<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[bayswong]]></title><description><![CDATA[bayswong]]></description><link>https://bayswong.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 19:20:57 GMT</lastBuildDate><atom:link href="https://bayswong.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[From Creating a Bot to Content Distribution: My Practice and Reflections on Using Grok Bot for Online Operations]]></title><description><![CDATA[I am bays, an independent developer responsible for both product development and content operations. Recently, I have used Grok Bot for trend monitoring…]]></description><link>https://bayswong.hashnode.dev/from-creating-a-bot-to-content-distribution-my-practice-and-reflections-on-using-grok-bot-for-online-operations</link><guid isPermaLink="true">https://bayswong.hashnode.dev/from-creating-a-bot-to-content-distribution-my-practice-and-reflections-on-using-grok-bot-for-online-operations</guid><category><![CDATA[AI]]></category><category><![CDATA[automation]]></category><category><![CDATA[Productivity]]></category><dc:creator><![CDATA[黄贝]]></dc:creator><pubDate>Mon, 16 Nov 2026 00:17:27 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/87f96240-ef55-49e7-9f88-3fcdeb0b0edf.jpg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>I am bays, an independent developer responsible for both product development and content operations. Recently, I have used Grok Bot for trend monitoring, source verification, article production, multi-platform adaptation, and product promotion, and have gradually formed a working approach based on AI execution and human review.</p>
<p>Compared with simply generating copy, I place more value on whether it can continue to complete subsequent tasks according to a clearly defined method: retrieving original sources, preparing versions in different languages, handling cover images, entering platform backends to save drafts, or turning product documentation into accessible web pages.</p>
<p><strong>Its core value is turning a method of work that has already been made explicit into tasks that can be executed repeatedly, reducing the gap between organizing information and carrying out specific operations.</strong></p>
<p>But the ability to execute is not the same as the authority to publish, still less as a platform's acceptance. Using tools to post content in communities such as Reddit, NodeLoc, and NodeSeek carries a risk of account suspension. These constraints are also a part of evaluating actual results that cannot be set aside.</p>
<h2>1. Creating a Google Trends Bot: Turning an Analytical Method into a Dedicated Assistant</h2>
<p>Recently, I created a Google Trends assistant to continuously track keywords and provide reference points for content topic selection and product direction.</p>
<p>The starting point of the creation process was defining the task in natural language, rather than first designing a software system. The initial requirements I provided included: tracking <code>ai video generator</code>, compiling related breakout queries daily, marking newly appearing queries, and comparing them with <code>GPTs</code>. The initial analysis scope was the United States, over the past 30 days.</p>
<p>I later adjusted the run time to 5:00 a.m. every day, including weekends. Once the monitoring target, output requirements, and execution time were made explicit, the task had a dedicated role that could keep running.</p>
<p>More importantly, the analytical method could continue to be iterated.</p>
<p>I later provided reference images and a keyword-mining method, and asked that the criteria for judgment be adjusted accordingly: when discovering new terms, look at popular and rising queries worldwide over the past 7 days and the past 1 day; when assessing an opportunity, then use a longer period to determine whether it is a new term, a second surge, or an older term heating up again, and check search results and related long-tail queries.</p>
<p>There is a detail that needs to be clarified here: the percentage increase shown among related queries is not the same metric as the interest curve obtained by looking up a term on its own. If the rules of analysis are not made explicit, even a result produced every day may still lead to persistent misreading.</p>
<p>Therefore, the process of creating this Bot actually consisted of three actions: defining the task, adding a method, and correcting it in light of the results.</p>
<p>It did not complete collection smoothly the first time. The first run encountered rate limiting on Google Trends, and no deduplication baseline had been established yet. After the criteria were later adjusted, an initial baseline was formed on November 14, and updates continued the following day. For keywords I did not understand, I could follow up in the same assistant without providing all of the background again.</p>
<p>The most immediate takeaway from this case is: <strong>creating a dedicated Bot can now begin with describing a piece of work.</strong></p>
<p>What to watch, which method to use, and when to deliver can all be set in natural language; reference materials and later corrections can also continue to be added. The barrier to creation has been lowered; the part that truly requires effort has shifted to whether the method itself is sound and whether the results are trustworthy.</p>
<p>What is easy is saving a method of work as a reusable assistant, not automatically obtaining a correct market judgment.</p>
<h2>2. Multi-Platform Content Production: Adaptation Is Not Turning a Full Article into a Summary</h2>
<p>Another repetitive task in content operations is converting the same article into versions that different platforms can use.</p>
