Does good AI work come from better prompting or from better systems?

A prompt works once, for one agent. A system of clear tasks, reusable skills and a fixed sequence of checks works again with every draft, before a human decides. The YMAD standard puts it almost word for word: quality comes from clear tasks, reusable skills, self-checks, review gates, a human decision and feedback loops. Such a system rests on six parts: clear tasks instead of vague instructions, skills an agent can reuse, a self-check before every submission, review gates that go through the draft once more, a human decision at the end, and a feedback loop that carries every correction back into the system. Three words carry the whole standard. An agent is an executing role. A skill is a reusable capability or rule that an agent loads when it needs one. A gate is a checkpoint or an approval point. The fourth word is bigger: a department is a managed area of work with a goal, a rhythm, roles, data, a quality standard and a feedback loop. Nine skills are named and written out, so nobody has to invent them again each time.

Agent, [[skill|skill]], [[tor|gate]]

Three words carry the whole standard. An agent is an executing role. A skill is a reusable capability or rule that an agent loads when it needs one. A gate is a checkpoint or an approval point. The fourth word is bigger: a department is a managed area of work with a goal, a rhythm, roles, data, a quality standard and a feedback loop. Nine skills are named and written out, so nobody has to invent them again each time.

A prompt is the instruction for one case. It can be clever and still leave nothing behind, because it ends with the session. A skill stays. It is written once, then loaded, checked and improved. That is also what the larger question is about, how single prompts turn into workflows and repeatable quality.

  • A fixed sequence of checkpoints that does not depend on the day.
  • Skills that are written once and loaded again and again after that.
  • A self-check the agent runs on its own work before handing it over.
  • Four clearly defined answers per checkpoint, each with a reason.
  • A human approval as a station of its own in the process.
  • A memory that keeps every correction and carries it into the next draft.

How a draft reaches approval

Content that leaves the building follows a fixed sequence of eleven gates, from Business Context to Feedback Memory. Brand Voice and Claims & Proof are the points where the guardian of the brand works, the first agent that produces nothing and only checks. Risk & Legal Preflight follows directly. Human Approval stands second to last, after Review Submission and before Publish Handoff. That makes approval a step of its own in the process.

Each single gate returns one of four answers. Pass, when the draft can continue. Revise, when it needs a change. Block, when it cannot continue. Escalate, when a human or a specialist has to decide. The word alone is never enough. Every answer carries a short reason, the concrete change or escalation, the affected passage in the output, and a confidence, from low through medium to high.

The point of all this bookkeeping shows up with the person who approves. They should see the results of every gate right next to the draft instead of piecing together afterwards what has already been checked. What that looks like in a live mandate is on the page about AI implementation.

When an agent escalates

One skill defines when an agent hands the decision on. Legal risks. Data protection risks. Any movement of money. Live publishing. An unclear promise to a customer. High reputational risk. Missing access. A contradictory briefing. The legal tone is a system rule too: the agent marks risk signals, missing evidence, data protection topics and platform rule risks, and it writes that this is not legal advice.

Another skill takes care of cost. Agents only carry out work that makes sense, pause when there is no assignment, and report what they used. The standard has one sentence about that which holds for people as well.

A department is allowed to pause. Idle time is not a fault. An agent that invents work in order to look busy is a governance problem.

The shorter path for technical work

A deployment runs through the same logic, only shorter. Seven stations in a fixed order, and memory stands at the end again.

  • Business Context
  • Technical Scope
  • Self-Check Before Done
  • Technical QA
  • Security / Secrets
  • Human Deployment Approval
  • Feedback Memory

The same pattern in my own work

I have been practising this pattern for longer than the YMAD standard has existed. A recording from 17 December 2025 shows my own work with a coding agent following exactly this sequence: it has to think a thing through and document it first, I read the documentation and check whether it fits, and only then does the actual build follow. That is a gate in miniature, months before it carried that name.

I said the same thing about my personal setup in the KI DeepDive. What matters is a setup in which context, access, skills and review processes work together reliably. Access is half of it, and there is a note of its own on why access beats the prompt.

The same holds for the offer itself. What a company gets is the working structure: sources, responsibilities, review packages, gates, approvals and feedback that flows back into the next version. Anyone who wants to settle the larger question first will find it under which system do I actually want to build.

From the YMAD standard

YMAD Quality Gates and Skills, internal ZukunftBilden standard document, status 30 April 2026. Core terms from the YMAD Department Library.

  • Eleven gates in a fixed sequence for external outputs, from Business Context to Feedback Memory
  • Seven gates for technical outputs, also ending at Feedback Memory
  • Each gate decides between pass, revise, block and escalate
  • Every decision carries a reason, the concrete change, the affected passage and a confidence from low to high
  • Nine named skills, among them the Escalation Skill, the Legal Check and the Cost-Aware Heartbeat
  • Brand Voice and Claims & Proof run before Risk & Legal Preflight
  • Review Submission and Human Approval stand before Publish Handoff

This note keeps growing

2026-09-03: Deepened with a recording of my own from December 2025: the same order in miniature, months before the YMAD standard.

2026-09-03: Deepened: the core terms agent, skill and gate, what belongs to every gate answer, the rules for escalation and cost, the seven stations for technical work, the link to my own setup, source list.

2026-09-02: Planted from the YMAD standard document on gates and skills.

ZukunftBilden GmbH · Salzburg · +43 681 81655313 · office@zukunftbilden.eu