Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning neural networks, debugging loss spikes or OOM, choosing architectures, or optimizing GPU throughput.
Use when the user asks for caveman mode, fewer tokens, brief responses, compressed communication, or otherwise explicitly requests a much shorter answer.
Use when facing 2+ independent tasks without a written plan, with no conflicting shared mutable state or sequential dependencies, where parallel delegation beats inline cost; otherwise inline. Planned tasks use subagent-driven-development.
Use when the user asks to establish shared project language, or project work exposes a conflicting, renamed, or deprecated domain term that needs active semantic modeling. Routine small tasks stay on the fast path.
Use when executing a written implementation plan across sessions or with review checkpoints. Small or single-slice plans stay inline. For same-session independent tasks, use subagent-driven-development instead.
Use when verified work needs integration or cleanup of an existing task-created branch/worktree, or the user explicitly requests merge, PR, or branch lifecycle handling.
Use when asked for first-principles or Occam's-razor review, or when high-risk decisions involve competing constraints, fallback growth, duplicate owners, or architecture direction risk. Ordinary bug fixes stay on the fast path.
Use when the user explicitly sets an Aegis goal with /aegis-goal, Aegis goal:, or asks to define goal, success evidence, stop condition, or task boundaries before work.
Use when receiving code review feedback before implementing suggestions, especially when feedback is unclear, risky, disputed, or technically questionable.
Use when the user asks to create, write, update, amend, supersede, or evaluate an ADR, architecture decision record, durable architecture decision, decision log, or baseline sync after architecture-changing work.
Use when requesting independent code review, after implementation slices, before merging high-risk work, or when verification exposes evidence, baseline, architecture, compatibility, or retirement uncertainty.
Use when executing a written implementation plan with independent tasks in the current session where delegation beats inline coordination cost; otherwise inline. Ad-hoc 2+ tasks without a plan use dispatching-parallel-agents.
Use when the user says `aegis:update`, asks to update or upgrade an installed Aegis method-pack, wants the latest Aegis version, or asks whether Aegis is current on this host.
Use when a coding task needs a concurrent checkout, unrelated dirty state blocks safe branch switching, or the user or repository explicitly requires a worktree.
Create animated flow diagrams from articles, workflow notes, architecture sketches, or process descriptions using a JSON specification and a local Python/Pillow renderer. Use when the user wants to turn source material into a static PNG + animated GIF diagram.
Pure API reference for reddapi.dev - authentication, all endpoints (vector search, semantic search, trends, subreddit lookup), request parameters, response schemas, and error codes, with no research-workflow framing. Use when the user wants raw endpoint documentation, is debugging a reddapi.dev integration, needs exact request/response field names, or asks for 'reddapi API reference', 'reddapi.dev endpoints', or 'reddapi error codes'. For guided research workflows and query playbooks, see reddit-research. For B2B lead scoring, see reddit-leads.
Use when multiple workflows duplicate the same operational logic, when deciding what belongs in actions vs shared services, or when refactoring repeated operational blocks across domain flows. Use when adding new features that share mechanics with existing ones.
Iteratively improves a PR (GitHub), MR (GitLab), or shelved changelist (Perforce) until Greptile gives it a 5/5 confidence score with zero unresolved comments. Triggers Greptile review, fixes all actionable comments, pushes/re-shelves, re-triggers review, and repeats. Use when the user wants to fully optimize a PR/MR/CL against Greptile's code review standards.
Iteratively improves a PR (GitHub), MR (GitLab), or shelved changelist (Perforce) until Greptile gives it a 5/5 confidence score with zero unresolved comments. Identical to greploop, but triggers reviews by tagging @greptile-apps, which bypasses Greptile's file-count limit on huge PRs that the plain @greptile mention refuses to review. Use when the user wants to fully optimize a large PR/MR/CL against Greptile's code review standards.
OpenRaiser/NanoResearch1.3k