<p>Recently, I carried out complete content preparation and distribution adaptation around Fable 5.2 and around "AI Makes Us Ten Times More Efficient. Why Are We Even More Exhausted?"</p>
<p>This work began with verifying sources. Taking Fable 5.2 as an example, I needed to compile public cases, original post sources, and the author's own account, and to distinguish facts that could be checked from promotional framing. In particular, "one-shot generation" could not be conflated: it might mean a single initial prompt, or it might mean that the entire process involved no subsequent human edits.</p>
<p>After entering the writing and distribution stage, I adopted explicit platform rules: the WeChat official account and Bilibili columns kept the full Chinese text; Dev.to and Hashnode used the full English text; Xiaohongshu was separately condensed into a version suited to card-style reading. Zhihu automatically syncs from the WeChat official account, so I no longer republish the same piece by hand; I handle Medium myself.</p>
<p>These rules came from corrections made in actual use. There had previously been a problem of the text being overly compressed after adaptation, so I explicitly required that, except on Xiaohongshu, a summary must not replace the original argument.</p>
<p><strong>Platform adaptation should adjust language, layout, and the form of assets, rather than deleting the article's analytical process by default.</strong></p>
<p>Cover images and illustrations also belong to the same deliverable. Different platforms require different sizes, and after cropping one must check whether the main subject remains intact; screenshots of sources and generated diagrams must be distinguished, and a schematic image must not be described as a measured result. For that Xiaohongshu article, eight text cards were made, plus one cover.</p>
<p>The advantage of Grok Bot is that these steps can continue around the same source text. I do not have to re-explain the background for translation, image-making, and draft-filling separately, nor prepare assets from scratch for each platform.</p>
<p>However, the main deliverable of these two rounds of work was drafts ready for review; one cannot claim on that basis that the pieces have been publicly published on all platforms. Reducing repetitive operations and proving distribution results are two different things.</p>
<h2>3. Product Content Operations: After Writing, Confirm That the Page Is Accessible</h2>
<p>Product operations should not treat "the article has been generated" as the sole standard of completion. Whether the content is accurate, whether the page can be opened, and whether a visitor can understand the product from it need to be checked together.</p>
<p>Taking CheapBuddy as an example, around the need to connect DeepSeek Harness, I had Grok Bot compile a tutorial and address a problem in which the blog page would not open properly. The article could subsequently be accessed on the official site, and it clearly stated that CheapBuddy is not an official DeepSeek Harness product.</p>
<p>This kind of task connects search demand, content production, and site maintenance. If a tutorial exists only as a document, or if the published entry point cannot be accessed, later distribution lacks an effective place to land.</p>
<p>The same is true of product directories. The Indie Hackers product page for Remove Anything is already public, but other directories include cases that remain under review after submission, with the public listing not yet appearing. When required information that can be verified is missing, one should pause or skip, rather than invent facts in order to complete a form.</p>
<p>Therefore, I divide results into different states: materials prepared, drafts saved, forms submitted, and pages made public. Only the last of these can be described as already live.</p>
<p>These tasks show that AI can take part in more than copy production, including concrete operations before content enters the actual operating environment. But whether a live page brings search traffic or conversions is not something for which there is currently attributable data, and it cannot be stated directly as a growth result.</p>
<h2>4. Execution Can Be Delegated; Responsibility for Accounts and Publishing Cannot Be Transferred</h2>
<p>In actual practice, what most often requires human confirmation is not the body text, but the account, the save location, and the publication status.</p>
<p>On Hashnode, the account initially signed in was a product account, whereas the article needed to be saved under a personal account, so a switch was required first. A Xiaohongshu draft was saved only locally in the current browser, and could not simply be opened by a link on another device for continued editing. A WeChat official-account manuscript also had an incorrect title after import; only after checking the body and confirming it was the same article was the original title restored and the cover added.</p>
<p>These experiences mean that acceptance cannot only ask whether something is "done"; one must also check which account was used, where it was saved, whether it remains complete when reopened, and whether an external submission has already occurred.</p>
<p>Before publishing, one also needs to check cover cropping, factual wording, sources, and any originality or AI-related disclosures required by the platform.</p>
<p>Community channels carry a higher account risk. Platforms such as Reddit, NodeLoc, and NodeSeek cannot be treated directly as outlets for automatic distribution. In the case of NodeLoc, the account first received a mute notice of a "suspected automated account," and later a suspension notice reading <code>Account suspended: AIGC</code>.</p>
