Safe Long-Run Mode (GPT-5.4 + Claude)
name: safe-long-run-mode-gpt54-claude
by bwiley1989 · published 2026-04-01
$ claw add gh:bwiley1989/bwiley1989-safe-long-run-mode-gpt54-claude---
name: safe-long-run-mode-gpt54-claude
description: Operate long-running AI tasks safely across GPT-5.4 and Claude by using model selection rules, phased execution, checkpoints, resumable workflows, API throttling discipline, and subagent isolation. Use when a task may run for a while, touch multiple files/systems, involve external APIs, browser automation, Azure, Orgo, or multiple subagents, or when the user asks about long autonomous runs, rate limits, reliability, or safe operating mode.
---
# Safe Long-Run Mode (GPT-5.4 + Claude)
Use this skill for tasks that may run long, span multiple systems, or risk losing progress if interrupted.
Core rule
Do not run long tasks as one monolithic attempt. Split into phases, write checkpoints, and keep the work resumable.
Model selection
Use **GPT-5.4** for:
Use **Claude** for:
Default to **GPT-5.4 first**. Escalate to Claude only when the task actually benefits from higher-quality judgment or tone.
Operating procedure
1. Scope before acting
Before starting, decide:
2. Break work into phases
Use phases such as:
1. gather / inspect
2. plan / write brief
3. execute / edit / build
4. validate
5. deploy or report
At the end of each phase, write artifacts to disk.
3. Always checkpoint
For long tasks, save progress in files:
Prefer a resumable workspace state over a perfect one-shot run.
4. Isolate long work
Use subagents when:
5. Throttle external systems
When interacting with Azure, Graph, Orgo, messaging providers, registries, websites, or any external API:
6. Prefer resumability over perfection
The goal is not "never fail." The goal is: if interrupted, resume with minimal loss.
System-specific guidance
Azure / cloud control planes
Browser / Orgo / GUI automation
Coding / documentation work
What to tell the user
When relevant, explain that safe long-run mode means:
Failure handling
If a long task is interrupted:
1. summarize completed phases
2. point to saved artifacts
3. identify exact next step
4. resume from checkpoint rather than restarting
References
More tools from the same signal band
Order food/drinks (点餐) on an Android device paired as an OpenClaw node. Uses in-app menu and cart; add goods, view cart, submit order (demo, no real payment).
Sign plugins, rotate agent credentials without losing identity, and publicly attest to plugin behavior with verifiable claims and authenticated transfers.
The philosophical layer for AI agents. Maps behavior to Spinoza's 48 affects, calculates persistence scores, and generates geometric self-reports. Give your...