zorost/ai-engineering-lab/using-zorost-skills

Using Zorost Skills

1 · Purpose

Make the agent choose the right skill for the task at hand, run it to its verification gate, and report honestly when no skill fits.

2 · When to use

• At the start of any non-trivial AI engineering task when this catalog is installed.
• When the agent is improvising a workflow that a catalog skill already covers.

Do not use this meta-skill for trivial one-line questions or file lookups. Answer those directly.

3 · Inputs

• The user's request, verbatim.
• The catalog table in reference/skills/agent-skills/README.md (names + descriptions).

4 · Procedure

1. Restate the task in one sentence, starting with a verb: "Build…", "Fix…",

"Decide…", "Deploy…".

1. Scan the catalog descriptions. Match on the task's stage, not its nouns:

deciding whether to fine-tune is fine-tune-readiness, not local-model-fit.

1. If exactly one skill matches, load its full SKILL.md and follow its procedure.
2. If two skills match, run them in lifecycle order: Define before Build, Build

before Verify, Verify before Ship.

1. If no skill matches, say so in one line, "No catalog skill covers X; proceeding

ad hoc", then proceed. Do not force a near-match.

1. While running a skill, obey its STOP conditions. A stopped skill is a success

state with a report, not a failure to hide.

1. After the skill exits, state which skill ran and what its Verify section produced.
2. If the skill's steps proved wrong or stale, open an issue or amend the skill file.

A wrong skill that stays wrong teaches the next run to be wrong.

5 · Anti-rationalization

6 · Red flags

• The agent claims a skill was followed but cannot name its Verify evidence.
• The same skill is loaded three turns in a row without its procedure advancing.
• A skill's STOP fired and the run continued silently.

7 · Verify

• The chosen skill is named in the run log.
• Its Verify section's evidence exists (file, score, test output, or STOP report).
• If no skill matched, the one-line "no skill covers X" note exists.

8 · ZoroLogistics example

User: "Make the BoL extraction prompt better." The agent restates: "Improve a scored artifact." Stage = Build-on-measured-loop → prompt-suite-versioning. It loads that skill, which requires the golden set and scorer, which eval-first-development already produced. Sequence: eval-first (already done) → prompt-suite-versioning → error-analysis-50 on the failures.

© 2026 Zorost Intelligence LLC · zorost.com