Video summary

Prompting is Dead. Use Loops Instead. (full guide)

Main summary

Key takeaways

Technology

Overview

This video argues that “prompt engineering” is becoming less important as newer large language models (LLMs) get better at autonomous, self-correcting workflows. Instead of repeatedly prompting, waiting, and providing feedback (where the human is the “middleman”), the recommended approach is loop engineering: let the model run repeated attempts, inspect its own output, and continue until a defined outcome is reached.

Core technological concepts / changes

  • Context / memory limitations

    • Traditional chat prompting struggles because LLMs have limited context windows and may effectively “forget” between sessions, causing errors and lower output quality.

  • Newer model behavior

    • Newer generations (mentioned: Claude “Fable 5” and ChatGPT “5.6 Sonnet”) are described as better at:
      • internal strategic thinking
      • running unsupervised loops
      • self-evaluation and self-correction

  • Shift in workflow

    • The speaker claims you can move from a turn-based human-in-the-loop workflow to being “outside the loop,” where the AI does work more autonomously.

Product / feature framing (Claude Code–style)

The guide centers on Claude Code–style functionality (and notes similar capabilities in ChatGPT), using commands/constructs such as:

  • /goal

    • Defines the mission and completion criteria (deliverables + what “done” means).

  • /loop

    • Sets the repeat interval/frequency for the work (e.g., every minute/hour/day).

  • Scheduling

    • /schedule can run tasks at specific times (example: every day at 6:00 a.m.).

  • Stopping / canceling

    • /loop cancel stops scheduled/recurring looping.
    • Also mentions a “Stop the loop” instruction.

  • Budget controls

    • Loops should include limits on iterations or token spend to prevent infinite-loop runaway.

Two types of loops

  1. Inner loop

    • Used for a single task where the model internally repeats/repairs work until quality is reached.
    • Example: “review my website,” where the model tries actions, checks its own work, fixes errors, retries, and then marks the job done.

  2. Outer loop

    • A recurring job that runs indefinitely or on a schedule.
    • Example use cases mentioned:
      • scanning competitor YouTube channels and maintaining an up-to-date log
      • daily market research to find posts that could affect a portfolio

Four-piece loop design framework (hands-off workflow)

To make loops reliable, the speaker proposes loops should include:

  1. Spec / goal What to build.

  2. Checklist What “done” means.

  3. Inspector An orchestrating “site manager” model that grades/approves output (configurable to inspect in the way you want).

  4. Budget / termination conditions Stop after N tries or after a token cost limit.

Live demonstrations / concrete examples

1) Website conversion audit (inner loop + /goal)

  • Goal: audit milesdeutch.com.au “like a conversion consultant.”

  • Output constraints include:

    • write findings to roast.md
    • list at least 10 specific issues
    • each issue includes:
      • severity
      • exact location on the page
      • a fix to ship today
    • top 3 issues ranked by revenue impact
    • issues must reference something actually on the page (not generic advice)
    • stop after 12 items/terms (as stated in subtitles)

  • How it works: The model reads HTML (not just visuals), produces actionable markdown, and the speaker can then ask it to “ship” changes (which requires connecting deployment/hosting tools like Vercel and a domain provider).

2) YouTube competitor scanning with a scheduled outer loop

  • Sets a 1-minute loop to fetch latest uploads from a channel (shown: “AI Edge”).
  • Creates/updates a document/template (niche MD) containing details like:
    • feed videos
    • audience
    • topics
  • The loop continues until the defined iteration count is reached (example shows “counter out of 8”).

Use-case menu / “loop ideas”

The speaker provides looping examples across business and personal life:

  • Content creators

    • daily ideation from performance/data
    • generating title options
    • competitor analysis

  • Marketing

    • automated scraping of ads/content that performs well

  • Business operations

    • agents in tools like Slack that summarize updates and compile blockers

  • Personal life automation

    • loop over downloaded files (e.g., bank statement CSVs) to clean up and categorize them

  • Agentic workflow framing

    • Loops are described as AI agents embedded in Claude Code
    • The speaker emphasizes they’re not separate “magic,” but more like digital employees that follow onboarding/context.

“Agents” analogy and onboarding requirement

Agents are compared to employees, which need:

  • clear onboarding (context/information)
  • defined tasks via /goal
  • a cadence via /loop / scheduling

Treating agents like real staff (with proper onboarding and constraints) is presented as key to achieving better results.

Main speakers / sources

  • Primary speaker: the video creator / “AI Edge” presenter (not named in subtitles)
  • Referenced models:
    • Anthropic Claude (Fable 5)
    • OpenAI ChatGPT (5.6 Sonnet)
  • Platform referenced: Claude Code (and similar /goal / /loop ideas in ChatGPT)
  • Referenced creator/company: “experts/builders from Anthropic” (cited as a source for the “loops are best” message)

Original video