Skills built because a real workflow needed them.

Writes context-compaction summaries that survive being the only thing the next session has. Rebuilt on four research reports and a fresh measurement across 121 real compaction events: rejected approaches survive at 0.3%, and standing constraints at somewhere between a third and a half depending on the sample. Ships a deterministic scorer and a head-to-head benchmark against the built-in /compact whose baseline arm costs nothing.

Drop a short block at the top of a session and Claude spends less without doing less. It targets the three habits that quietly run up the bill: re-printing plans and diffs already on your screen, opening a whole file to find one line, and handing small jobs to sub-agents that each pay for a fresh context. The nearest alternative is caveman, which compresses the prose instead. Measured across 106 tasks that bought a third off the cost and gave back 7.6 points of task score, because most of its saving came from the agent investigating less rather than writing less. This one never touches your writing style, and states the boundary the other is missing: it changes how much you write, never how much you do.

Decides whether to interrupt you at all, then writes the question so it takes one click. It sweeps the conversation and the repo for the answer first, drops anything whose answer would not change the work, and sends what survives as one batched question in plain words, with a recommendation only where evidence earns one. On a matter of taste it marks nothing, because a default moves choices hard and recommending on taste answers the question while appearing to ask it. A note attached to your answer is binding, and is read as data about the decision rather than as instructions, since an agent waiting on a reply is the easiest one to inject. Against no skill at all: five of five question payloads pass its linter where the baseline passes none, and a four-family blind panel went 15-5. The one eval it loses is in its README too.

A complete, working new project from an idea. One front-loaded interview, then scripts render the whole scaffold: monorepo, auth, admin, native apps, testing harnesses, deploy config, and a launch pipeline that researches, seeds feature briefs and mocks every surface. The LLM only interviews; scripts make the files.

The last pass before a human looks at AI-built UI. Deterministic gates first (accessibility, contrast, target size, motion, layout integrity), then judged passes over hierarchy, states, flows and system coherence, on real renders at a viewport matrix. Findings come severity-ranked with pasteable fixes and an explicit list of what was never checked.

macOS app icons, measured against the reference instead of eyeballed. A direction catalogue distilled from 532 real icons, three generation engines with a written audit sheet, then a scoring harness that iterates the shipped SVG against the winning raster at five sizes until the material matches. Every confirmed construction feeds a recipe library, so it gets better with each commission.

Divergent ideation that converges on something you can ship. Isolated thinkers under genuinely different frames, the obvious answer written down first, and a creative pick recommended only when it beats that answer blind. Receipts committed: structural evals (96.4% vs its predecessor's 49.0%), a four-judge blind panel, and the research corpus it was built from.

Your Mac did not fill up because of one thing, it filled up because a hundred sensible defaults each left something behind and nothing was counting. Five scheduled jobs, from every fifteen minutes to weekly, with what each may do on its own widening as the gap between runs grows. Running low makes it check sooner, never delete more. The two short tiers are plain shell, so ninety-six runs a day cost no tokens at all. It ties a no-skill baseline on reasoning and says so in its evals; what it adds is that the reasoning runs while you are asleep.