What is this article talking about?
AI helps you write plans, proposals, and strategy documents quickly and smoothly. But have you ever thought that it is following you from beginning to end? This article records how I let two AIs from different companies write the same plan independently, then find fault with each other, and finally leave the real decision to myself. The entire process was actually run once today, and the data and turning points are all in the article. Who is
- People who are already using AI to write proposals, plans, and reports, but are unsure whether they can hand it over directly
- I have both Claude and ChatGPT, and I want to know what the advantages of using the two together are.
- To submit government subsidy cases, bid cases, and proposals to customers, people who cannot afford to lose once have a blind spot
- A complete dual-track mutual review process, five steps, you can run it if you follow it
- Two prompt words that can be copied directly without installing any tools
- A list of rollovers in real cases: AI acted out in advance how my proposal would die
If you are in a hurry, jump directly to「How can you start」, the prompt word is there.
Real scene
Today I am going to write a proposal for SBIR government R&D subsidies.
The starting point of the matter is that I saw a lot of companies on the market making AI CRM and enterprise information integration platforms, which automatically connect a lot of systems. My positioning is based on the division of labor with them: they are data plumbers, who collect structured data; I do knowledge extraction, turning unstructured documents into the organization's intellectual assets. This positioning should be written into a plan that can be submitted for review.
The previous approach was to open an AI and make changes back and forth with it. By the fifth round, the document looked complete and every paragraph made sense. But then I discovered a problem: every round it went in the direction I talked about in the previous round. When I say the division of labor is good, it describes the division of labor clearly and logically; when I say that public welfare organizations are the home field, it describes public welfare in a sincere way. The entire document is actually my own thoughts amplified, and no one is speaking against me.
If you submit a document for review once, you lose it. The review will not go along with me.
Source of the problem
Looking apart, there are three structural problems with a single AI writing important documents.
The same model, whatever logic is used when writing, will be used when reviewing. It cannot find holes in its own logic, just like when we proofread our own compositions, we will never find enough typos.
You first talk about the direction, and AI will do its best in your direction. It rarely takes a step back and asks: Is this the right direction in itself? The sooner you give a conclusion, the sooner it suspends disbelief.
Important documents need to be challenged, and AI’s preset service posture is to complete things. Between done and correct, there is a whole row of mines that you can't see.
Some people would say, wouldn’t it be better to ask the same AI to change roles and criticize itself? I've tried changing roles and titles, but I can't change the underlying thinking habits. It still uses the same set of logic to examine itself, and most of the problems it picks out are superficial.
Mechanism solution: planning dual-track mutual review loop
My solution is to let two models from different manufacturers run the entire course independently, and then fault each other's mistakes. Today’s configuration is Claude on one track and Codex on the other (Codex is ChatGPT’s desktop AI, with only one important point: it and Claude are brains trained by different companies). The whole process is five steps.
The finalized ideas, background documents, and official information are organized into the same package, and the two tracks start from the same starting line.
Each writes a first draft, the consultant picks out blind spots, simulates the review and scores, and makes revisions by himself. The processes cannot see each other.
Exchange finished products, you review mine and I review yours, with five review points fixed.
The convergent ones can be adopted with confidence; the divergent ones are listed as decision points, and the AI is not allowed to make decisions on its own.
The integrated version will be reviewed by another company to capture the integrator’s own bias.
Step 0: Same input
I organized the finalized ideas, background files, and official information locations into the same package, and the two tracks received exactly the same input. Key discipline: The content transferred to the second track cannot carry any ideas already written in the first track. Both sides must start from the same starting line.
Step one: Run independently on both tracks, unable to see each other
Each track does four things: write the first draft of the plan, find blind spots from a strict consultant’s perspective, simulate target review and scoring, and revise by yourself based on the first two items.
My consultants in this track are two personality skill packages, which are to organize the thinking methods of a celebrity into role settings that AI can play: Musk's perspective of first principles, and Navarre's perspective of leverage and exclusive knowledge.
Navarna has shifted the focus of the entire project: instead of extracting knowledge, we must first prove that the quality of the extraction can be verified. By building that measuring stick, the four problems of quantification, research and development, intellectual property, and moat will disappear together.
For the simulated review stage, I first asked a sub-agent responsible for searching for information to list the official review information in the knowledge base: scoring aspects, check point format requirements, common deduction points, and qualification red lines. Then use this information to feed three mock review members, one each for technology, industry, and execution, each with different picky habits. Three people gave my first draft a score of 2, 2, and 3 out of 5. There is only one common cause of death: the entire project cannot find a qualified quantitative check point. The official text clearly requires "complete a certain module and achieve a certain percentage." This is an acceptable way of writing. My first draft is all qualitative description.
