Pick your harness.
Same skills, three delivery mechanisms. Editable files if you want to tune them, managed plugin if you don't, runtime-native if you run Hermes.
Pulls the kit as plain SKILL.md files into your repo. Read them, edit them, fork them — they're yours. Works with any agent that reads instructions.
Installs as a managed plugin. Auto-updates when we ship a new skill or tighten an old gate — you never run a stale verifier.
Adds a single skill straight into the Hermes runtime by slug. Per-skill granularity — install one today, the rest when you earn them.
pressure-test → spec → tickets → receipts → seal.
Five skills that chain. Run them in order and a vague idea comes out the other end as sealed, shippable work.
The agent attacks your plan before you've fallen in love with it.
Survivors become a written spec — interfaces, invariants, done-tests.
The spec is sliced into tickets an agent can grab without colliding.
Tickets get built in order, the tree staying green between slices.
The run seals into a ProofPacket — evidence, not adjectives.
Each output feeds the next input — the chain improves as you tune it.
Twelve we reach for daily.
Every card is a slash command, named for what it does for you. The "It's working if" line is the skill's own done-test — the agent knows what success looks like before it starts.
Find the holes in the plan before you build.
Common questions
When do I run it? Before any build you'd hate to redo. Thirty seconds of pressure beats a weekend of rework.
Does it ever stop asking? It stops when your plan survives three rounds without a change.
Turn a vague idea into a real spec.
Common questions
Is a spec always required? For anything that touches money or production, yes. For a throwaway script, no.
How long is a spec? One page. If it's ten, pressure-test it again.
Slice the work into tasks you can pick up today.
Common questions
How thin is a ticket? Thin enough to verify in one sitting. Thick tickets hide rework.
Who sequences them? The skill marks blockers and parallel-safe lanes; you make the call.
Ship the task with a proof attached.
Common questions
Where do the tests come from? From the spec's done-tests — never from vibes after the fact.
Can it parallelize? Yes, when tickets don't share files. It refuses the ones that do.
Lock the claim in so you can prove it later.
Common questions
What's in a packet? Inputs, outputs, gate results, timestamps, and the hash chain linking them.
Can a packet be faked? Touch any link and the chain breaks. Forgery is the one thing it can't do.
Try to prove done false before the user does.
Common questions
Isn't verifying the builder's job? Builders build. Breakers break. Same agent, both jobs, is how green lies happen.
What if it passes? Then — and only then — you're done.
Show the proof next to the claim.
Common questions
What counts as proof? A real command's real output, run now. A screenshot of last week's green is a story.
What if evidence is stale? Stale evidence is a failing gate. Re-run or remove the claim.
Block the ship until it is real.
Common questions
When do I run it? Before every merge, deploy, or “we’re done” claim.
What if one gate is flaky? Flaky is failing. Fix the gate or the claim — do not skip.
Catch when an old claim stops being true.
Common questions
Is this monitoring? Yes — for claims, not vanity uptime graphs.
What does it watch? Sealed hashes, owners, and the conditions they depend on.
Get a real number, not a story.
Common questions
Who is this for? Anyone about to bet money, time, or reputation on a maybe.
Does it need historical data? It needs honesty. History makes it sharper.
See how often the calls are actually wrong.
Common questions
What's a good Brier score? Lower than last quarter. Calibration compounds like interest.
Why should I care? An uncalibrated agent lies to you with a perfectly straight face.
Catch ghost seats and double charges.
Common questions
When do I run it? Before approving any vendor invoice tied to agents, tokens, or seats.
Does it replace finance? No. It gives finance a sealed exception list instead of a PDF shrug.
Type the slash command. Keep the receipt.
Free kit, MIT, sealed runs. When you outgrow generalists, the marketplace is one door down.
install-a-whole-playbook
Git-imported skill packs with org promotion. Install a sequenced playbook once, grant it per workspace, and keep versions sealed — not one-off paste prompts.