Why You Must Still Write Down Your Ideas in the Age of AI

Twelve years ago, I wrote an article whose core idea was simple: if you have an idea in your head, write it down. Back then, there was no ChatGPT, or at least ordinary people didn’t have to worry that a machine could write a more complete article in a few seconds after they’d just thought of an idea.

Twelve years later, I handed that old article to a large model and asked it to reorganize it in light of today’s AI. A few seconds later, it produced a new draft with a clear structure and fluent language. To be honest, many parts were indeed better written than the original.

This is a bit awkward: if AI can already write faster and better than me, why should I still write an article persuading others to write themselves?

That old article aimed to answer a question: why write down ideas that emerge during research? The “research” I referred to wasn’t just about academic papers. When someone gathers information, compares explanations, or tries to solve problems in work and life, they are broadly doing research.

Twelve years later, I no longer just record and exchange ideas on Douban and Zhihu. GitHub, Quora, LinkedIn, and X have also become platforms where I participate in discussions. Platforms, languages, and readers have changed, but for knowledge to accumulate, someone still needs to write down what they know.

So, if I could only leave one piece of advice, I would still say: please, definitely write down the ideas in your head. Writing is not just about producing text; it’s also the process of transforming experience into knowledge, and vague feelings into judgments.

The More AI Organizes, the More Important Human Original Records Become #

If an engineer solves a fault but doesn’t record the cause and resolution process, subsequent people may still repeat the same mistakes. If a researcher finds problems with a certain explanation but leaves no evidence or reasoning, others cannot continue the discussion based on it.

Large models can quickly organize existing materials, but they don’t know about experiences that have never been recorded. They don’t know the unspoken concerns in a meeting when a project failed; nor do they know why a technical solution that looks reasonable in documentation becomes difficult to run in a real system. Only when those who experienced it write down their observations and judgments can these experiences become knowledge usable by others – and by future machines.

Therefore, the better large models are at reorganizing text, the more crucial human first-hand records become. What is truly scarce today may not be grammatically correct paragraphs, but rather what a person has experienced, what they observed from it, and what different explanations they proposed.

What Truly Needs Vigilance is Outsourcing the Entire Thinking Process #

Large models do improve the efficiency of some writing tasks. A preregistered online experiment had 453 university-educated professionals complete moderately difficult, career-related writing tasks. Participants who received ChatGPT assistance reduced their average completion time by approximately 40% and improved output quality by about 18% [1]. These results come from specific simulated tasks and cannot directly represent all real-world work, but they are sufficient to show that AI can increase the efficiency of some writing tasks. There’s certainly no problem with letting AI summarize meeting minutes, revise emails, or condense materials. There’s no need to deliberately waste time that can be saved.

However, “getting an article faster” and “completing a thought better” are not the same thing.

To exaggerate a bit, rather than worrying about AI making me unemployed, I’m more concerned about being used as an “idiot” by AI first. It’s not that AI suddenly lowers my intelligence, but that answers come too quickly, making it easy for me to skip analysis, comparison, doubt, and expression.

MonkeyUser’s comic 《Deprecated》 pushes this concern to an absurd extreme: AI doesn’t continue to get smarter, but humans continuously degenerate after handing tasks over to machines, thus both sides reach a “singularity” in a different way. The comic below is adapted from this premise.

I also discussed similar issues in 《Will Humans Stop Thinking? Dune’s Sixty-Year-Old Warning and Human Dignity in the Age of AI》. What Dune truly warns against is not machines beginning to think, but humans gradually abandoning judgment in convenience and outsourcing responsibility to machines.

A blank page originally exposed difficulties in thinking. Can evidence support a conclusion? What is the relationship between two viewpoints? Am I truly opposing the result, method, or premise? Large models can now quickly fill the blanks, using complete structures and fluent language to hide these problems. The article seems finished, but the author may not have truly formed a judgment.

In 2025, researchers surveyed 319 knowledge workers and collected 936 instances of their use of generative AI at work. The results showed that, after controlling for factors such as task type and individual characteristics, respondents with higher confidence in generative AI reported less engagement in critical thinking for related tasks. This is a correlation, not sufficient to prove that AI causes reduced thinking. Respondents’ descriptions also indicated that they applied critical thinking more to information verification, answer integration, and task supervision [2]. Using AI does not equal stopping thinking. Some work simply shifts from “writing out the answer oneself” to “judging whether this answer can be trusted.” Even if typing and preliminary organization are skipped, verification and judgment do not automatically disappear.

