Vibe Loop Engineering

A practical loop-engineering kit for vibe coders who use AI agents and still want clean, reviewed, shippable work.

Built by Mustafa
Instagram: @mustafaiscoding · X/Twitter: @mustafa30

The idea

Prompting an AI coding agent one task at a time works until the project gets real.

Then you need a loop.

A loop is a repeatable system that tells the agent what to watch, what context to keep, what it can change, how to verify the work, and where it must stop for you.

This repo is a builder-friendly version of that idea. Less theory. More shipping.

Star growth

The Vibe Loop

Part	What it means	Good default
Goal	the outcome the loop is trying to produce	one sentence, user-visible
State	durable memory outside the chat	STATE.md, issues, PR notes
Run	where the agent does work	branch or worktree
Verify	proof the work is real	tests, lint, screenshots, curl, review
Gate	where risky work stops	human approval before merge/deploy/delete
Learn	what gets written back	short notes, not a diary

Maturity levels

Level	Name	The agent can	The agent cannot
L1	Report only	inspect, summarize, recommend	edit files, comment, commit
L2	Assisted PRs	branch, edit safe files, run checks, open PRs	merge, deploy, touch secrets
L3	Trusted chores	run allowlisted actions with logs and rollback	improvise outside policy

Most projects should start at L1. L3 is not the flex. A calm repo is the flex.

Quickstart

Copy the loop files into your project:

git clone https://github.com/mustafa3252/vibe-loop-engineering.git
cd vibe-loop-engineering
cp templates/LOOP.md /path/to/your/project/LOOP.md
cp templates/STATE.md /path/to/your/project/STATE.md

Or install the skills with the open skills CLI:

npx skills add mustafa3252/vibe-loop-engineering --all

List the included skills:

npx skills add mustafa3252/vibe-loop-engineering --list

First loop to run

Start with a daily hygiene report. It is useful and low risk.

Use the loop-brief and state-keeper skills.
Loop: daily hygiene
Level: L1 report-only
Goal: summarize repo health and recommend the next smallest useful action.
Inputs: open issues, open PRs, failing checks, TODOs, recent commits.
Allowed actions: read only.
Never do: edit files, comment on GitHub, create branches, deploy, delete anything.
Done means: a short report with evidence and one recommended next action.

Included skills

Skill	Use it for
loop-brief	turning a vague automation idea into a safe loop brief
state-keeper	keeping STATE.md short, useful, and current
loop-auditor	checking whether a loop is safe enough to run
launch-loop	polishing docs, demos, screenshots, and launch copy before release

Loop patterns

Pattern	Start level	What it does
Daily hygiene	L1	reports repo health and next actions
PR babysitter	L1	watches PRs and explains blockers
Bug replay	L2	creates a reproduction before a fix
Launch polish	L2	cleans README, examples, screenshots, and release copy
Dependency scout	L1	reviews update risk without blind upgrades
Content loop	L1	turns shipped work into posts and devlogs

Files you copy into a real project

File	Why it exists
templates/LOOP.md	tells agents what loops are allowed
templates/STATE.md	stores tiny project memory
templates/loop-report.md	keeps reports short and comparable
examples/github-actions/daily-hygiene.yml	example scheduled L1 loop
examples/prompt.md	copy-paste prompt for your first run

Safety rules

• Start report-only.
• Put secrets, auth, billing, deploys, deletes, and migrations behind a hard human gate.
• Prefer branches and PRs over direct commits.
• Require evidence in every report.
• Kill loops that create noisy work.
• Keep state short enough that the next agent will actually read it.

Example report

# Daily hygiene report

Status: yellow

Evidence:
- CI failed on PR #42 because `npm test` timed out.
- 3 issues have no labels.
- README install command still references the old package name.

Recommended next action:
Open one docs-only PR that fixes the README install command.

Do not automate yet:
CI timeout needs a human look before retries are useful.

How this is different

This is not a giant automation framework. It is a tastefully small set of rules, templates, graphics, and skills for people building with AI agents right now.

The focus is not “more autonomy.” The focus is better loops:

• smaller scope,
• clearer state,
• better proof,
• cleaner handoffs,
• more human control where it matters.

Brand

Built with taste by Mustafa.

• Instagram: @mustafaiscoding
• X/Twitter: @mustafa30

If this helps you ship cleaner with AI, star the repo and tag me with what you built.

Attribution

This repo is an original Mustafa-branded take on agent loop workflows. It was inspired by public discussion around loop engineering and by Cobus Greyling's open-source loop-engineering project, but the content, graphics, templates, and skills here are written for our Vibe Ship workflow.

License

MIT. Use it, remix it, and ship something useful.