Make Your Own Loop — the A–F Method

The 14 steps taught you the organs. This page is the surgery manual: six moves, A through F, that turn "I keep doing this by hand" into a loop you can actually trust.

When to reach for this page

The moment you notice you're doing the same agent-assisted task for the third time. Three is the magic number — once is just work, twice is coincidence, and three times is a loop you haven't built yet.

The method

A · Pick the task and its shape

State the task in one sentence, then let the way it ends choose the heartbeat for you:

• The work ends (a list empties, a suite goes green) → conditional run-until-done.
• The work repeats (every morning, every PR) → schedule or event.
• The work happens once → no loop. Just do it. (The most common mistake in the book: looping a one-off because loops are fun to build.)

B · Write the stopping condition as a spec

Before you touch a prompt: one sentence a machine can verify. "Every box in state.md is checked." "npm test exits 0." If you can't write that sentence, you don't have a loop task yet — you have a wish. Keep sharpening the task until the stop practically writes itself.

C · Assemble the six parts

Fill in the table — every row, even when the honest answer is small:

D · Add the three stops and the guardrails

Success (straight from B) · limit (max runs — pick a number that would embarrass you if it were ever hit) · no-progress (3 unchanged beats → stop and log). Then bolt on the guardrails: budget caps plus the 80% tripwire, a kill switch, and one log line per beat.

E · Prove it, then let go — one level at a time

L1 report-only for real, watched runs → L2 assisted (writes, human reads every output) → L3 unattended (writes, human samples). One promotion per proven level, and automatic demotion on any incident. Never skip a rung because the streak was looking good.

F · Improve the loop, not just the work

When a beat lets you down, fix the skill, the rubric, or the spec — never the individual output. Output fixes evaporate by morning; loop fixes compound. (This is the seed of hill-climbing: see Step 12.)

Worked in 90 seconds

Task: "changelog entries pile up unwritten." A: it repeats per merge → event loop. B: "every merged PR since the last beat has a changelog line." C: body = CHANGELOG.md only; spine = last-processed PR number; checker = a script (does every merged PR number show up?); gate = the release manager reads it before tagging. D: limit 10/day; no-progress = 3 beats stuck on an unprocessable PR → escalate. E: one week of L1 drafts-as-comments before it's ever allowed to commit. F: the entries came out too terse → fix the skill's template, not Tuesday's individual entry.

Use it against real paper

• The method applied to a full project, decision by decision: worked-example-dependency-sweeper
• Blank six-part table + checklist: loop-design-checklist
• A ready-made kit to copy and fill: scaffold-from-template
• Which heartbeat/pattern fits: pattern-picker
• Should this be a loop at all: decision-framework

Course context: this method is the bridge from Part 5's build to loops of your own design — and the capstone assessment asks you to walk A–F cold.

Sources: the A–F method is original to this course, synthesizing Panaversity's Loop Engineering: A Crash Course (S1), stopping-condition-as-spec from Spec-Driven Development (S3), and the L1-report-only-first rule of the cobusgreyling/loop-engineering reference repo (S7, MIT). Full attribution: resources/sources.md.