The Loop Engineering Playbook

Part 1: The Task Brief

Every top operator converged on the same 3-line brief. You don't need paragraphs. Three short lines beat ten vague ones.

Weak vs. strong, in your style

"Don't tell it what to do. Give it success criteria and watch it go." Andrej Karpathy, January 2026. Declarative prompts keep agents working longer and better than step-by-step orders.

The 7 operator laws (consensus of 3+ top sources)

1. Always give a "done when." A test command, an expected output, a screenshot target, a metric. Anthropic's #1 tip. Without it, "looks done" is the agent's only stop signal, and that's how you get confident garbage.
2. One unit per ask. Hashimoto: "don't draw the owl in one mega session." If the task has 4 parts, hand over part 1 or say "plan first." Kitchen-sink prompts are the #1 named anti-pattern.
3. Point, don't describe. "Do it like the Aria config" beats 3 paragraphs describing style. Name a file, a past PR, a live page, a git commit. Reference code is the strongest prompt there is.
4. Two-strike rule. Corrected the same thing twice in one session? The context is polluted. Say "fresh session" and restate the task with what you learned. Grinding a dirty context wastes your time.
5. Correction = rule, never just a fix. When the agent screws up, say "make that a rule." It goes to the corrections file permanently. Cherny does exactly this: "write it to CLAUDE.md and Claude runs forever." A correction that isn't persisted is a correction you'll repeat.
6. Plan-first when you can't describe the diff in one sentence. Say "plan first" and review the plan before any code. One-sentence diffs, just say do it.
7. Vague is fine when vague is the point. "What would you improve here?" or "research X" needs no structure. The brief format is for build tasks where correction is expensive.

For big features: say "interview me"

The official Anthropic workflow for spec-sized work: the agent interviews you (edge cases, tradeoffs, out-of-scope), writes SPEC.md, then a fresh session executes the spec. Time spent making the spec precise pays off more than time spent watching the build.

Part 2: The Loop Layer

The shift the whole industry made this year: stop prompting tasks one at a time, start designing loops that prompt the agent for you.

"I don't prompt Claude anymore. I have loops prompting Claude and figuring out what to do. My job is to write loops." Boris Cherny, creator of Claude Code. Armin Ronacher, creator of Flask, says the identical sentence.

The loop menu (what you say, what runs)

Anatomy of every real loop

The six-part anatomy, identical across Anthropic's harness design and Cherny's loop guidance:

TRIGGER -> every 15 min / overnight / on CI failure
DOER    -> the agent works ONE task
CHECKER -> a separate model or test suite grades the output
STOP    -> success signal, or N iterations, or $X spent (whichever first)
MEMORY  -> progress file read at the start of every run, updated at the end
SKILLS  -> frozen project knowledge, kept short ("a bloated rules file
           gets paid for on every beat of the loop")

The checker is the load-bearing part. The 2026 iron law: the worker never grades its own homework. A separate model, a test suite, or CI declares "done", never the agent that did the work. Verification collapse is real: an agent grading itself will delete the failing test and call it done. And every loop carries an iteration cap AND a budget cap, or it's not a loop, it's a runaway.

Copy-paste templates (work verbatim in Claude Code)

/loop [verifiable end state], only touching [scope], stop after [X iterations],
use [skill] for the work, use verifier agents at [checkpoint], and keep a
memory file of all your work.

You have a maximum of 20 attempts. If all tests pass, report "TASK_COMPLETE"
and stop. If you hit an error you cannot resolve after 3 retries, report
"TASK_FAILED: [reason]" and stop.

The 10-80-10 model split: first 10% (planning) on the smartest model, 80% grunt work on cheaper models, final 10% review back on the smartest model.

The overnight cadence (the compounding habit)

Every top operator runs the same rhythm. Hashimoto spends the last 30 minutes of each workday kicking off overnight agents.

1. End of day: queue the loops. "ralph the moldscanner backlog overnight", "loop the PR babysitter."
2. Morning: review outcomes. Merge green PRs, kill bad branches. You review outcomes, not keystrokes.
3. Every correction becomes a rule. So tomorrow's loops are smarter than today's.

Part 3: The Cheat Sheet

The whole playbook on one screen.

