workflows

ALWAYS load before finalizing ANY written prose, and whenever text is suspected of being machine-written - "does this sound like AI", "make this sound human", "clean up the AI-isms", "did a bot write this", "check this draft before I send it", "review this for AI tells", "this reads like ChatGPT", "review this for mannered prose", "is this too writerly", "is this student paper AI-generated", or before handing back any draft, memo, email, or article you wrote. Use proactively even if the user never mentions AI writing.

ALWAYS use for ANY legal citation question, even if the user never says 'Bluebook' - 'cite this case', 'how do I cite a statute', 'is this footnote formatted right', 'id. or supra here', 'cite a law review article', 'what signal goes here', 'format these footnotes', 'short form for this cite', 'block quote this', 'where does the ellipsis go', 'is this quote altered right', or writing or checking any citation or quotation in a legal manuscript. NOT for auditing a whole manuscript's footnotes (use bluebook-audit) and NOT for rendering these rules in Typst (use docx-typst).

This skill should be used when the user asks to 'audit footnotes', 'check Bluebook formatting', 'audit citations', 'run footnote audit', 'check my footnotes', 'bluebook audit', or needs systematic Bluebook compliance checking of a law review manuscript.

ALWAYS use when a draft's citations need checking against what the sources actually say — 'check citations', 'check my cites', 'verify cites', 'cite-check', 'run citation review', 'are my citations grounded', 'does this source actually support that claim', 'does source X support claim Y', 'what does source X say about Y', 'did I cite that right', 'make sure the cites hold up before I send this'. Use proactively before a draft goes out. NOT for Bluebook formatting (use bluebook / bluebook-audit) and NOT for detecting citations that do not exist at all (use source-verify).

ALWAYS use before running the `consensus` CLI or hitting consensus.app - "search Consensus", "consensus search", "find RCT papers", "randomized trials on X", "systematic reviews / meta-analyses of X", "clinical papers on X", "empirical papers in top journals only", "papers in these specific journals", "filter by journal quartile", "most-cited papers on X since 2018". Use even when the user names a filter (study type, journal, year range, citations) without saying "Consensus". NOT for a broad multi-source literature sweep - use the research skill, which calls this one.

This skill should be used when the user asks to 'extract patterns from this session', 'save what we learned', 'create a skill from this workflow', 'learn from this conversation', 'capture reusable knowledge', or wants to turn session patterns into reusable skills.

This skill should be used when the user asks to "get the complaint PDFs", "harvest court documents", "pull the docket", "download the opinions", "get filings from Lex Machina", "Docket Alarm", "Bloomberg Law docket", "Delaware Chancery filings", "who filed this case", or needs court filings, docket entries or written decisions out of a commercial legal database.

Use when the user says 'CourtListener', 'Free Law Project', 'RECAP', 'PACER data', 'court opinions', 'get the opinion text', 'federal dockets', 'delch', 'Delaware Chancery opinions', 'state court opinions', 'demand futility opinions', 'judicial financial disclosures', 'oral arguments', or needs case law, dockets or opinion full text for empirical legal research. Also use when someone assumes CourtListener is federal-only, or hits a token or human-verification wall on it.

Use when "CRSP is stale / out of date", "CRSP only goes through December", "my panel ends in December", "why do my returns stop", "extend CRSP to today", "fill forward CRSP with LSEG", "current stock prices for my CRSP panel", "up-to-date returns for permnos", "splice CRSP and LSEG", "CUSIP to RIC for a CRSP universe", "backfill the CRSP gap", or any request to carry a CRSP daily/monthly stock series past CRSP's last data date using LSEG/Refinitiv. Use proactively whenever a CRSP panel has to reach a date after CRSP's annual cutoff, even if the user never mentions LSEG — CRSP returns a short panel without erroring, and a naive concat fabricates the seam return. NEGATIVE ROUTING: a CRSP query that stays inside CRSP's coverage, or CIZ table and column names, is crsp-v2; LSEG session, quota, entitlement or general symbology questions are lseg-data; raw WRDS PostgreSQL access and connection patterns are wrds — this skill covers only the join between CRSP and LSEG.

Use when "CRSP CIZ", "CRSP v2", "CRSP flat file format 2.0", "crsp.dsf_v2 / msf_v2", "StkDlySecurityData", "StkMthSecurityData", "StkSecurityInfoHist", "stocknames_v2", "DlyRet / MthRet / DlyPrc / MthPrc", "SHRCD or EXCHCD equivalent in new CRSP", "SIZ to CIZ migration", "CRSP data after 2024", "CRSP delisting returns", "CRSP cumulative adjustment factors", "CRSP index INDNO / INDFAM", or any CRSP stock/index query where the legacy SIZ column names no longer exist.

