Skills library for a research team
Below is a snapshot of a research team’s shared skills library.
A snapshot of the full skill files is here.
| Name | Description |
|---|---|
| adversarial-prompt | Red-team a claim or document from five independent angles at once, surfacing empirical, logical, methodological, historical, and strongest-opposition critiques. |
| ask-many-contexts | Compare the same prompt under full project context, base context, and zero context to test whether added scaffolding improves responses. |
| ask-many-models | Send the same prompt to multiple AI models in parallel, then synthesise their responses to surface consensus, disagreements, unique insights, and confidence level. |
| ask-many-times | Send the same prompt to multiple fresh model instances and compare where answers converge or diverge. |
| ask-many-ways | Generate several prompt variations, run them all, and compare how wording affects the result. |
| robustness-stress-test | Run identical, variant, and leading prompts to test stability, sensitivity, and sycophancy. |
| blind-review | Strip identifying information from a document before evaluation, then compare blinded and unblinded assessments to detect prestige or author bias. |
| back-of-envelope-brief | Draft or revise a structured technical brief with rough calculations, tables, assumptions, and formatted derivations. |
| prompt-chain | Run multi-step prompt chains inside an agent conversation, including workflows defined in a chain file. |
| consensus-check | Investigate a claim by independently searching for evidence for and against, then synthesise the state of the evidence. |
| model-debate | Run a structured debate between two models with a moderator, producing a transcript and map of cruxes without declaring a winner. |
| decompose | Break a complex question into answerable subquestions until each leaf is a search, calculation, lookup, or concrete action. |
| deep-review | Run a comprehensive document review using parallel roles such as editor, fact-checker, calculation auditor, red team, extensions reviewer, and simulated readers. |
| diagrams | Create diagrams, flowcharts, charts, graphs, infographics, and data visualisations for research, papers, or presentations. |
| epistemic-audit | Chain multiple review methods into one full epistemic audit: adversarial review, consensus check, steelmanning, premise audit, and blind review. |
| explore-tree | Recursively branch out from an input, question, word, or document to map promising directions and unexplored territory. |
| fact-check | Fact-check research documents by tracing verifiable claims to sources, assessing consensus, and producing severity-rated findings. |
| feedback | Capture feedback on the tooling, installation, individual skills, bugs, or feature ideas. |
| handoff | Checkpoint an in-flight project into a handoff file so future sessions can resume without losing state. |
| new-project | Scaffold a new research project folder with a standard layout for memo, notes, tasks, drafts, research, and code. |
| premise-audit | Surface hidden assumptions in an argument and rate how load-bearing each one is. |
| research-sprint | Run a multi-agent research fan-out and consolidate the results into one research-notes file. |
| skill-help | Print a quick-reference table of available skills and help answer questions about when to use them. |
| steelman-duel | Generate the strongest possible case for each side of a debate using separate agents that cannot see each other’s work. |
| style-guide | Generate or review writing against a house style for abstracts, blog posts, research notes, papers, threads, and public-facing content. |