Skill-Driven Loop Development

The precondition discipline for building an Agent Loop: never build a loop without battle-tested skills behind it. A loop is only as reliable as the parts it drives, so its steps should be execution skills you have already validated by hand — not fresh, unproven instructions the loop discovers at run time. The structure separates two kinds of skill: a single orchestration skill that kicks off and drives the whole loop (it holds the goal, how to complete it, and how to verify), and the execution skills it calls — each a single, specialized, already-battle-tested job. The reason skills matter is that they encode how you personally want a task done: without a /analyze-workout skill, a weather-check loop just says "it's raining, cancel your run"; with one that records you love running in the rain, the answer flips. And because a skill can be upgraded to emit its own verification (approved / not-approved, or a 1–10 score), skill-driven loops are what let even abstract, non-quantifiable goals become verifiable — the bridge from "is this good?" to a concrete signal the loop can gate on.

This is distinct from its neighbours: Reusable Workflow Library is about cataloging and sharing workflows; skill-driven loop development is the narrower rule that the loop's steps must be pre-validated before you automate them. Loop Engineering is the role shift; this is the build discipline that keeps the shift from producing brittle loops.

Claims

Don't build a loop without battle-tested skills behind it — a loop inherits the reliability of its parts. best practice — context: composing loops for real, repeated work; automating unproven steps just runs an unproven process faster. The bar is "already validated by hand," mirroring the Agent Loop two-occurrence rule for what counts as proven.
Separate the orchestration skill (one skill that drives the whole loop: goal, how to complete, how to verify) from execution skills (each a single specialized, pre-validated job it calls). best practice — context: structuring a skill-driven loop; the split keeps the loop's control flow in one legible place and its work in reusable, testable units.
Battle-tested skills encode how you want a task done — that personalization is what makes loop output correct rather than merely plausible. observation — the /analyze-workout example: the same trigger yields opposite advice depending on whether a skill carries the operator's actual context. This is Evergreen vs Volatile Context and personal-dedup applied inside the loop.
A skill can be upgraded to carry its own verification element (approved/not-approved or a score), which is the bridge that turns an abstract goal into a verifiable one. principle — durable: a loop needs a real acceptance check (Agent Loop); when the goal isn't a metric you manufacture a checkable signal by giving the reviewing skill a concrete verdict to emit. See LLM-as-Judge for making that verdict trustworthy.
Data ingestion is itself skill-driven: build and test a skill that processes raw data exactly how you want before wiring it into an automated pipeline, then have an orchestration skill run the validated utility skills together. best practice — context: a second, independent source (Austin Marchese) applies the pre-validate-then-automate discipline to a Self-Improving System's inflow — "99% of people get this wrong" by automating an untested ingestion step; a /data-ingestion orchestration skill calls sync-claude-sessions, sync-ecosystem-data, and sync-curated-content utility skills.
Close the loop by hand to improve the skill: after you go back-and-forth to fix an output, say "based on this conversation, improve this skill" rather than waiting for the automatic loop. best practice — context: keeping a skill sharp under real use; the corrective session is training data for the skill, and deferring it to the auto-improve loop wastes the signal.

Agent Loop — the artifact this discipline feeds: skills are the execution steps and the acceptance check inside the loop.
Loop Engineering — the role shift (designing loops instead of prompting); skill-driven development is the build rule that keeps designed loops reliable.
Reusable Workflow Library — the adjacent, broader move: catalog and share workflows; here the point is narrower — pre-validate the steps before automating them.
LLM-as-Judge — how the verification element a skill emits is kept honest (calibrated, ideally independent).
Loop Training Mode — the runtime complement: even skill-driven loops are run supervised first before they go autonomous.
Distillate: Loop Engineering, Illustrated: Triggers, Skills, Verification, Memory.
Distillate: How to Build a Self-Improving System with Claude Code — skill-driven data ingestion + orchestration-over-utility skills, applied to a self-improving system's pipelines.
Distillate: You're the Problem, Not Claude — Six Fixes to 10x Output — the same author's augment-then-automate ordering: "only after you've systematized it manually is it time to automate; automating something you haven't systematized manually is wasting your time."
Self-Improving System — the system these validated skills compose into.

Skill-Driven Loop Development

Claims

Related

Linked from