v1.2.0 — Let the AI handle recovery and finish authorized work
Release date: 2026-10-04
In one sentence: Give the AI responsibility for technical choices, troubleshooting, and delivery, with fewer repeated confirmations and unnecessary steps while retaining scope and safety safeguards.
1. What this version does for you
You state your needs, use the result, and give feedback. The AI handles research, methods, operations, checks, and repairs. This version makes that division explicit so everyday work can continue more smoothly.
- Less repeated confirmation: authorization remains valid while goals, scope, and impact stay unchanged. Necessary execution outside the sandbox uses platform-permitted channels. If platform approval is needed, the AI requests it formally without asking for the same permission again in chat.
- The AI tackles limits first: when a tool fails or the environment differs, it checks the cause and possible partial writes, then finds an authorized, safe remedy. Only necessary conditions it cannot supply come back to you, with the smallest action needed.
- Effort goes into results: KISS, YAGNI, and fulfilling your requested outcome with the least rework are explicit. Reuse valid plans and evidence; extra files or steps alone do not add ceremony. Deliver once agreed requirements are met.
Safety checks follow the current action’s actual risk. Public release, spending, and access changes still need appropriate authorization. Reconstructable temporary artifacts created by the task may be cleaned up only under narrow conditions.
2. Should you update?
- Recommended if you already use these instructions and encounter repeated approval requests, agents stopping at tool failures, or simple work turning into lengthy process.
- You can wait if your current workflow runs smoothly and you are not ready to replace lasting instructions. The previous release remains available.
3. How to use it
- Open the project or workspace where you want to use it.
- Choose the English complete instruction or Traditional Chinese complete instruction, and replace the previous version in full. Do not paste both together.
- Save, start a new task, and try one small real job to check that the AI reads relevant material, completes necessary changes, and verifies the result.
4. Known limits
- Instructions do not replace tool permissions or safety mechanisms. Platform approval requirements still apply.
- Checks cover bilingual meaning, focused text scenarios, and the guides. Real permission failures, long tasks, cross-model behavior, and results across all platforms still require confirmation in actual use; there is no universal performance guarantee.
- The cleanup exception excludes user or other-agent files, deliverables, and required evidence.