agentic-loop

your prompt one-shots. an agent loops. that's the entire difference.

this repo is one prompt that turns a normal request into an agentic loop — no framework, no orchestration layer, just a prompt shaped like a loop.

the anatomy

every agentic loop prompt has six parts:

1. role — a job, not a persona ("you are my writing loop")
2. state — the variables the loop runs on (topic, audience, style, goal)
3. ordered steps with a gate — "do not produce the final output immediately" blocks the one-shot reflex
4. a critic with a named checklist — vague critics rubber-stamp; named failure modes catch things
5. a rewrite rule — every flagged issue gets fixed or defended, no silent skips
6. a stopping condition + handoff — end with the deliverable, the weakest remaining part, and what to test next run

the prompts

• GENERATOR.md — a meta-prompt that builds a new loop for any job you describe. the generator is itself a loop.
• .claude/skills/loopgen — the generator packaged as a Claude Code skill. type /loopgen <your task> and it runs the five steps and hands back a ready-to-paste loop.

swap the steps and the critic's checklist to loop anything — code review, landing page copy, cold emails. or let the generator do the swapping.

the skill

if you use Claude Code, you don't have to paste the meta-prompt by hand:

/loopgen reviewing pull requests on a typescript codebase

it derives the domain-specific failure modes, designs the divergence step, builds the six-part loop, stress-tests it, and delivers the loop plus an example fill plus the one failure mode most likely to slip through.

the shiporskip marketplace

this repo is also the shiporskip Claude Code plugin marketplace — a single home for the tools i ship. add it once:

/plugin marketplace add scottshapiro142/shiporskip-plugins

(you add the repo shiporskip-plugins, and install from the marketplace named shiporskip.) then install any of the tools:

tool	what it does	install
loopgen	turn a task description into a ready-to-paste agentic loop	/plugin install loopgen@shiporskip
roast	adversarial council that stress-tests an idea → GREEN LIGHT / RESHAPE / KILL + the cheapest 48-hour test	/plugin install roast@shiporskip
flip-check	should you buy this secondhand item to resell? BUY / WALK + max price (ships an eBay comps script)	/plugin install flip-check@shiporskip

once installed, each tool's slash command (/loopgen, /roast, /flip-check) works in every project.

or clone and use loopgen here — it also lives at .claude/skills/loopgen/, so running Claude Code inside a clone of this repo exposes /loopgen with no install step.

maintainer notes:

• each tool is a plugin under plugins/<name>/. edit the copy there, then bump version in its plugins/<name>/.claude-plugin/plugin.json so installed users pull the change on /plugin marketplace update.
• loopgen also has a project-skill copy at .claude/skills/loopgen/ (for clone-and-run). keep it identical to plugins/loopgen/skills/loopgen/SKILL.md when editing.

how to use it

1. fill in the four state variables
2. paste into your model of choice
3. don't skip to the final output — the option-scoring and the critique pass are where the value is
4. the critic will attack the model's own work. it is usually right.

why this works

a one-shot prompt gives you the average of everything the model has seen on that topic. the loop forces divergence (multiple options), judgment (scoring), and revision (critique → rewrite) — the three things a competent human does instinctively and a one-shot prompt skips entirely.

built by

@ScottShapiroUXD — building shiporskip.io and nexuscli.dev in public.

if this saved you a bad draft, the follow button is right there.