Workflows, skills and hooks

Into the guide: scripted multi-agent workflows, skills that fire before the work starts, and a hook fault that only shows with two sessions open.

This is the update edition. If you’ve read Mastering Claude Code and you’re back to catch up, this is the diff: just what changed this week, so you don’t have to go through the whole guide again. There are three changes worth your attention, and one piece of housekeeping at the end.

A second kind of skill

The first change is in the skills chapter. Until now the guide described skills as procedures you invoke: a deploy, a style pass, a sequence you’d otherwise re-type. There’s a second kind worth knowing about. Some skills encode how to approach a class of work, and their whole value is that they fire before the work, not during it. A brainstorming skill that runs before you build any new feature, to pin down what’s actually wanted. A test-first skill that insists the test exists before the code. A debugging skill that makes you find the cause before proposing a fix. Call these process skills. The catch is that the discipline only works if it’s consulted before you charge in: a process skill you remember only after you’ve picked an approach has done nothing. And you don’t have to write these yourself; they increasingly arrive as installable plugin packs, which is also the fastest way to learn what a good skill looks like from the inside.

Hooks that leak across sessions

The second change is a hooks failure mode that only shows up once you run more than one session at a time. The moment a hook writes a file (a flag that unlocks a guard for one operation, a lock that says this session owns the inbox, a marker that stops work being done twice), that file is shared by every session running at once. Give it a fixed name and two sessions collide: a one-shot unlock you granted in one terminal satisfies a guard in another, or a flag left behind by a session that has since died blocks a live one. The fix is dull and general: put the session’s own identifier in the filename, so a bypass stays where you granted it.

Handing over the control flow

The third change is the largest, and it’s in the multi-agent chapter. The guide already described five patterns for getting a second pair of eyes: the critic loop, the planner and executor split, the second opinion, parallel fan-out with synthesis, and build-and-grade. All five are things you drive by hand: you spawn an agent, read what it says, decide, spawn the next. That’s right while you’re still feeling out a problem. But once the shape of the work has settled (fan out across this list, verify each finding, keep the survivors), you can hand the control flow itself to a script. That script is a workflow.

A small vocabulary covers most of it. A pipeline runs each item through every stage on its own, so one finding can be at the verify stage while another is still being discovered. Nothing waits for the slowest item at each step, and that’s the sensible default. A barrier is the opposite: it collects every result from a stage before the next one starts, which is right only when the next stage genuinely needs the whole set, to merge duplicates, or to stop early if the count came back zero. Loop-until-dry keeps spawning finders until a round or two turns up nothing new, which is the honest way to size a search whose answer you can’t know in advance. Adversarial verify spawns independent sceptics for each finding, briefed to refute it, and keeps only what survives a majority. That’s the step that stops a confident wrong result from reaching you. And a judge panel generates several independent attempts, scores them, and synthesises from the winner.

Two cautions came with it. Reach for a workflow when the work is large and its shape is known. When the shape isn’t known yet, do the reverse: scout by hand, discover the work-list, then hand it over, because automating a guess only produces a great deal of confident guessing. And there’s a posture worth naming on its own: exhaustive mode, the deliberate choice to spend tokens buying confidence, fanning out wider and verifying harder than the answer strictly needs, because this particular decision is expensive to get wrong. It’s the right call for the irreversible and the costly, and the wrong default for a quick lookup.

Housekeeping

One cosmetic note for anyone reading the guide on the web. The site’s display headline used to be set in a serif typeface; it’s now set in the same sans-serif as the rest of the site, for consistency. Nothing in the content changed with it.

That’s the week. As always, the full written guide, with tables, diagrams, and the copy-paste prompt for spinning up a fresh assistant that already knows all of this, lives at claude-guide.sauer.com.au. Share it with anyone who’d find it useful.

The cards