Writing Is Not Just Output, It Is Also Part of Thinking #

In 《Exploration and Generation: A Meaning of Life》, I interpreted ’exploration’ as discovering new possibilities and ‘generation’ as making these possibilities stand firm and allowing others to further develop them. Writing is precisely a step from exploration to generation: it’s not just about outputting already formed conclusions, but also allowing an idea to gradually stand firm through recording, dissemination, and verification.

Twelve years ago, I summarized the role of writing as memory, understanding, and forming opinions. Looking back today, I prefer to describe them as preserving reasoning, testing understanding, and forming opinions. These three things have not lost their meaning due to the advent of large models.

Writing preserves reasoning. People tend to remember the final conclusion but forget what materials were used, what explanations were excluded, and what conditions were necessary for the conclusion to hold. Writing down the working process preserves the path by which an answer was formed. Only when rereading months later can you distinguish between facts and mere speculations from that time.

Writing tests understanding. We often think we’ve understood something until we try to articulate the problem clearly, only to find gaps between concepts. Does a case truly support a universal conclusion, or does it only illustrate a possibility? When comparing two solutions, what does each sacrifice? Writing these questions into complete sentences forces us to confront areas that were originally vague.

Writing forms opinions. An opinion is not just a pretty sentence. It includes what facts a person has chosen, how they interpret those facts, what they have overlooked, and what consequences they are willing to bear. The better AI is at generating answers, the more a person needs to judge what questions are worth pursuing. Personally writing forces one to make choices: What do I truly care about? What do I want to explain? Why is this question worth occupying my and the reader’s time?

The part that truly belongs to you is usually the part the model initially doesn’t know: your experiences, confusions, hesitations, and where you diverge from common answers.

How I Now Divide Labor with AI #

My current rule for myself is simple: if the task is merely to convey already established information, such as adjusting the tone of an email or organizing a fixed-format report, I can directly let AI start polishing it.

If the task involves analysis, judgment, or choice, I usually first organize the material using my accustomed STAR method. The STAR method here differs slightly from the common version used in job interviews; I’ve changed the T from Task to Target:

• Situation: What happened? What facts are established? What are the constraints?
• Target: What problem do I want to solve, or where do I want things to go?
• Actions: What actions have I already taken? If things haven’t started yet, what are the available options?
• Results: What actually happened? How big is the gap between the result and the target?

This framework can be used for both post-mortem review and pre-planning. If something hasn’t happened yet, Target still represents the desired state; Actions don’t have to pretend to be fixed, but can list several options, along with their costs, risks, and reversibility; Results can then be changed to the expected outcomes of each option. I can use these expected outcomes to judge how far each option is from the target and what costs need to be borne. At this point, AI can help me discover information gaps, compare solutions, and deduce risks, but it cannot judge for me which target is worth pursuing, much less bear the consequences of choices for me.

If I am just researching a problem, or an idea has just emerged, I use a lighter version: what I observed, my temporary understanding, and what I haven’t figured out yet. Three sentences are fine, and it doesn’t matter if it’s poorly written. After writing, I then let AI propose counterexamples, check concepts, adjust the structure, and condense repetitive content.

STAR is not a prompt template that everyone must copy. It’s just a checkpoint I set for myself: I first provide the real-world situation, goals, and preliminary judgments, and then let AI participate in organizing and deducing. Others can also create their own checklists. The key is not which acronym to use, but to think clearly about what tasks can be handed over to AI and what judgments cannot be outsourced before opening the chat box.

By the way, this article also used AI. I used GPT-5.6 Sol (Medium inference intensity) from Codex. The old article from twelve years ago provided the raw material; the core judgments and decisions on what to revise this time were mine; AI helped me adjust the structure, check for repetition, find relevant research, and improve expression.

