skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
~/agent-patterns / 10 field-tested cycles

work that
knows when to stop.

A loop is a contract: what repeats, what changes, and what ends the cycle. Preview the behavior, then copy the prompt into Claude Code, Codex, Cursor, or your agent of choice.

observe→act→verify↺until done ≠ forever

What is an AI agent loop?

An agent loop repeats a defined cycle of observing, acting, and verifying until an explicit condition is met. The stop condition prevents a useful iterative process from becoming an endless one.

How do you turn a loop into an agent skill?

Copy the loop's prompt, adapt its inputs and finish condition to one repeatable job, then package it in SKILL.md with a clear trigger description and outcome test. The agent skill guide covers the complete format and publishing process.

10 / 10 specimens
loop 01 · Reliable multi-step implementation

Plan / Execute / Verify

Turn a goal into a plan, perform one safe step, then verify the result before continuing.

  1. plan
  2. execute
  3. verify
stop when: Every acceptance check passes
codinggeneral
plan-execute-verify.loop● live
iteration 01PLAN
›$ inspect auth flow and select the smallest safe change
loop 02 · Debugging and regression repair

Test / Fix / Retest

Keep the failure reproducible while applying the smallest fix that makes it disappear.

  1. test
  2. fix
  3. retest
stop when: The regression test and surrounding suite pass
codingdebugging
test-fix-retest.loop● live
iteration 01TEST
›✗ expected 401, received 500 for an expired session
loop 03 · Evidence-backed research

Research / Synthesize / Critique

Gather primary evidence, combine it into an answer, then attack the weak claims.

  1. research
  2. synthesize
  3. critique
stop when: Important claims are supported or clearly qualified
researchwriting
research-synthesize-critique.loop● live
iteration 01RESEARCH
›+ 6 primary sources · 2 conflicting benchmarks
loop 04 · Design, copy, and solution quality

Generate / Evaluate / Refine

Create a candidate, score it against explicit criteria, then improve only what scored poorly.

  1. generate
  2. evaluate
  3. refine
stop when: The candidate clears the quality threshold
creativequality
generate-evaluate-refine.loop● live
iteration 01GENERATE
›+ landing headline candidate A
loop 05 · Dynamic operations and incidents

Observe / Orient / Decide / Act

Continuously update a decision from changing conditions instead of following a stale plan.

  1. observe
  2. orient
  3. decide
  4. act
stop when: The environment reaches a stable target state
operationsstrategy
ooda.loop● live
iteration 01OBSERVE
›latency p95 1.8s · error rate 7.2%
loop 06 · Editing and code review

Review / Revise

Review work against a fixed brief and revise concrete mismatches until none remain.

  1. review
  2. revise
stop when: No material mismatch with the brief remains
reviewwriting
review-revise.loop● live
iteration 01REVIEW
›! intro delays the main claim by 180 words
loop 07 · Tool-using agents

Reason / Act / Observe

Choose a tool from the current evidence, use it, and feed the real result into the next decision.

  1. reason
  2. act
  3. observe
stop when: The requested outcome is verified
agentstools
reason-act-observe.loop● live
iteration 01REASON
›need the shared caller before changing validation
loop 08 · Rate limits and flaky services

Retry / Backoff

Retry transient failures with increasing delays while preserving a hard attempt limit.

  1. attempt
  2. classify
  3. wait
stop when: The call succeeds or the retry budget is exhausted
reliabilityapi
retry-backoff.loop● live
iteration 01ATTEMPT
›! 429 rate limited · retry-after 2s
loop 09 · Deploys, spending, and external actions

Draft / Approve / Continue

Pause before consequential actions and continue only after a person approves the exact change.

  1. draft
  2. approve
  3. continue
stop when: The approved action completes or approval is denied
safetyhuman-in-loop
human-approval.loop● live
iteration 01DRAFT
›> deploy commit 8d31f2 to production · affects web
loop 10 · Parallel multi-agent work

Delegate / Integrate / Check

Split independent work, integrate the returned evidence, and check the whole result together.

  1. delegate
  2. integrate
  3. check
stop when: The integrated result passes shared checks
agentsparallel
delegate-integrate.loop● live
iteration 01DELEGATE
›+ routes audit · accessibility audit · data audit