A 30-minute video is too long for one conversation. This guide walks the 8-step Skill that locks the narration and one look, then gives each 3-minute segment to a fresh sub-agent, five at a time, and polishes the cut in an empty-context session.
The folder structure behind my AI agent team, the Herdr setup where Grok Build, Claude Code and Kimi Code work side by side, and the two live tasks from the video: an article published by voice and a three-model front-end test.
Eight folders of plain Markdown, a CLAUDE.md index in each, and seven prompts that make your agent interview you. Every prompt, template and command from the video, in order.
Subagents run well-defined tasks in isolated contexts so the main session stays clean. One dispatch rule: high intermediate noise, short final output. Start read-only, grant the fewest tools possible, run three real tasks. Copy-paste config plus the hidden costs in workflow automation software.
Plugins bundle skills, commands, subagents, hooks and outside connections into one installable unit - 30 seconds to install, which is exactly why the security check matters. Browse the official directory, verify a plugin loaded, judge its reach, and package your own workflow automation software.
160,000+ Skills exist, so availability stopped being the constraint and selection took over. A SKILL.md that searches registries, the web, and GitHub in parallel, deduplicates, and ranks every candidate across six dimensions with install commands attached.
A prompt you retype is not a habit, it is an undocumented procedure. Saving it changes who is allowed to run it — which is the part that matters.
A written instruction depends on the model remembering; a checkpoint does not. Eight lifecycle moments where your script fires automatically, where the config lives, how to write one that blocks dangerous commands, why three is the ceiling, and how to debug workflow automation software hooks.
Permission prompts are not friction - they are what makes file-editing safe for real projects. Six modes compared, allowlist and denylist rules, the five-layer config priority chain, the OS sandbox as a second defence, and the one setting business process automation should never use.
The trigger words most guides still teach stopped doing anything. Here is the live mechanism: effort levels as the primary control, ultrathink as a single-turn boost, a decision framework by task difficulty, and the point where deeper business process automation stops paying.
More tokens buy less accuracy past a point, and the fix is subtraction rather than capacity. Compaction, sub-agents, external memory, and the rule for when starting a fresh session beats carrying everything forward. Built around how an AI agent for business actually degrades.
Empty or stuffed with boilerplate - both end the same way. Weeks spent auditing real files on GitHub, from 185K-star behavioural principles to a maintainer's commit rules. Four-layer scope, five teardowns, six templates by role, and an anti-pattern audit for business process automation.
A complete video editing web app, built by someone who never wrote a line of code by hand. Install through cloud deploy, memory architecture, boundary-first development, hooks, Docker builds, and five principles that transfer - where business process automation builds rather than connects.
Installed is not commissioned. Six capabilities ship switched off, and the gap between paying for them and using them is where most of the value quietly sits.
Most guides install it and stop, or dump every feature at once. Neither says what to learn first or why the order matters. Thirteen knowledge points in tiers with real dependencies, three paths by role, six milestones, and the four that are load-bearing in any AI assistant for business.
When I figure out how to get something real done with AI, you get the write-up — every step, including the parts that didn't work.