Part 4: What the Top Guys Actually Do

The apex tier, with their documented workflows. They all flipped to agent-first within the same window (December 2025, when model capability "crossed a threshold of coherence"), and they all converged on the same method independently.

Andrej Karpathy (OpenAI co-founder, ex-Tesla AI director, now at Anthropic)

The most credible witness in the industry because he was the biggest skeptic. October 2025: agents were "net unhelpful," he hand-built nanochat with autocomplete. January 2026: "I rapidly went from 80% manual coding in November to 80% agent coding in December... easily the biggest change in 2 decades of programming." April 2026 he named the discipline agentic engineering. His documented rules:

1. Declarative, not imperative. "Don't tell it what to do, give it success criteria and watch it go." This is the "done when" law from Part 1, stated by the top guy in the field.
2. Watch them like a hawk on important code. Agent mistakes look like "a careless, impatient, but very knowledgeable junior developer." Multiple agent windows on one side, an editor on the other purely for reviewing diffs.
3. The autonomy slider. Match the leash length to trust: small edits on a short leash, well-scoped repo work on a long one. Break work into chunks, validate sequentially.
4. The verifiability law. "Traditional software automates what you can specify. LLMs automate what you can verify." Tasks with an automatic pass/fail signal are where agents shine; everything else stays jagged.
5. Agents implement, humans generate. His 8-agent research-org experiment (Feb 2026) found agents "very good at implementing well-scoped ideas, bad at creatively generating them." Ideas, architecture, and taste stay human.
6. "You can outsource your thinking, but you can't outsource your understanding." The agentic engineer "designs specs, supervises plans, inspects diffs, writes tests, creates evaluation loops, manages permissions, and preserves quality."

Boris Cherny and the Claude Code core team (Anthropic)

1. Cherny (creator): "I don't prompt Claude anymore. I have loops prompting Claude. My job is to write loops." Brief = goal + constraints + acceptance criteria. Every correction becomes a permanent rule. 259 PRs in one month without opening an IDE.
2. Adam Wolff (engineer, ex-Robinhood Head of Eng): "90% of the code we ship to production is written by or with Claude." "When implementation cost goes to zero, the feedback loop becomes everything. The speed of learning becomes the only competitive advantage."
3. Cat Wu (founding PM): scale gradually. Never kick off a run on 100,000 items; test on 1, iterate the prompt, scale to 10, then expand. And stop stuffing prompts with examples and "don't do X" lists, current models work better without them.
4. Team stats (verified): 5 PRs per engineer per day, 60 to 100 internal releases daily, 80%+ of Anthropic engineers use it daily.

Jarred Sumner (Bun creator): the biggest loop run ever documented

May 2026: ported Bun, 535,496 lines of Zig, to ~750,000 lines of Rust in 11 days using 64 concurrent Claude instances, ~50 workflows, and 6,502 commits. The rules that made it work:

1. 2+ adversarial reviewers per implementer. "The reviewer's only job: find bugs and reasons why the code does not work." Not style, not opinions, just breakage.
2. An untouchable ground truth. 1,380,000 pre-existing test assertions, which the agents were never allowed to modify, served as the verifier. 99.8% pass at publication.
3. The human runs the org, not the code. "For most of those 11 days, I monitored workflows, manually reading outputs to check for issues." He held merge authority and audited that tests genuinely executed.

The rest of the tier, one line each

Part 5: Our Upgrade Backlog

We already run 80% of this: the ralph.sh fresh-context loop, /loop, crons, the 7-agent team, and the corrections-to-rules pipeline. The research surfaced 5 upgrades, in value order.

1. Independent judge on loop completion. Our loops currently trust the worker's own "complete" signal. Add a separate cheap-model judge pass before accepting done.
2. Budget caps. Iteration caps exist; dollar and token caps don't yet.
3. PR-gated loop variant. Each iteration opens a PR, blocks on CI green, merges, continues. Fits our always-use-PRs law perfectly.
4. Sprint contracts. Done-criteria written to a contract file and agreed BEFORE the loop starts, so the goal can't drift mid-run.
5. Test ratchet. Loops may never weaken checks. Verification only tightens.