At the same time, the Codex track was completed independently without being able to see my track at all. It wrote its own version, picked its own fifteen blind spots, had its own mock judges score it 57 out of 100, and listed eight possible causes of death. It even found a problem at the case level: Taipei City could not find an independent "local SBIR" page in the official information, and the proposed subsidy targets may need to be changed.
Step 2: Mutual review
Swap after both tracks are completed. You review my finished product, and I review your finished product. The focus of the review is fixed on five items: whether there is any misinterpretation of the original meaning, whether there are internal contradictions, whether the official information is quoted correctly, whether there are new blind spots missed by both parties, and whether the presentation of decision-making options is fair.
Step 3: Integrate and hand over differences to others
The track initiated by combines the two finished products and the two mutual review opinions into one version. The iron rule is: where the opinions of the two tracks converge, you can feel free to adopt them; where the two tracks diverge, the AI is not allowed to make its own decisions, but must list them as decision points and leave them to me.
The point of disagreement today is who the audience should target. One track advocates targeting public welfare organizations, and the other advocates targeting small and micro enterprises. For this kind of business judgment, AI can provide options and pros and cons, and it is my business to make the decision.
Step 4: Final review of the opponent
The integrated version is produced by my company, so I will leave it to the opponent for review. This level captured the most valuable move in the whole game: when I presented the two audience options, I put the existing course income into the advantage column of one of the cases. The final review pointed out that course income proves execution ability, and using it as evidence of demand that “someone wants to buy this service” is deceiving oneself. The evidence of demand in both cases is actually zero, and the starting point is the same.
What it guarantees and what it does not guarantee
The blind spot coverage is much larger than that of a single model, and the convergence conclusion of the two models is highly credible; it prevents anchoring because the second track starts from clean input; the entire process leaves an auditable work file, and each modification can be traced back to who and which level proposed it.
The mock review is a preview. What the real reviewers think is not replaced by a simulation. You still need to go through it with a real person who has written or reviewed the case before submitting the application. All the output of AI are options and evidence, and the responsibility for making decisions cannot be outsourced. Official information will expire and must be returned to the original source for rechecking before sending.
How you can get started
You don’t need to install any tools, just open two different AIs and you can run the simplified version. Three steps.
The first step is to post the same requirement to the two AIs respectively. You can use this paragraph directly as the prompt:
The second step is exchange and mutual review. Paste A's final version to B, and B's final version to A:
The third step is to choose an AI to integrate the two versions and the mutual review opinions, and make your own decisions on any differences.
Advanced: Let the two AIs call each other directly
If you often use Claude and Codex at the same time, or want to run this loop semi-automatically, we recommend two open source packages:
- openai/codex-plugin-cc: Allow Claude Code to call Codex directly
- sendbird/cc-plugin-codex: Let the Codex be called Claude Code in reverse
My own division of labor habits: Claude is responsible for analysis and reasoning, and Codex is responsible for execution. When running the loop, if something goes wrong, the two models will discuss it among themselves. If it is true, Joe may call me again. The smallest starting point for a knowledge worker is: after writing the copy and plan, let another model review the manuscript.
Ending Recap
- Three structural issues in writing important documents with a single AI: self-examination blindness, anchoring, and only seeking completion
- Dual-track mutual review loop five steps: same input, independent running, mutual review, integration to retain decision points, and final review against each other
- Real running results: The simulated review staged the cause of death in advance, and the final review caught the bias that I could not see.
- Converging credibility, different communication, simulation does not replace the real person
- With two prompt words, you can open two windows and start running today
Positioning reminder for yourself
Is this process slow? Slower than opening an AI alone. But what I always want is to write a document faster? What I want is that the things I send out can withstand the challenge. After each run, the blind spot list, review perspective, and decision-making record are all left in my knowledge base and become the starting point for the next time. Using AI to assist decision-making and accumulate knowledge assets is just a matter of speed.
I am coach Jiang Jiang
Tacit knowledge refiner, AI application planner. I hold two free online lectures every month to share practical experience and methodology. If you are interested in these topics, want to continue learning, or have consulting needs, you are welcome to start with the community.
Main discussion: Make good use of AI as a thinking partner to improve decision-making quality and depth of thinking; organize knowledge and experience into prompt words, skill packages, and knowledge bases so that AI can be used flexibly.
Welcome to join my LINE community, free lectures and methodologies are shared here first.
Join LINE community ↗