Skill, agent, or ecosystem

Before building a capability, decide how much machinery its quality needs — not how much its plumbing needs. Getting this wrong is the failure Alex named directly: a working frontend, a working pipeline, and output that is quietly rubbish with no way to tell where it broke.

The three depths

Skill — one judgment, one pass. Quality is mostly a matter of instructions, so it lives in a SKILL.md a strategist can tune without a deploy. *Extracting features from a transcript.*

Agent — a worker plus a skill. Needs tool access, structured output, retries, or a budget. Still one judgment, but it has to interact with the system. *Reconciling new claims against the knowledge base.*

Ecosystem — several agents with different objectives, plus a critic, plus revision. Reach for this only when the test below says so.

The test

Can a single pass be wrong in more than one independent way?

If yes, one skill will not fix it — it will *average* the failure modes. A single prompt asked to nail the hook, the structure, the proof and the voice produces something mediocre at all four, and no amount of prompt editing separates them, because the model is trading them off internally where nobody can see it.

Scripting is the clear case: a script can fail on the hook, on structure, on using a claim the knowledge base does not support, and on sounding nothing like the client — independently. That is an ecosystem, and Alex has said so.

A second signal: would two competent humans in different roles disagree about it? If a strategist and an editor would argue over the same draft, that argument has to happen inside the system or it happens in the review queue instead.

The rule that prevents the bad outcome

Do not ship a UI for a capability whose depth was downgraded without recording it.

Building a thin version to unblock something is legitimate and often correct — the alternative is building nothing for weeks. What is not legitimate is letting the thin version *look* finished. A button that produces scripts implies the scripting system exists.

So when depth is downgraded, it goes in ../quality-debt.md with what is thin, what it should be, and what specifically will be bad about the output. That register is the answer to "we circled back and forgot".

Attribution: knowing where it broke

Bad output is only fixable if it can be traced to a stage. Every generated artefact records what produced it — skill name and version, worker, model, and the inputs it was given.

Without this, "the scripts are bad" is unactionable: it could be the extraction, the angle, the skill, or the model. With it, the same complaint points at one file.

This is cheap to add while artefacts are few and expensive to retrofit once there are thousands, so it goes in with the first ecosystem, not after.