This skill should be used when the user asks to 'set up data analysis for our database', 'extract tribal knowledge about dataset', 'generate data skill', 'document this dataset', 'what does this column mean', 'create data dictionary', 'help me understand this data schema', 'capture domain knowledge about our data', or needs to create a reusable data context skill from dataset expertise.

ALWAYS use when prose needs to stop sounding machine-written — 'de-AI this', 'make it sound less like AI', 'this reads like ChatGPT wrote it', 'remove the AI-isms', 'de-tic this draft', 'humanize the prose', 'fix the AI writing tells', 'less AI-sounding', 'this doesn't sound like me', 'too many em-dashes and tricolons', 'it reads robotic', 'just flag the AI tells in this'. Use as the standard AI-prose pass before any draft ships, even if the user only says 'clean up the writing'. NEGATIVE ROUTING: this is the skill that EDITS the draft, and it runs detect-only on the same scorers when asked to scan rather than fix; reading the tic tables or asking what counts as an AI tell is ai-anti-patterns; validating a candidate phrase against the human corpus or adding a linter rule is ai-tic; grading a whole draft against the domain register and prose-quality rules is the writing-reviewer agent; judging whether text WAS AI-written is nobody's job here — this renders no verdict on authorship.

ALWAYS load before attempting a broad web-grounded research report - "deep research this", "do a deep dive on X", "comprehensive research on X", "thorough investigation of X", "write me a research report on X", "everything you can find about X", "I need background on this industry/market/regulation", "the academic sources came up empty, look wider". Use even when the user never says "deep research" but wants a synthesized multi-source report. NOT for finding academic papers or citations - use the research, consensus and google-scholar skills first; this is the paid last resort.

Use when the user says "build this feature", "implement X", "add support for", "add a flag for", "fix this bug properly", "this is broken, fix it", "refactor X", "write the tests for this", "/dev", or hands over any code change that should ship with a real failing-then-passing test rather than a quick edit. Use proactively the moment a request implies changing code — before reading the codebase to understand the bug, since that reading is part of the run. NEGATIVE ROUTING: a change whose acceptance cannot be one command that fails now and passes later is /work; a dataset, table, figure or number is /ds even when code produces it; a skill, workflow, hook or plugin manifest in this repo is skill-creator, workflow-creator or plugin-creator; a typo or one-line fix is done inline.

Use when "query Dewey Data", "deweydata.io", "SafeGraph places/patterns/spend", "Advan foot traffic", "POI / points of interest", "mobility data", "dataplor", "Veraset", "PassBy", "crypto/Bitcoin ATM locations", "GovFiles", "business entity / Secretary of State registry data", "OpenCorporates alternative", "LinkUp job postings", "job-postings panel", "Exchange Data International / EDI", "WCA / RCAN corporate actions", "global corporate actions (dividends, splits, mergers, tenders)", or any pull from the Dewey Data academic marketplace (UVA/NYU Platform Subscription) via the deweypy/deweydatapy client, DuckDB, or the Dewey MCP server.

Use when an EXISTING .docx, .doc, .pptx or .xlsx has to become a PDF or PNG — 'convert docx to pdf', 'docx to pdf', 'render this Word doc', 'word to pdf', 'export docx as pdf', 'make a pdf of this docx', 'pdf from the docx', 'render the document to PDF', 'the PDF has the wrong page count', 'why did the layout reflow'. Use proactively before handing any Word document to a human as a PDF, and INSTEAD OF hand-rolling soffice/libreoffice, even when no one says 'render'. NEGATIVE ROUTING: NOT for editing docx content (use the generic 'docx' skill), NOT for building a docx from markdown (use 'law-review-docx' for Bluebook footnotes or 'law-econ-docx' for author-date), NOT for building one from a Typst source (use 'docx-typst'), NOT for a damaged docx (use 'docx-repair'), NOT for inspecting or diffing individual slides (use 'pptx-render').

