Twenty-two original skills for agents that have to prove what they did. Each one answers a single question, states when it must not fire, and ends in a checkpoint someone else can re-derive.
Why these exist
An agent that reports success is not the same as an agent that produced it. Most of these procedures were written after something in our own production went quiet rather than loud: a write that returned 200 and destroyed a published article body, a judge that scored a truncated file five out of five, a content run that published 270 pages and earned four clicks with every gate green.
So they share four rules, and the rules are why the set holds together rather than being twenty-two unrelated files:
- A success status is not evidence. Any step that claims something landed reads the state back from whatever owns it.
- Guards detect on form or size — a count, a byte range, a computed style, a difference in whole minor units. Never on a green check.
- Loops carry a budget and an exit. A recovery path with no maximum is a hang wearing a recovery’s clothes.
- Done means validated, not produced.
These are ours
Every file is written from scratch. Where a third-party pattern informed one, it is credited inside that skill as inspiration and nothing more — measured seven-word overlap against every source file consulted is 0.00%, and the measurement is a deterministic script, not an opinion. That matters if you intend to use them commercially: you are not installing somebody else’s repackaged work.
The twenty-two
Value and planning gates
- baseline-before-build — Before I build a batch of similar items, which of them are worth building at all?
- plan-and-prove — How do I break this project into tasks whose completion can be proven, not just reported?
Verification core
- read-back-pattern — How does an agent verify that an API write actually happened, when the platform reports success either way?
- change-review-gate — This change is about to land — which of my findings block it, and which are only worth a comment?
- exact-math-verification — How do I make an AI’s math exact and provable instead of approximately right?
Publishing and editorial
- verified-publishing-gate — How do I gate AI-generated content so only verified work goes live?
- seo-preflight-audit — Does this article pass a scored on-page SEO audit — and exactly what must be fixed before it is allowed to publish?
- published-content-sweep — Which of my already-published pages are silently non-compliant, and am I measuring the thing the reader actually sees?
- discovery-article-voice — How do I make AI-written articles read like a person documenting a real finding instead of generic AI advice?
- draft-revision-loop — How do I revise a draft into clearer, tighter prose while provably preserving its meaning and the author’s voice?
- comms-send-gate — How do I draft a business email whose ask is extractable in one read, and that passes a send-readiness gate before it goes out?
Media production
- image-production-gate — How do I produce an article image and prove the delivered file meets spec before it ships?
- deck-from-source — How do I build a slide deck from a source document so that every slide is traceable back to a specific passage and provably legible in the rendered output?
- slide-deck-fit-gate — How do I build a slide deck programmatically and prove every slide renders without overflow, cutoff or overlap before delivery?
- transcript-repurposing — How do I turn a long video or audio source into timestamped written content whose quotes provably match a hash-locked reference transcript?
Platform and data
- wp-rest-safe-write — How do I change content through a REST API that returns 200 for writes that did not land, without destroying what was already there?
- jetengine-headless — How do I inspect and build a JetEngine data layer on a WordPress site from outside, with no browser and no admin session?
- bulk-harvest-reconcile — How do I pull structured data out of thousands of remote records with parallel workers, and afterwards prove that what I have is complete and correct?
- design-token-system — How do I stop a batch of generated pages from drifting apart visually, and prove mechanically that none of them bypassed the shared appearance layer?
Business operations
- payment-integration-hardening — Which parts of this payment flow will answer OK while the money state is actually something else, and what do I read back to catch each one before it ships?
- subscription-billing-ops — After a billing period has run, which subscriptions no longer agree across payment provider, access system and ledger — and in what order do I repair them without moving money twice?
Agent infrastructure
- agent-memory-layer — When an agent recalls something it wrote down in an earlier session, how does it decide whether that claim is still safe to act on?
How they are checked
Four layers, and each was added because the one before it missed something. A judge reads the file and scores it — it sees form. An independent reader hunts logic defects — it finds guards that read fields no step produces. A systematic second pass re-reads the fixes — it finds defects the first pass introduced. Then the skill is run on real input, which is where the rest surface: three layers of review passed a slide checker that inspected thumbnails where the text was four pixels tall.
Getting them
The skills install as one plugin that loads them on demand, so nothing sits in your context until the description matches what you are doing. They are part of membership rather than sold separately.
We are still in early access and there is no self-serve checkout yet. Register and get in touch, and you get the library as it stands, labelled honestly. What membership covers →
Members reach the twenty-two, their versions and the current package on the originals page inside the member library.
22 skills, verified against their own checkpoints. Last built 2026-09-04.