AI Workflow Retrospective Skill for ChatGPT: Propose Small Verified Changes From Failures
Summarize supplied errors, corrections, and evaluation regressions. Capture triggers, impact, minimal proposed changes, and validation while preserving approval for durable guidance.
中文Install in ChatGPT
Requires a ChatGPT workspace where you can create skills. This is a text adaptation of the pinned GitHub source.
- Review the inputs and outputs in the pinned saynext_codex_skills/.agents/skills/saynext-retrospective-learning/SKILL.md.
- In ChatGPT, open Plugins, Skills, Create, and Create with editor.
- Write the skill rules using the workflow described here. Provide failure examples, corrections, existing guidance, and evaluable inputs.
- Save the skill and check the fictional example in a new chat.
Example task
A fictional text tool three times treats unknown terms as facts. Existing tests only cover typos. Propose a change and two acceptance cases.What to check
Record the failure and trigger. Suggest handling unknown terms with an eval that preserves confirmed wording and flags unresolved terms.
Original SKILL.md on GitHub · Updated
Frequently asked question
Can a one-time preference become a global rule?
Check scope and recurrence. Propose a local rule or test before promoting it to enduring guidance after review.