough, preparing a submission or replication archive, asking what CONSORT, JARS, or DA-RT require, or documenting a deviation from a pre-analysis plan. Writing that plan before data collection goes to pre-registration-writing.

The English summary is being prepared. Showing the original description.

Runs a multi-model orchestration workflow led by the session's Fable or Claude Opus model. Sends menial tool-calling work to Haiku subagents, judgment work to Opus subagents at an effort matched to each task, and high-stakes or fresh-perspective questions to a GPT-6 Astra Codex peer. Detects the lead from session context; `--lead fable` or `--lead opus` overrides. Use to orchestrate, delegate, fan out, get a decorrelated Codex opinion, run a blind Opus-Codex cross-check, or act as tech lead.

The English summary is being prepared. Showing the original description.

Reads model uncertainty from token log-probabilities. Covers collecting logprobs and aggregating multi-token labels, confidence tiers and margins for triage, calibration assessment with ECE, Brier scores, and reliability diagrams against human labels, responsible downstream use of confidence, and reproducibility archives. Use when the user asks about classifier confidence by item, logprobs, top-k tokens, calibration, ECE, Brier, reliability diagrams, or routing low-confidence cases to human review. Codebook and validation design go to text-classification.

The English summary is being prepared. Showing the original description.

Runs a deliberative two-model committee of GPT-6 Astra and Claude Opus under a selectable chair, Fable by default, with `/model-committee-astra`, `/model-committee-opus`, and `/model-committee-sol` selecting the others. Use when a consequential decision has several defensible options and two model families should propose, critique, revise, and cross-rank them against a predeclared rubric before converging, as in research design, interpretation, architecture, manuscript strategy, or evaluation design. Not for factual lookups, independent-coder reliability, brainstorming, routine implementation, or final professional judgment.

The English summary is being prepared. Showing the original description.

Scaffolds or audits a social-science replication package and evaluates the manuscript and archived research objects against FAIR principles. Scaffold mode writes the folder structure, README, master.R, figure-table crosswalk, codebook, LICENSE placeholder, and .gitignore; audit mode grades the package against a pre-release checklist and audits data, code, materials, identifiers, licenses, access restrictions, and availability statements. Use when setting up or repairing a package, checking one before submission, auditing FAIR compliance, or drafting availability statements. Adapted from Yusaku Horiuchi’s replication-package-guide.

The English summary is being prepared. Showing the original description.

Provides a pre-submission manuscript audit covering argument, numerical consistency, references and DOIs, writing, figures, CONSORT flow, pre-registration, and the replication archive. Use before paper submission or when a draft needs an adversarial, quote-grounded review with a journal-readiness checklist. Add `--codex` to run the specification independently on Claude and Codex, GPT-6 Astra, and cross-check findings.

The English summary is being prepared. Showing the original description.