These 3 skills changed my life

Summary

Theo explains how three Markdown-based AI-agent skills—Soul Loop, Full Send, and Full Stop—grew out of repeated coding habits and now shape much of his development workflow. His broader lesson is to identify what you trust agents to do, encode those expectations in reusable instructions, and build verification and recovery safeguards around their work.

The three skills

• Soul Loop — iterative code review: Uses an external AI reviewer to inspect changes in a bounded loop before or sometimes after a pull request is opened. The goal is to make the smallest changes needed to address valid findings while resisting scope creep and pushing back on minor or irrelevant suggestions. The skill limits review rounds, avoids nested reviewers, tracks findings, and can record the reviewer’s verdict on the PR. Theo says this helps catch issues before other review tools or teammates see the change, reducing review cycles and preventing unnecessary code from reaching a PR.

• Full Send — autonomous work through merge: Gives the agent responsibility for completing a task, verifying it, opening and monitoring a PR, addressing CI and review feedback, and merging once explicit requirements are met. These requirements include evidence that the change works, passing checks, resolved or documented review findings, and a mergeable, up-to-date branch. The agent should use an appropriate question tool only when genuinely blocked, since ordinary messages may go unnoticed while a thread is unattended. Theo reports that 27 of 32 recent Full Send runs merged without requiring him to inspect them again. He emphasizes that this depends on knowing which changes are safe to delegate and having safeguards to catch failures.

• Full Stop — autonomous work, but no merge: Follows much of the same verification and PR-maintenance process as Full Send, but stops when the PR is ready for Theo’s approval. He uses it for riskier changes where he wants the agent to do the work and resolve feedback without merging independently.

Workflow and tools

• Theo keeps skills and machine setup in a “fleet” repository, making it easier to synchronize skills and configuration across computers.
• In T3 Code, he treats the sidebar as an inbox: active work should stay out of the way while it is progressing, and completed or abandoned threads should be settled. The aim is to focus attention on work that needs a person.
• T3 Code’s Orchestrator V2 can delegate reviews to other agents using their native harnesses. Theo sees value in cross-model review because models have different tendencies: he finds OpenAI models thorough, Claude capable of catching unusual risks, and Gemini more prone to hallucination.
• A CLI proxy can centralize model access and usage monitoring across tools. Theo cautions that providers’ subscription terms differ: he says to keep Claude subscriptions within Claude Code, while OpenAI models can be used in other tools.
• Skills should have descriptions that act as activation triggers, rather than simply explaining what the skill does. Theo prefers trigger phrases over manually invoking slash commands, partly because some coding harnesses make it difficult to invoke multiple skills together.

Other reusable skills

Theo also describes skills for taking over existing PRs while preserving contributor credit, modernizing outdated PRs by rebasing and checking whether the work is still needed, and managing PRs.

He recommends treating these files as editable workflow instructions: update them when an agent repeatedly makes a mistake or behaves in an unwanted way.

Sponsor segment

The video’s sponsor, WorkOS, is presented as an authentication service for handling enterprise requirements, such as custom SSO setups through an admin portal, as well as MCP authentication. Theo also describes AuthKit and AuthMD, a standard developed with other companies for agents to sign up for services on a user’s behalf.

Other announcements

Theo mentions a future T-Stack project for sharing his skills, and Mega Dev, a course about using AI to ship software in real company workflows.

His closing point is not to copy his exact Markdown files, but to build a workflow around what agents can reliably do, how their work can be verified, and how mistakes can be contained.

Main speaker/source: Theo of the T3.gg channel. Tools and services discussed: T3 Code, Claude Code, OpenAI/Codex, Gemini, and sponsor WorkOS.