blakecyze

Use when the user asks for a code review, audit, pre-PR check, quality sweep, or pattern analysis of a diff, branch, module, or codebase. Also use when the user asks to "check" or "look over" their code for issues.

Use when the user asks to update, prune, audit, or create AGENTS.md, CLAUDE.md, or other agent context files. Also use when the user says their AI agents keep making the same mistakes and they want to fix it at the context level.

Use when the user asks for a nuclear review, a thermonuclear audit, a deep maintainability sweep, a structural audit of the whole codebase, or a "harsh" pass on a branch. Stricter and more structural than /kanso-audit.

Use when generating, reviewing, modifying, or discussing any code. Sets standing anti-dilution rules for all code-related work. Auto-loads; not directly invoked by the user.

Use when the user asks to clean up, tighten, simplify, de-bloat, de-slop, or refactor code. Also use when acting on findings from /kanso-audit or /kanso-nuclear. Behaviour-preserving only.

Use when the user wants a task done well rather than fast, wants a rough or underspecified request sharpened before it runs, or asks for a careful, principled pass at a piece of work.

Use when rewriting, sharpening, or producing a prompt for Claude or another current frontier model. Sets standing rules for getting better LLM output through better prompts. Loaded by /kanso-task; not directly invoked by the user.

Use when the user asks to write a pull request description, open a PR, draft PR notes, or summarise a branch for review.

Use when the user wants to run plans written by /swarm-improve, execute a plans/ backlog, or dispatch mechanical work to a cheaper model. The pair of /swarm-improve: it planned, this executes.

Use when the user wants to audit a whole codebase for latent work (bugs, performance, tech debt, missing tests, what to build next) and capture it as self-contained plans a cheaper agent can execute later. Spends the capable model on judgement, not execution.…

Use when orchestrating subagents, fanning work out in parallel, or running any /swarm-* skill. Sets standing rules for when to fan out, when not to, cost ceilings, and who judges. Auto-loads; not directly invoked by the user.

Use when the user wants to try several independent approaches to the same problem and keep the best one. Best-of-N generation. Expensive by design; respects the cost ceiling.

Use when the user wants to map a large codebase, gather context from many files at once, or investigate a problem space in parallel. Read-only fan-out. Never writes.

Use when the user wants to break a large task into parallel workstreams, plan a multi-agent job, or decide what can run concurrently. Produces an orchestration plan, does not execute it.

Use when the user has one artefact (code, doc, design, plan) and wants it hardened by attacking it from several angles at once. Parallel critics, then merge. Not to be confused with bulldoze, which generates many solutions.

Use when the user wants prose made shorter, tighter, or denser without losing meaning: "make this concise", "tighten this", "cut the fluff", "too wordy", "trim this down", "say it in fewer words". Strips throat-clearing, dutiful summaries, hedging, puffery…

Use when the user wants prose to sound human or to stop reading as AI: "humanise this", "make this sound human", "does this read as AI", "remove AI tells", "de-slop this", "make it less ChatGPT". Runs read-only audit (flag every tell with location and…

Standing anti-tell rules for prose generated while mimesis is installed: no em dashes, no kill-list slop, no negative parallelism or participle tails, varied rhythm, a point of view. Auto-loads; not directly invoked by the user.

Use when the user wants text written or recast in a specific named human voice: "make this sound like an expert / a practitioner", "make this warmer", "make this blunter / sharper", "make this more persuasive / sell this", "make this plainer / simpler", "give…

Use when the user wants text written for a machine reader, not a human: "write a prompt for", "compress this for an LLM", "turn this into instructions for an agent", "make this a system prompt", "compile this into a prompt". Produces dense, structured,…

Use when the user wants to know if an interface looks AI-generated, or wants the machine-made design tells stripped: "is this design AI-generated", "de-slop this UI", "does this look vibe-coded", "why does this look like every AI site". Read-only audit of…

A note library that remembers how hard problems got solved, so the next session does not re-derive them. Recall runs when a non-trivial task starts: it reads a bounded index, then greps notes for the task tools, errors and domain terms. Capture writes one…