AgenticAI is a skill I built for one specific feeling: you have a goal, you don't want to babysit the AI through it, and you don't trust a one-shot answer that says "done" with no proof. You type /agenticai <goal> and it stops being a single assistant — it runs a whole virtual company around that goal until it's actually delivered or genuinely blocked on something only you can decide.
How it actually works
It's a loop with roles. A Project Manager turns your goal into a Definition of Done and breaks it into milestones. A Developer builds each one. Then the part that makes it worth running — an adversarial QA Griller whose entire stance is that everything the Developer claims is a lie until proven. It runs the build, hits the endpoints, feeds hostile inputs, re-runs the old tests, and writes a PASS/FAIL verdict with evidence. "The code looks correct" is a FAIL by definition. Specialists (security, SEO, docs, release) only get pulled in when the goal actually needs them.
The pros
Evidence over vibes. Nothing advances on "looks good" — only on executed output. That's the whole reason it exists: it kills the confident-but-wrong result that normal AI coding ships by default.
It survives crashes. State lives in an Obsidian vault, not the chat, so a context compaction or a closed laptop doesn't lose the run — it resumes from state.md and even carries lessons forward between runs.
It doesn't pay Opus to run `ls`. The PM assigns cheap models to mechanical chores and the strong model to design, security, and the final grill. It's model-agnostic too — the discipline holds whether you run it on Fable, Sonnet, or Opus.
The cons
It's slow and token-hungry by design. Grill cycles and parallel agents cost real money. For "fix this typo" it's absurd overkill — the skill itself refuses to run on ordinary one-off requests, and that's deliberate.
It can hand you a documented failure. There are hard caps — 5 grill cycles and 2 design rethinks per milestone. Hit them and it ships a "Partial Delivery" with honest limitations instead of a finished feature. That's the honest move, but you still finish the last mile yourself.
It's an autonomous loop with repo access. It branches, commits, and can install other skills from a registry (which it treats as untrusted input and reads before trusting). Still — that's a lot to hand a machine. Run it where a bad commit is revertable.
Grab it
Here's the entry file. One honest caveat: the full skill also ships companion references/ files — the mindset rules, the role charters, and a machine toolbox — that do a lot of the heavy lifting. This SKILL.md is the orchestration spec that ties them together, so it's the core, not the entire folder.
Skill
agenticai