ALWAYS use for ANY substantive empirical task whose output is a dataset, table, figure or number - "analyze this data", "build the panel", "merge these datasets", "run the regression", "profile this dataset", "clean this file up for analysis", "make me a summary table", "pull the data and check X", "replicate this paper's table 2", "build the ETL pipeline", "/ds". Use proactively even when the user asks casually and never says "analysis". NEGATIVE ROUTING: grading empirical work that already exists is the ds-reviewer agent; writing the results up as prose is writing-econ; spreadsheet file mechanics - reading, writing or reformatting an .xlsx - is xlsx, while building, merging or modelling the dataset inside it is ds; a one-line lookup against a file already in hand is answered inline, since this skill runs a full clarify/plan/dispatch/human-review lifecycle.

Use when a court opinion that is not in the casebook has to become a casebook-style excerpt for students — "add this case to the addendum", "elide this opinion for class", "cut this case down to four pages the students can read", "make a casebook excerpt of X", "this case isn't in Choi, put it in the reader", "excerpt the Howey analysis from this opinion", "the TM mentions a case we don't have", "I want to teach this new decision next week". NEGATIVE ROUTING: compiling an addendum whose readings are already written is teaching:generate-addendum; retrieving an opinion with no elision is workflows:courtlistener.

Run ALL delegated agent work through the CLIProxyAPI wrappers. Use INSTEAD OF the Agent tool, subagents, the Workflow tool, and in-session agent teams — for any "delegate this", "run these in parallel", "fan out", "have an agent review/investigate/search", "use a subagent", "spawn a team", "get a second opinion", "run this workflow script", or any task you were about to hand to a background agent. Also use for explicit mentions of cliproxy, codex-code, gemini-code, claude-code, farm out, or delegating work to another model. NEGATIVE ROUTING: anything phrased as a new, background, separate or companion SESSION — 'spawn an agent', 'spawn a background claude', 'kick off claude in <dir>' — belongs to agent-spawn, which fires first and carries this work inside its prompt; messaging a session that already exists is agent-msg; designing, repairing or auditing a workflow is workflow-creator; and a `work`, dev or ds dispatch goes through work-dispatch.sh, never a hand-written farm.sh --workflow line.

Use when linking or deduping datasets by entity NAME because no shared key exists — 'fuzzy match', 'fuzzy name matching', 'entity resolution', 'record linkage', 'match company or person names', 'match these on the firm name', 'just merge these on company name', 'dedupe entity names', 'name-based join', 'these two files have no common id', or any use of char n-gram TF-IDF, cosine similarity on names, `sparse_dot_topn`, or RapidFuzz at scale. Use proactively whenever a join is about to be keyed on a name string, even if the user never says 'fuzzy'. NEGATIVE ROUTING: identifier bridging that a real link table already covers (gvkey ↔ permno via CCM, cik ↔ gvkey via wciklink) is the wrds skill — come here only for identifiers that genuinely do not cross vendors; the surrounding merge, sample-coverage and pipeline work is ds, not this skill.

Use when the user says 'run this prompt over all the documents', 'process thousands of PDFs', 'extract fields from every filing', 'bulk LLM job', 'submit a batch job', 'use the Gemini Batch API', 'upload files to Gemini', or 'this is too many to do one at a time' - any large-scale LLM extraction or classification over many files. ALWAYS load before writing Gemini batch code, including the small test run.

This skill should be used when the user asks to "search Google Scholar", "find academic papers", "scholar search", "lookup papers", "find citations", "academic search", "search for papers by author", "find journal articles", "get BibTeX", "cite this paper", "download paper", or needs to search Google Scholar for academic literature via the scholar CLI tool.

Use when work has to keep going with NO session alive — "run this in a loop overnight", "grind loop", "start the loop and go to bed", "keep iterating until the check passes without me", "a fresh agent every iteration", "loop on this for days", "the orchestrator session is burning tokens on status reports", "wake it only when the grid job lands", "what has the loop done so far", "stop the loop", "it keeps re-diagnosing the same dead item". Use proactively when a hill-climb needs many cheap passes over hours or days and each pass costs more to supervise than to run. NEGATIVE ROUTING: a stopping condition for a session that stays ALIVE — a Stop-hook hold, a cron heartbeat, an overnight brief — is the work hold; one delegated piece of work handed to another model and waited on is farm-out; starting a persistent interactive session is agent-spawn. This skill owns the out-of-session loop, the append-only journal that is its whole memory, and the shell gate that decides when a model call is worth spending.

