Level 7Lesson 0โฑ๏ธ 120 min

Level 7 Capstone

Design your first real loop end-to-end - with an interactive helper that builds your loop spec and STATE.md as you go

Stop Prompting. Start Designing.

Across nine lessons you learned the whole stack: when a loop is worth building, the five building blocks, the memory file, the minimum viable loop, and every way loops fail quietly. This capstone turns that knowledge into one concrete artifact - a real, build-ready design for a loop you could run.

You do not need to write any code to pass this capstone. The goal is to designa loop correctly - because in loop engineering, the design is the engineering.

Your Mission

Take the candidate task you have been carrying since Lesson 56 and produce a complete loop design for it. A passing submission has five parts:

1. THE TEST     - show your task passes the 4-condition test
                  (and name the gate that proves "done")

2. THE SKILL    - a one-page SKILL.md: how to think, where to
                  look, and a "never do" list

3. THE LOOP     - the Minimum Viable Loop: automation + skill +
                  state file + gate, with the build order

4. THE SAFETY   - your red-team notes: how it could fail quietly
                  (Ralph Wiggum, drift, leaks) and the defense for each

5. THE METRIC   - how you'll measure Cost per Accepted Change,
                  and your stop-the-loop threshold (e.g. <50% accepted)

๐Ÿ† Capstone Project Helper

Use the interactive helper below to build parts 1, 3, and the spine of your loop. Tick the 4-condition test, pick a starter pattern (or go custom), fill in the fields, and it generates a ready-to-use loop-spec.md and STATE.md you can copy straight into your project. Everything updates live as you type.

๐Ÿ› ๏ธ Loop Design Helper

Fill this in to design your first loop. It generates a ready-to-use loop spec and STATE.md you can copy out.

Step 1 ยท The 4-condition test

A loop only makes sense if all four are true. Be honest.

0/4 conditions met - until all four are true, keep this task as a manual prompt. The missing piece is your real first project.

Step 2 ยท Describe your loop

Fill in all four conditions and every field above to unlock your build-ready verdict.

Step 3 ยท Your generated files

loop-spec.md
# Loop Spec - CI failure triage

## Trigger (automation / heartbeat)
Every morning at 7:00am

## Skill
Use a SKILL.md for "CI failure triage":
  - how to think about the task
  - where to look / fix patterns
  - a "never do" list of hard rules

## Maker / Checker (separate agents)
- Maker: drafts the change in an isolated git worktree
- Checker: a SEPARATE agent verifies it against the gate below

## Gate (objective check - "done" means this is true)
All tests pass and the linter is clean

## Stop condition (/goal)
Every failing test in the target folder is either fixed (tests pass) or escalated to a human

## Hard cap (safety)
Max 20 minutes or 200k tokens per run

## Human gate
A human reviews the diff and approves before anything merges or ships.

## Blocked zones (never touch)
Architecture, authentication / cryptography, payments / billing.
STATE.md
# STATE - CI failure triage

Last run: (not run yet)

## Done
- (nothing yet)

## In progress
- (nothing yet)

## Lessons learned
- (record gotchas here so future runs don't repeat them)
How to use the output: The generated loop-spec.md is your design on a page - hand it to a teammate and they'd understand your loop in one read. The STATE.md is the starting memory file you drop in your repo root. Together they're two of the five required parts, done.

Finish the Other Three Parts

Part 2 - The Skill (from Lesson 59)

Expand the SKILL.md you drafted: how to think about the task, where to look / fix patterns, and a "never do" list of at least three hard rules.

Part 4 - The Safety / red-team (from Lesson 64)

Write one way your loop could declare "done" while failing, one rule that could drift, and one permission risk - each with its single defense.

Part 5 - The Metric (from Lesson 63)

State how you'll track Cost per Accepted Change and the acceptance rate at which you'd turn the loop off. If you can't measure acceptance, say how you'll start.

What "Great" Looks Like

A strong submission:
  • Picks a genuinely small, repetitive, machine-checkable task - not "rewrite our app."
  • Has a gate that is an objective check (a real test/build), not "an agent reviews it."
  • Separates maker from checker, and has a hard cap plus a human approval before merge.
  • Is honest about where it is in the build order - most good answers are still at "make it work manually."
  • Treats the loop as a multiplier on the engineer's judgment, not a replacement for it.

What You Have Learned in Level 7

  • L56 - From prompter to loop designer; agent = model + scaffolding + loop
  • L57 - The 4-condition test; when NOT to build a loop
  • L58 - Automations (the heartbeat) + worktrees (parallel without chaos)
  • L59 - Skills: write project knowledge down once
  • L60 - Connectors: let the loop touch your real tools
  • L61 - Sub-agents: separate the maker from the checker
  • L62 - The state file: memory that outlives the chat
  • L63 - The Minimum Viable Loop and Cost per Accepted Change
  • L64 - When loops go wrong: quiet failures, comprehension debt, security
The leverage moved, and so did your job. You can now look at a piece of repetitive engineering work and decide - honestly - whether it should be a loop, then design that loop so it checks itself, remembers its progress, and fails loudly instead of quietly. That is loop engineering. The tools will keep changing; the judgment you built here will not.

Build the loop. Stay the engineer.

๐Ÿง  Check your understanding

5 quick questions on Level 7. Answer each, then check your score.

1. In loop engineering, a "loop" is...
2. Before building a loop you should ask...
3. The "checker" sub-agent exists to...
4. A state file gives a loop...
5. The Minimum Viable Loop principle says...
๐Ÿ’ฌ

How was Level 7?

Your feedback helps improve this course. Takes 30 seconds.

Overall rating *

Select a star rating to submit.

โ† When Loops Go Wrong