AI Research Workflows

Use when you need to understand how an existing codebase works and/or survey external prior work (papers, methods, tools) before planning or changing something. Traces function call paths, maps module dependencies, searches academic papers and documentation, compares library alternatives. Triggers: research the codebase, how does X work, where is X implemented, what tools or libraries exist for X, prior art on X, has this been done before, survey approaches to X.

Use when an existing implementation plan needs changes before or during execution. Triggers: update the plan, change the plan, add a phase, revise scope, incorporate experiment results.

Use when work context must transfer to another session or person. Triggers: create a handoff, hand off this work, summarize context for the next session, prepare a handoff.

Use when an approved implementation plan (a docs/rse/specs/plan-*.md file) exists and the next step is writing the code. Triggers: implement the plan, execute the plan, start building, run the plan.

Use when the best technical approach is genuinely uncertain and a head-to-head prototype comparison is needed before committing to a design. Triggers: should I use X or Y, compare approaches, benchmark, which is faster or simpler.

Use when research or scientific code must be trustworthy — verifiably correct, regression-safe, and numerically stable. Triggers: harden this code, is this numerically correct, add regression tests vs known results, make the research code robust.

Use when a result, experiment, or analysis must be reproducible by others or by a future session. Triggers: make this reproducible, capture provenance, pin the environment for this result, why can't I reproduce X.

Use when a feature, refactor, or multi-file change needs to be designed before coding. Triggers: plan X, design the implementation, how should we build X, create an implementation plan.

Use when starting or continuing ANY research-software task — understanding or modifying code, surveying prior art, planning, experimenting, implementing, validating, reproducing, or hardening — and before reading code to "just check", before editing, or before answering "how does X work". Triggers: research workflow, which skill should I use, where do I start, guide me through the workflow, structured development.

Use when an implementation is (claimed) complete and must be checked against its plan before shipping. Triggers: validate the implementation, verify it matches the plan, is the implementation correct, check before PR.