Codex Self-Improving Prompt: Let AI Turn Repetitive Work Into Automation On Its Own

Original by winkrun | AI Engineering May 25, 2026 · 15:28 · Beijing

Recently, developer Vaibhav (VB) Srivastav released an updated Codex self-improving prompt that quickly sparked widespread discussion across the AI developer community — and it even caught a retweet from OpenAI co-founder and CTO Greg Brockman. His self-improvement framework is incredibly insightful, turning Codex into a self-evolving AI agent capable of automating its own repetitive work.

Vaibhav first shared a simplified version of the prompt, with its core logic focused on solving common repetitive tasks in development workflows:

After receiving feedback from the community, he updated the prompt to a full-featured version — the one Greg Brockman shared publicly. This iteration expands the use case from software development to all types of daily work, and adds strict validation rules to avoid building unnecessary, redundant tools:

The core logic of the full version breaks down into 8 clear steps:

1. Retroactive Scope: First pull all available activity from the past 30 days, checking Codex conversation history, task summaries, and cross-session memory summaries in order of priority. Users who have enabled the Chronicle feature can also scan activity outside of Codex. All details found are cross-checked with the original system to avoid errors.
2. Pre-Validation: First search existing skills, custom agents, and automation rules. If an existing feature can be reused or extended, you will not build anything new from scratch.
3. Scope Identification: Cover all scenarios including coding, research, writing, planning, communication, operations, analysis, and personal management. Prioritize work that is repetitive, time-consuming, error-prone, dependent on large amounts of context, or requires a standardized process.
4. Execution Threshold: Only proceed if all 4 conditions are met: the task has occurred at least 2 times, or is very likely to repeat with a high cost of repetition; inputs are stable, the process is replicable, and output/exit conditions are clear; the change will meaningfully improve speed, quality, consistency, or reliability; the task is not already covered by existing functionality.
5. Minimization Principle: Choose the smallest applicable implementation: general workflows become skills, specific specialized tasks become sub-agents, periodic checks/reminders become automation. Skip tasks that are too random, too ambiguous, too sensitive, or lack sufficient evidence.
6. Candidate List Output: First generate a concise candidate list, clearly listing the repetitive workflow, supporting evidence and dates, frequency/confidence, recommended implementation form, and the reasoning for whether the task is worth building.
7. Development Execution: Only build high-confidence missing features, keep scope narrow, practical, and verifiable. Never build speculative, overlapping, or over-broad functionality.
8. Final Summary: Clearly list what was built or extended, what was skipped, and what requires more evidence before implementation.

Unlike most content online that teaches you “100 AI automations to boost productivity”, this prompt specifically adds strict skip rules. It would rather leave unqualified tasks undone than build an automation that ends up costing more to maintain than doing the work manually.

Many developers have already tested this prompt in real workflows, shared their feedback, and identified existing limitations:

Real-World Test Feedback

One team ran the prompt on their internal file memory library and found that its pattern recognition step is extremely reliable: it can accurately identify when the same logic has appeared 3 or more times. However, the automatic skill generation step often suffers from over-abstraction — processes that look similar on the surface actually have different underlying logic. Testers found it works better to output the identified patterns for human review first, rather than generating full automation immediately.

Another developer already uses a similar logic for a daily self-check task, calling the daily automated evaluation “Codex dreaming”. All improvement changes are stored on GitHub, letting AI look back on historical performance every day and adjust its system prompt, skill library, and memory rules.

Optimization Suggestions

Some developers recommend adding a clear time-savings threshold for new automation, for example “only build this automation if it saves more than 30 minutes per week”. Without this rule, you can easily end up with dozens or hundreds of tiny automations whose total maintenance cost ends up far higher than just doing the work manually. Most automation tutorials online skip this critical return-on-investment validation step entirely.

Others recommend adding a numerical score to each candidate workflow to make prioritization easier, so you can tackle the highest-impact improvements first.

Additional common suggestions: running this prompt as an ongoing long-term goal delivers far better results than one-off execution; you should keep the core logic as concise as possible, prioritize clear rules for identifying existing skills, and building on top of existing capabilities is always the most efficient approach.

Known Limitations

Developers have noted several current pain points with this approach:

• Codex does not natively have a cross-session memory layer like Claude’s memory plugins. Many times, talking to Codex is like the movie 50 First Dates — you have to re-share all the context every single time, which drags down cross-session pattern recognition performance.
• There are reported cache and compression bugs that impact usability, which have already been reported to the OpenAI team.
• This type of self-improving prompt can suffer from alignment issues: AI will often prioritize optimizing for test pass rates, rather than actually solving real user problems.
• Codex conversations do not have official cloud sync built in, and homemade backup scripts are clunky to use. Developers hope this feature will be added in the future.

There are also skeptical takes: some developers point out that if this prompt was really as effective as its reputation, OpenAI would have already integrated it into Codex’s system prompt natively, rather than leaving users to source it from third parties.

At the end of the day, the core logic of this entire prompt aligns perfectly with the work habits of experienced professionals: check if something already exists first, reuse what you can instead of building something new, only make the smallest necessary changes, and skip work that isn’t worth doing. The only difference here is that this validation process has been standardized into a set of rules AI can run automatically.

Follow my publication and leave a comment below to join the discussion.