This does not contradict the article’s main argument. The distinction is not whether AI was used, but whether I first had my own observations and judgments, whether I verified the information, and whether I was ultimately willing to take responsibility for the views in the article. This division of labor may not get me an article the fastest, but it allows AI to help me write better, rather than skipping the thinking process for me.

I’m glad I didn’t let that idea stay in my head twelve years ago just because I didn’t write well enough.

If I hadn’t written it down back then, no matter how smart today’s large models are, they wouldn’t have this old article to work with.

So, if you have an idea in your head, just write a few sentences first. It’s okay if it’s not well written – AI can always help you revise it, but it cannot, twelve years from now, retrieve a self that left no record whatsoever.

Related Articles #

• 《When a Computer Science PhD No Longer Automatically Appreciates in Value: Choices, Training, and Career Paths in the AI Era》: Why problem definition, evidence judgment, and responsibility become more important when candidate answers, code, and drafts are increasingly easy to generate.

• 《Achieving Clarity of Mind: The Times Are Changing, What Do Effort and Choice Mean?》: Beyond getting the answers right, why people also need to choose problems worth long-term commitment.

References #

[1] Noy, S., Zhang, W. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 2023. https://doi.org/10.1126/science.adh2586

[2] Lee, H.-P., et al. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025, 2025. https://doi.org/10.1145/3706598.3713778

Appendix: My English Tech Blog Writing Rules #

The main text ends here. Below is the SKILL.md I developed while reviewing the English edition of this article, provided for readers who wish to copy and adapt it. It does not ask AI to form opinions on my behalf. Instead, it helps AI examine an article’s central argument, structure, prose, evidence boundaries, and citations.

---
name: english-tech-blog-writing
description: Draft, review, and edit evidence-led English technology blogs while preserving the user's authorial voice and maintaining a clear argument, explicit evidence boundaries, and accurate citations. Use for English technical blog writing, structural review, bilingual adaptation, or editorial refinement.
---

# English Tech Blog Writing

Preserve the author's viewpoint, judgment, and personal voice. Explain complex technology clearly to knowledgeable readers while respecting facts, uncertainty, and the limits of the evidence.

## Determine Task Type

- **Draft a New Article:** State the central claim in one sentence, then build the argument around it. Draft directly unless essential facts are missing.
- **Edit an Existing Draft:** Make editorial-level changes by default. Preserve the structure and viewpoint unless the structure obstructs the argument.
- **Review:** Identify the exact passage, problem, reason, and proposed direction. Separate required corrections from optional improvements. If the user asks only for a review, do not rewrite the draft without approval.
- **Check Terminology:** Check technical accuracy, usage, audience fit, and consistency throughout the article.
- **Adapt from Another Language:** Preserve the final source article's substantive meaning, evidence limits, citations, images, and authorial voice. Keep the section structure aligned unless a change is needed for natural English or reader context, and make any material divergence explicit. Write natural English rather than translating sentence by sentence, and briefly explain cultural context when readers need it.

## Audit the Argument Hierarchy

For a full review, structural edit, or rewrite, work from the top down:

1. State the article's central claim and identify the question it is trying to answer.
2. Summarize the job of each section and check whether it advances the central claim.
3. Identify the main point of each paragraph and check whether it serves its section.
4. Find repeated arguments, logical gaps, misplaced material, and paragraphs with no clear function.
5. Decide whether the title and section structure need to change before editing paragraphs and sentences.
6. Reorganize only when the existing structure obstructs the argument. Do not rewrite a useful paragraph merely because another style is possible.

## Build the Article's Throughline

- Express the central claim in one sentence. It must be arguable, not merely a topic label or slogan.
- Begin with a concrete situation the reader can understand, then introduce the real question promptly. Do not overload the opening with background, definitions, or sweeping claims.
- Prefer a progression such as "evidence or observation → mechanism → consequence → judgment." Add limitations, counterexamples, or engineering conditions where they matter.
- Let each section answer one clear question. Sections should form a causal, progressive, or contrasting sequence rather than a stack of loosely related material.
- Make every technical detail serve the throughline. Remove terms, examples, and summaries that do not change the reader's understanding or judgment.
- End by returning to the central claim and drawing limited, actionable implications. Do not mechanically recap the article or force a grand conclusion.

