llm-public-utils

Analyze code quality and perform targeted refactoring against established coding principles. Use when identifying and fixing "slop" in a codebase, running dedup sweeps, or reviewing code for maintainability violations.

Stress-test apps (especially webapps) — confirm reliability, comb happy paths, cause mischief, and hunt bugs in code — via agent browser, API scripts, or both. Use whenever the user wants to find bugs, edge cases, knots, break an app, harden UX or APIs, QA a flow, or catch issues before users do — even if they only say "try the app", "test this", "hit the API", or "see what breaks".

Research a new coding principle and merge it into the deslop skill reference. Use when extending the deslop principles library with a missing, battle-tested software engineering concept.

End-to-end repo housekeeping pass — branch sync, submodule hygiene, stuck-deploy recovery, doc triage, E2E smoke. Use at session start or when repo/deploy state is unclear. Reports surgically; only applies safe fixes.

Break down complex coding tasks into detailed, actionable markdown planning documents. Use when planning new features, refactors, or multi-step implementations before writing code.

Deliver harsh, multi-perspective critiques of text. Use when the user wants brutal editorial feedback that surfaces missing angles, weak arguments, and best-practice violations.

Apply concise, DRY, modular Python coding style with extensive logging and type hints. Use when writing or reviewing Python code where the user wants genius-level, self-documenting, deduplicated implementations.

Bring up a local/hybrid dev stack (Postgres + vector store + FastAPI + reverse proxy + SPA locally, optional remote LLM provider). Use to iterate on frontend, backend, and agent changes without a cloud redeploy.

Search PubMed and display structured paper metadata. Use when the user wants literature search results with titles, authors, abstracts, DOIs, and PubMed links.

Reformat academic papers while preserving exact text, applying template styling, and extracting figures. Use when converting a .docx paper to a journal template with validation.

Break large files into focused, single-responsibility pieces. Use when a file exceeds size thresholds (~250 LOC frontend, ~400 LOC backend) or mixes multiple responsibilities.

Separate overlay code into upstream metal and project slag. Use when deduplicating provider-agnostic improvements from project-specific overlay code in submodule/overlay workflows.

Iteratively compress text by combining phrases while preserving meaning. Use when reducing word count without losing intent; show incremental work at each pass.