Loopops protocol

LoopOps Protocol

Use this skill when reusable agent behavior should become more than a better prompt.

The goal is to choose the smallest durable artifact that improves future runs:

prompt -> skill -> checklist -> script -> workflow -> loop

Do not promote work upward unless the extra machinery buys reliability, reuse, or cheaper future execution.

Artifact Ladder

Classify the candidate method first:

• Prompt: one-off instruction for this run.
• Skill: compact reusable context, procedure, or tool adapter loaded on demand.
• Checklist: ordered human/agent validation sequence.
• Script: deterministic repeatable action with parameters.
• Workflow: ordered tool/process execution with gates and evidence.
• Loop: workflow plus evaluator, memory write, retry policy, and stop rule.

Prefer the lowest artifact that handles the task reliably.

Loop Suitability Gate

Promote a workflow into a loop only when all are true:

• The goal has an observable completion condition.
• Progress can be evaluated without trusting the worker agent blindly.
• A retry is likely to improve the result.
• Cost can be bounded by iteration, time, token, or tool budget.
• Failure has a clear stop, rollback, or human-escalation path.

If any condition is missing, keep it as a supervised workflow or checklist.

Open vs Closed Loop

Use an open loop when:

• the solution space is unknown,
• discovery or exploration matters,
• multiple valid approaches may exist,
• the output is research, architecture, strategy, or design direction.

Use a closed loop when:

• acceptance criteria are known,
• tests, checks, or review criteria exist,
• regression risk matters,
• the task can be decomposed into verifiable steps.

Open loops need tighter budget limits. Closed loops need stronger evaluators.

Supervisor Contract

Every loop must define:

Goal:
Completion condition:
Loop type: open | closed
Worker action:
Evaluator:
Evidence:
Memory write:
Retry policy:
Budget:
Stop rule:
Rollback or escalation:

The evaluator can be a test command, lint/typecheck/build, browser check, static analysis, script, human review, or a separate model. Prefer deterministic evaluators when available.

Memory Policy

After each iteration, save only durable signal:

• current hypothesis,
• changed files or artifacts,
• verification result,
• blocker or failure mode,
• next intended action,
• reusable lesson if it generalizes.

Do not append raw logs, long transcripts, secrets, private data, or one-off observations to durable memory.

Skill Evolution

After a loop completes, update skills only with lessons that are:

• procedural,
• reusable,
• testable,
• not already covered by higher-priority instructions,
• worth their token cost.

Good updates:

• missing repo-specific command,
• repeated failure mode,
• verification step that caught a real issue,
• script or workflow that replaced repeated manual work,
• rejected path that should not be retried.

Bad updates:

• generic coding advice,
• motivational wording,
• model-behavior hacks,
• single-incident preferences,
• rules that increase context without reducing errors.

Cost Controls

Every loop must have at least two:

• max iterations,
• max wall-clock time,
• max token/tool budget if measurable,
• narrowest useful verification first,
• summarized memory instead of raw logs,
• stop after repeated identical blocker,
• human approval before external, destructive, expensive, or privacy-sensitive actions.

If cost cannot be bounded, do not run autonomously.

Compiler Rule

When the same step is repeated three times, decide whether to compile it:

• repeated prose instruction -> skill rule,
• repeated checklist -> workflow,
• repeated shell/Python/JS snippet -> script,
• repeated verification pattern -> evaluator,
• repeated failure recovery -> stop hook or escalation rule.

Prefer scripts and deterministic checks over longer instructions.

Final Report

When designing or updating a loop-capable method, report:

• chosen artifact level,
• why higher levels were rejected or accepted,
• supervisor contract if a loop is used,
• cost controls,
• memory writes,
• validation performed,
• skill/workflow/script files changed.