safe-long-run-mode-gpt54-claude（safe long RUN mode gpt54 Claude 搜索）

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:

• coding
• docs
• research
• file-heavy transformations
• multi-agent delegated work
• repetitive build tasks
• long internal work where cost and throughput matter

Use Claude for:

• strategic judgment
• sensitive decisions
• nuanced synthesis
• client-facing polish
• brand voice refinement
• high-trust orchestration

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:

• what the final deliverable is
• which systems/tools will be touched
• what can fail or throttle
• what must be saved after each phase

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:

• draft outputs
• notes
• reports
• partial results
• tracker entries
• checkpoint summaries

Prefer a resumable workspace state over a perfect one-shot run.

4. Isolate long work

Use subagents when:

• the task will take more than a few tool calls
• multiple files/systems are involved
• external APIs are involved
• failure should not pollute the main session
• specialized work can be delegated cleanly

5. Throttle external systems

When interacting with Azure, Graph, Orgo, messaging providers, registries, websites, or any external API:

• batch reads when possible
• avoid tight polling loops
• serialize risky writes
• respect retry/backoff
• avoid one-item burst loops when a bulk operation is possible

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

• validate auth first
• create foundational resources first
• verify after each layer
• log resource names/IDs
• do not chain long destructive commands blindly

Browser / Orgo / GUI automation

• use explicit goals and stop conditions
• capture screenshots at checkpoints
• bound retry counts
• save artifacts locally
• prefer API/CLI over GUI when equivalent exists

Coding / documentation work

• create a brief/spec first for complex tasks
• write files in chunks
• validate after each major change
• leave notes for resume if work is unfinished

What to tell the user

When relevant, explain that safe long-run mode means:

• cheapest adequate model
• phased execution
• saved checkpoints
• subagent isolation
• controlled API usage
• resumable progress

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

• Read references/checklist.md for a reusable pre-flight checklist and model routing matrix.