Use this skill when the user asks to add news headline cards, 'Last Week Tonight'-style cards, headline slides, news quote slides, or media quote cards to a Typst presentation. Also use when the user wants to modify existing headline cards (add/remove cards, change quotes, swap logos, fix layout issues). Trigger on: 'add a headline', 'news card', 'LWT card', 'headline slide', 'quote card', 'media quote', 'add a quote from [publication]'.

Use when submitting jobs to UVA HPC (Rivanna/Afton), writing Slurm scripts (sbatch/srun/squeue), converting SGE to Slurm, running compute on any Slurm-managed cluster, or building WRDS data pipelines with polars on HPC. Triggers: 'submit to HPC', 'sbatch', 'squeue', 'slurm job', 'run on Rivanna', 'run on Afton', 'HPC array job', 'convert SGE to Slurm', 'polars on HPC', 'WRDS from HPC'.

ALWAYS use before touching a notebook as text or running one headlessly - "convert this notebook to a .py", "turn this script into a notebook", "get this ipynb into git", "the notebook diffs are unreadable", "keep the .py and .ipynb in sync", "run this notebook end to end", "execute the notebook with these parameters", "papermill", "jupytext", "add an R/Stata/SAS kernel", "share data between kernels". Use even when the user only says "just run the notebook" - never claim a notebook ran without executing and inspecting it.

Use this skill to BUILD a Word manuscript from LAW AND ECONOMICS markdown — author-date citations plus a reference list, double spaced throughout, Latin Modern typography — for the Journal of Law and Economics, Journal of Legal Studies, JLEO, ALER, or an econ-flavored job market paper. Triggers: 'build the L&E paper', 'law and economics Word template', 'JLE submission docx', 'Journal of Legal Studies format', 'JLEO manuscript', 'author-date Word document', 'econ job market paper in Word', 'Chicago author-date docx', 'make the submission docx for JLE'. NEGATIVE ROUTING: NOT 'law-review-docx' (Bluebook footnotes, TOC and small caps — different citation model), NOT 'docx-typst' (source is Typst, not markdown), NOT the generic 'docx' skill (edits docx content), NOT 'docx-render' (renders an existing docx to PDF/PNG only).

Use this skill to BUILD a formatted Word document from LAW REVIEW / legal MARKDOWN drafts — Bluebook footnotes, TOC, small caps, styled tables and vector figures. Triggers: 'build the law review document', 'make the Word version of my article', 'turn my markdown draft into Word', 'export the article to docx', 'compile/finalize the draft', 'make the submission docx for the law review', 'apply the law review template', 'the figure is blank in Word', 'fix the widows in the compiled PDF'. Use proactively once a legal markdown draft is finished and a Word file is implied but unnamed. NEGATIVE ROUTING: NOT 'law-econ-docx' (author-date citations plus a reference list, for JLE/JLS/JLEO/ALER — different citation model), NOT 'docx-typst' (source is Typst, not markdown), NOT the generic 'docx' skill (edits docx content), NOT 'docx-render' (renders an existing docx to PDF/PNG only), NOT 'docx-repair' (fixes a cloud-editor-damaged docx).

Use when the user says 'get this case from Lexis', 'pull it off Lexis+', 'export the opinion from Nexis', 'use Lexis instead of Westlaw', or names Lexis/Lexis+/Nexis/Protege as the retrieval source. NOT IMPLEMENTED — this skill exists only to route that request to workflows:westlaw, which is implemented. It retrieves nothing itself.

Use when the user asks to 'look at', 'analyze', 'describe', 'extract from', or 'what's in' media files like PDFs, images, diagrams, screenshots, or charts. Triggers include: 'what does this image show', 'extract the table from this PDF', 'describe this diagram', 'what's in this screenshot', 'analyze this chart', 'read this image', 'get text from this PDF', 'summarize this document', or requests for specific data extraction from visual or document files. Use for interpreted content, not literal file reading (Read tool).

Use when "query LSEG/Refinitiv", "fundamentals or market data from LSEG", "ESG scores", "management guidance / guidance reports / GR", "RIC/ISIN symbology", "corporate governance or activism (poison pills, campaigns)", "M&A or IPO deals", "syndicated loans or project finance", "PE/VC investments", "joint ventures", "municipal bonds", "Lipper fund details", "stock screening (fscreen)", "Refinitiv news", "Workspace web client", "Codebook", or any use of the `lseg.data` Python API. (academic loan/PE data: prefer the wrds skill.)

