Source: raw/Stop_Prompting_AI_Agents._Build_Loops_Instead..md (auto-caption transcript)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=JMOI5OimVYs · Duration: 20:12 · Uploaded: 2026-08-12 · No sponsor named.

“If you are still typing prompts into Claude Code, waiting for it to build, testing the build yourself … you are the agent.” Sharbel gives a four-question filter for when a loop is worth building, a four-part loop anatomy, and a real build: moving a Notion card to “filmed” starts his existing YouTube-edit skill, which produces a low-resolution preview, checks it against the skill’s checklist, and waits for his approval before a 4K render. His main addition to the loop literature is training mode: the first runs stop at every step for feedback, and approval gates are removed only as the loop earns trust.

Key Takeaways

  • Four filter questions. Does the task repeat? Is there a clear definition of done? Can you afford the token cost of retries? Does the agent have the tools (browser, files, tests)? Four yeses make a loop candidate; otherwise just prompt.
  • Four parts of a loop:
    • Trigger: manual, scheduled, or event-based (a ticket, a failed test, a Notion status change).
    • Execution skill: a real procedure such as a web-design or fact-checking skill, not “make my app better”.
    • Goal plus verifier, which must live together: a goal without a verifier never knows it is done; a verifier without a goal checks forever.
    • Output plus memory: a changed codebase or task, and a log of what was tried and learned so the next run does not start cold.
  • Build loops from workflows you already do by hand. His YouTube-edit skill already downloads raw footage, fixes audio, cuts silences, removes double takes and adds animations. The loop only makes it repeatable.
  • Verifiers should be measurable. Machine-checkable: black frames, audio-video drift, words clipped mid-speech, matching the plan. Human-owned: does it feel right, did he ramble, should humour stay.
  • Training mode. Run one stops and asks at every step. By run 5 there might be two checkpoints, by run 20 the goal is one. “The autonomy increases only when the loop proves that it deserves it.”
  • Cheap preview before expensive output. Low-resolution preview first; 4K only after approval.
  • Claude Code for coding loops, Hermes for operator loops. Hermes can sit in the background and stay quiet when nothing needs him.
  • Loops not to build: vague goals (“make my videos better” is “a wish”), outputs the agent cannot verify (it will not know when to stop), expensive work with no preview, and external actions (upload, publish, email, delete, send to client) without an approval gate.

The build

  1. Prompt to Claude Code: build a Notion-triggered YouTube editing loop. Trigger: a page moves from “To film” to “Filmed”. Context: the page holds title, script and raw footage links. Execution skill: the existing YouTube-edit skill, “Do not invent a new editing style.” Loop: detect the change, collect footage, run the skill, make a low-resolution preview, run the verifier, fix and re-preview on failure, show preview plus edit summary plus verifier report on a pass, wait for approval, then render 4K. Plus stop-when, stop-and-ask-if and do-not rules, and “start building this loop in training mode step by step”.
  2. Claude Code asked where his Notion token was, built the first part and asked him to move a card. It downloaded all 11 footage files, asked more questions, then ran for 37 minutes before returning a preview.
  3. The first preview missed his on-screen animations, which are part of the skill; the second added them but showed 15 sections where the video had about 8. Both were caught in training mode.

Try It

His template (screenshot it and give it to your agent): “I want you to work in a loop with the goal [exact outcome], the trigger [what starts it], the context [files, pages, constraints, examples], the execution skill [skills used], the verifier [how output is checked], the loop [steps], training mode, stop when [condition], stop and ask me if [condition], do not [forbidden actions], memory/log.”

  1. Pick a workflow you already run by hand with a skill.
  2. Write two lists: what a machine can verify, and what you must judge.
  3. Start in training mode and remove one checkpoint only after several clean runs.