Author a new task for the eval suite — JSON spec + reference implementation + verified validation. Activate when the user asks to add an eval task, write a benchmark task, or extend the suite. Encodes the 6 hard-won pitfalls from prior post-mortems.
Add multi-language variants (Python / Go / C / C++ / Rust / Zig — 6 supported languages) to an existing eval task. Activate when the user wants to make an eval task language-agnostic or extend the variant rollout. Companion to eval-task-author.
Design a public API contract — function signatures, types, error model, example usage — BEFORE implementing. Activate when adding a new public function / endpoint / library, when changing an existing public surface, or when the user says "design an API for". Frontier models do this naturally; local models often code first and accidentally lock in bad shapes.
Produce a design doc BEFORE writing code for any non-trivial change — new subsystems, public API changes, anything >500 lines or crossing module boundaries. Activate when the user asks for a design, when you find yourself about to invent a non-trivial structure on the fly, or when "let me just start coding" would be premature.
Orient yourself in an unfamiliar codebase BEFORE editing. Activate at the start of any non-trivial task in a repo you don't already know cold — wrong file, wrong style, wrong layer is the most common local-model failure mode and this skill prevents it.
Clean commits, atomic units, meaningful messages, no random staging. Activate any time the user is about to commit, push, or open a PR — local models often stage too aggressively or write messages that don't explain why. Frontier models do this naturally.
Read, structure, and synthesize a large input (10K+ line PR, 30+ files, paper, transcript, log dump) without drowning. Activate when the input doesn't fit in working memory — when "just read it linearly" would consume your context budget without producing structured understanding. This is Kimi 2.6's natural strength encoded as a method.
Measure-driven performance work. Activate when the user asks to make something faster, when you spot what looks like a perf issue, or when adding code in a hot path. Encodes the discipline frontier models follow: profile first, optimize the actual hotspot, re-measure, never speculate.