ALWAYS load before editing ANY .py file that contains @app.cell or marimo.App — 'edit this notebook', 'add a cell', 'fix the notebook', 'why is this cell not updating', 'my notebook won't run', 'convert this ipynb to marimo', 'turn my Jupyter notebook into marimo', 'start a notebook', 'export the notebook to HTML', 'run marimo', 'the cell says variable already defined', 'marimo edit'. Use even when the user just says 'the notebook' — hand-editing a marimo file breaks its cell signatures and DAG.

ALWAYS use for ANY NotebookLM work - "make me a notebook on X", "create a notebook", "add this PDF/source to the notebook", "what notebooks do I have", "make a podcast/audio overview of these", "generate a video overview", "study guide", "briefing doc", "FAQ from these sources", "outline from the notebook", "slide deck from these sources", "ask my notebook about X", "summarize the sources in there", or any mention of NotebookLM or `nlm`. Use even when the user says "make a podcast of these papers" without naming NotebookLM. NOT for open-web research reports - use deep-research instead.

ALWAYS use when an executed notebook has to be checked or has gone wrong - "debug this notebook", "why did the notebook fail", "which cell errored", "read the traceback from the ipynb", "the notebook ran but the output looks wrong", "did all the cells run", "papermill/jupytext/marimo execution failed", "inspect the notebook outputs". Use proactively before claiming any .ipynb executed successfully — never assert a notebook works without checking its outputs for tracebacks.

ALWAYS use before building anything that joins proxy votes to ownership — 'build the N-PX panel', 'proxy voting panel', 'how did index funds vote on this item', 'passive vs active voting', 'fund-level votes with ownership attached', 'voting by block', 'link ISS funds to CRSP', 'risk.voteanalysis_npx', 'N-PX votes by meeting/agenda item', 'institutional ownership per proposal', '13F or S12 holdings joined to votes', 'build the ownership panel'. Use proactively instead of writing a local pipeline — the panel already exists end-to-end on the WRDS grid. NEGATIVE ROUTING: querying the underlying ISS, Thomson or 13F tables directly, WRDS connection patterns, and the data-defect reference are the wrds skill; generic data building, merging and profiling with no proxy-vote leg is ds — this skill owns the assembled panel and its DAG, not raw table access.

Use when the user says 'organize my notes', 'clean up the vault', 'tidy the vault', 'where should this note go', 'file this note', 'move these notes', 'rename this note', or 'the vault root is a mess' - and ALWAYS before creating a new note anywhere in the Obsidian vault, so it lands in the right PARA folder under the right name. NOT for searching notes or appending to the daily note - use the obsidian skill.

Use when the user asks to "add paper", "paperpile add", "fetch PDF for", "find and add", "search paperpile", "find in paperpile", "paperpile search", "label paper", "trash paper", "download paper", "paperpile index", "edit paper metadata", "update paper title", "fix paper author", "paperpile edit", "find PDF online", "search google for PDF", "resolve PDF", "fetch PDF for citation", "get full-text for DOI", "resolve cite to PDF", or any request to manage their Paperpile library or resolve a citation to a local PDF.

This skill should be used when the user asks to 'find repeated feedback', 'what do I keep correcting', 'capture this pattern', 'DRY my prompting', 'stop repeating myself', 'turn this into a check', 'automate this correction', or when the same type of feedback has been given 3+ times across sessions.

ALWAYS use when URLs in a manuscript need permanent archived copies — 'perma these links', 'archive the URLs in my footnotes', 'add perma links', 'the journal wants perma.cc cites', 'make sure these links don't rot', 'archive this site before it changes', 'snapshot these URLs', 'perma.cc', or when cite-checking a draft with web sources. ALSO use when perma returns 'you've reached your usage limit' despite an institutional account, or when asking which perma folder an account may archive into.

This skill should be used when the user asks to 'create a plugin', 'scaffold a plugin', 'set up plugin structure', 'new plugin', 'edit the plugin manifest', 'wire plugin hooks', 'validate plugin structure', 'audit my plugin's enforcement', or needs plugin-level work spanning multiple components. Use proactively whenever plugin.json, marketplace.json or hooks/hooks.json is being edited, even if the user never says 'plugin'. NEGATIVE ROUTING: never invoke plugin-dev:create-plugin, plugin-dev:plugin-structure or plugin-dev:plugin-validator directly — this skill is the wrapper that adds the enforcement audit they lack. Creating or editing a single skill, even inside a plugin, goes to skill-creator; designing, repairing or auditing a multi-phase workflow goes to workflow-creator.