Use the following sequence when it suits the subject, but do not force it into every article:

- A real-world change or reader's question
- How the system used to work
- Which part the new technology changes
- The mechanism behind that change
- Engineering practices, business results, or user impact
- Capability boundaries, risks, and unresolved problems
- Practical implications for practitioners or general readers

## Write Natural English

- Prefer concrete subjects and strong verbs when they make agency, responsibility, or causality clearer. Use the passive voice when the actor is unknown or genuinely unimportant.
- Avoid excessive nominalization, long noun stacks, inflated abstractions, marketing language, and phrasing that reads like a literal translation.
- Use transitions only when the argument genuinely turns, qualifies, advances, contrasts, or reaches a consequence.
- Avoid mechanical signposts such as "firstly," "secondly," "finally," and "in conclusion" when the structure is already clear.
- Reduce empty parallelism, slogan-like claims, throat-clearing, repetitive summaries, and excessive subheadings.
- Vary sentence length naturally. Let each paragraph make one main argumentative move rather than compressing several levels of causality into one sentence.
- Prefer familiar words when they express the same meaning as fashionable or needlessly specialized vocabulary.
- Briefly explain platforms, institutions, events, and expressions that may be unfamiliar outside their original cultural context. Add only the context needed for the argument.
- Preserve restrained and conditional judgments. Do not turn them into stronger claims merely to make the prose sound more confident.
- Follow the author's established spelling and punctuation convention. If none is evident, choose one English convention and apply it consistently.

## Handle Terminology, Technology, and Evidence

- Introduce an unfamiliar technical term with a concise plain-English explanation. Define an acronym at first use unless it is already standard for the intended audience.
- Do not replace an established technical term with an awkward paraphrase merely to avoid jargon. Keep terminology consistent across the title, summary, headings, diagrams, and body text.
- Distinguish facts, research findings, engineering practices, reasonable inferences, the author's views, and forecasts.
- Do not present a research prototype as a generally deployed product capability, correlation as causation, or an individual case as an industry-wide pattern.
- For quantitative claims, state the measurement, population or sample, time period, and scope of applicability. When sources conflict, explain the difference rather than hiding it selectively.
- In a technical example, make the input, mechanism, and output clear. Explain what the mechanism does for general readers instead of listing implementation terms.
- Verify rapidly changing claims about models, products, prices, policies, standards, or markets against current authoritative sources.

## Use Citations

- Use sequential citations such as `[1]` and `[2]`, numbered by first appearance.
- Place a citation immediately after the specific fact or claim it supports. Do not leave several sources hanging at the end of a paragraph with no clear mapping.
- List references in order of first appearance. Prefer research papers, official documentation, standards, and primary data.
- News and industry reports may support market events and case descriptions, but should not replace technical or scientific evidence for a mechanism.
- Paraphrase sources instead of using unnecessary long quotations. Keep any verbatim quotation brief and faithful to its context.
- Do not invent authors, titles, dates, DOIs, URLs, page numbers, or access dates. Mark anything that could not be verified.

Maintain consistent reference format, for example:

[1] Author. "Article or Paper Title." Journal, Conference, or Institution, Year. URL or DOI.

## Editorial Check

Check each item before delivery:

- [ ] Does the title accurately promise the article's content, avoiding overly strong conclusions?
- [ ] Does the opening enter a concrete problem quickly and transition naturally to the first section?
- [ ] Does the core judgment run through all sections, rather than appearing only at the beginning and end?
- [ ] Does each section answer a question and propel the next section?
- [ ] Are actors, pronoun references, causal relationships, and time frames clear?
- [ ] Is unfamiliar cultural context explained briefly where necessary?
- [ ] Is the English natural, and are spelling, punctuation, and terminology consistent?
- [ ] Are facts, research findings, inferences, and the author's views clearly distinguished?
- [ ] Do citations support the adjacent claims, and do citation numbers match the reference list?
- [ ] Are there remaining repetitions, slogans, translation artifacts, mechanical summaries, or unnecessary technical details?
- [ ] If the article was adapted from another language, are its substantive meaning, evidence limits, citations, and images aligned with the final source, and are any structural differences deliberate?

Deliver the complete, usable main text by default. Only provide modification notes, comparative versions, or multiple options if requested by the user.