Use when the user asks to "render pptx", "show pptx slide", "compare with pptx", "pptx to image", "export pptx slide", "original slide", "show me the original", "what does the pptx look like", or needs to extract a specific PPTX slide's content for visual comparison.

ALWAYS use for ANY query against the user's own reading library — "search Readwise", "find highlights", "what did I highlight about X", "get quotes from my reading", "search my annotations", "find that thing I saved about Y", "pull the full text of that article", "what do my notes say on this", "add my highlights to a notebook", "add tagged documents to notebook". Use even when the user never says "Readwise" but is asking about something they read or saved. NOT for web or academic search — this searches only their curated library.

Use when the user asks a QUESTION of their reading library rather than for a keyword lookup — 'ask readwise', 'chat about my highlights', 'what do my notes say about X', 'summarize my reading on X', 'what have I read about proxy advisors', 'compare what I've read about X vs Y', 'what's my take on X based on what I've read', 'pull together my highlights on X'. Use proactively for synthesis across many documents. NOT for finding a specific highlight by keyword — use readwise-search instead.

Use when the user says 'add to reader', 'save this article', 'read this later', 'put this in Readwise', 'list my documents', 'what's in my reading inbox', 'archive that article', 'tag this document', or 'delete that from Reader' - anything that lists, saves, updates or deletes documents in the Readwise Reader library. NOT for searching highlights (readwise-search) and NOT for bulk cleanup of stale documents (readwise-prune).

Clean up stale Readwise Reader documents. Use when the user wants to declutter their reading library, remove old unread articles, or manage Reader inbox. Triggers on "clean up readwise", "prune reader", "delete old articles", "declutter reading list".

ALWAYS use when the user wants something out of their own reading library — "what did I highlight about X", "search my highlights", "find in readwise", "my notes on X", "did I save anything on Y", "what have I read about Z", "find that quote I saved", "pull quotes from my reading", "who was it that said X". Use proactively before answering from memory whenever the user refers to their own reading or annotations. NOT for conversational summarizing across highlights (readwise-chat), listing or saving Reader documents (readwise-docs), or cleanup (readwise-prune).

ALWAYS use when code and data have to leave the machine as a self-contained archive for someone who did not write them — 'build the replication package', 'package this for counsel', 'send the workpapers', 'production archive for the opposing expert', 'replication archive', 'they want the code and data', 'put together what they asked for in the document request', 'make a workbook for the client', 'archive this analysis so someone else can check it', 'write the runbook', 'they need to be able to rerun the pipeline', 'document how the data was collected'. Use proactively whenever an expert report, a submission or a document request implies handing over the underlying work, even when the user only says 'zip it up'. NEGATIVE ROUTING: building the analysis itself goes to `ds`; a Word deliverable goes to `law-review-docx` or `docx-typst`; a single spreadsheet with no archive around it is plain `xlsx`.

ALWAYS use for ANY request to find academic papers - "find papers on X", "what does the literature say about X", "is there research on this", "who has written about X", "find me cites for this claim", "literature search", "any papers in the journals I like", "top journals only", "recent work on X", "has anyone studied this". Use proactively before answering an empirical or doctrinal question from memory, and never run scholar/consensus/paperpile by hand in sequence - this skill parallelizes them. NOT for open-web report writing (deep-research).

Use when the user says "set up workflows", "workflows setup", "is workflows installed correctly", "verify my install", "check the workflows install", "why isn't my agent's guidance loading", "my agent dispatch got denied", "/setup", or asks whether this plugin's agents, their preloaded skills and the main-thread guard's allowlist actually resolve on this machine. Use proactively right after installing or updating the plugin — a preload that fails to resolve is logged to the debug log only, and the run reads exactly as if it had loaded. Machine-level and idempotent; offers one optional user-tier plansDirectory write, always behind a question. NEGATIVE ROUTING: setting up a course's teaching plugin is teaching:setup and checking the Codex CLI is codex:setup; actually upgrading to a newer plugin version is plugin-update — this skill only checks what is installed, it never installs.

This skill should be used when the user asks to 'create a skill', 'improve a skill', 'edit a skill', 'add a skill to a plugin', 'add enforcement patterns', 'add Iron Laws or fact rows', 'fix a skill description', 'audit skill enforcement', or needs to substantially create or edit any SKILL.md file — including a single skill inside a plugin. Use plugin-creator only for plugin-level work (manifest, hooks wiring, multi-component scaffolding).