<p>These notices can be recorded, but one cannot infer from them that all suspensions were caused solely by AI assistance. The promotional nature of the content, the state of the account, community rules, and the platform's judgment of automation may all affect the outcome.</p>
<p>A human click to publish is not a waiver of responsibility, still less a guarantee of account safety. For communities of this kind, one must first judge whether the mode of participation complies with the rules and whether the content genuinely serves the discussion; after a restriction is received, the related actions should be stopped, rather than adding further publishing activity.</p>
<p>That a tool is capable of an operation, and that a platform permits that operation, are questions at two different levels.</p>
<h2>5. AI Is Changing How Information Is Processed, and Platform Governance Must Adjust as Well</h2>
<p>These practices ultimately point to a larger tension: users are increasingly using AI to read and process information, while many centralized platforms still design their rules around traditional human ways of visiting and publishing.</p>
<p>It is entirely reasonable for platforms to protect real users and genuine discussion. Mass spam, false identities, and advertising disguised as ordinary discussion consume community trust and crowd out normal exchange. Restricting these behaviors is fully justified.</p>
<p>But governance needs to distinguish: <strong>a real user using AI to assist their work, and using automation to manufacture fake participation, are not the same thing.</strong></p>
<p>A person who uses AI to read materials, organize sources, and translate an article, and who takes responsibility for the content after review, does not thereby lose their identity as a real user. Conversely, marketing content entered entirely by hand does not inherently have value.</p>
<p>Authenticity should be judged more by sources, facts, identity, and responsibility, rather than by whether every word was typed by hand.</p>
<p>As AI develops, more tool involvement will appear in retrieval, screening, reading, synthesis, and expression. A user may first have AI compare sources, then decide which pages to visit; they may also first form structured notes, then organize them into a public article. Platforms cannot assume that all valid participation in the future must fully reproduce past ways of operating.</p>
<p><strong>My judgment is this: if centralized platforms rely only on blocking AI to maintain the old model, and do not change their mechanisms for content distribution and participation, they will eventually be left behind.</strong></p>
<p>The question is not whether platforms need rules, but what the rules are aimed at. If "whether AI was used" is treated as the main criterion, responsible information work is easily mixed together with malicious flooding; if one only judges whether the writing "looks like AI," it is also difficult to truly identify whether the content is trustworthy.</p>
<p>A more reasonable direction is to recognize human–AI collaboration while strengthening constraints of responsibility. For example, provide a clear means of authorized access, set frequency limits on bulk behavior, require necessary disclosure of sources and conflicts of interest, and have accounts continue to bear responsibility for what they publish. These are my suggestions for the direction of governance, not features that existing platforms already all have.</p>
<p>For platforms that depend on content and user participation, long-term competitiveness should not consist only of keeping new tools out; it also includes whether they can continue to provide trustworthy content, effective discussion, and a reasonable return under new ways of processing information.</p>
<h2>Closing</h2>
<p>This period of use has given me a more concrete understanding of Grok Bot's value: saving an analytical method as a dedicated Bot, carrying an article through into complete platform versions, and turning product documentation into accessible pages.</p>
<p>These changes have not eliminated judgment and review; they have made them more important. When the cost of execution falls, the focus of the work should shift to whether the method is sound, whether the facts are reliable, and whether the mode of participation is appropriate.</p>
<p>I will not summarize the experience with unverified multipliers of efficiency or growth figures. For me, the change that has already occurred is this: repetitive operations can be handed off, while topic judgment, responsibility for the content, and the decision to publish remain in my own hands.</p>
<p>Users need to adapt to the changes in capability that tools bring, and platforms likewise need to adapt to changes in user behavior. Protecting authenticity and accepting AI assistance are not in conflict; what is truly difficult to sustain is treating the two as a choice of only one or the other.</p>
<hr />
<p>References for cases in this article:</p>
<ul>
<li><a href="https://cheapbuddy.cc/blog/deepseek-harness-cheapbuddy">CheapBuddy tutorial on connecting DeepSeek Harness</a></li>
<li><a href="https://www.indiehackers.com/product/remove-anything-2">Remove Anything's Indie Hackers product page</a></li>
</ul>
]]></content:encoded></item><item><title><![CDATA[Fable 5.2 and the Code-Driven Creative Workflow]]></title><description><![CDATA[A closer look at code-driven creative workflows, the limits of “one-shot” claims, and my own Saltline recreation with GPT-6 Astra.]]></description><link>https://bayswong.hashnode.dev/fable-5-2-and-the-code-driven-creative-workflow</link><guid isPermaLink="true">https://bayswong.hashnode.dev/fable-5-2-and-the-code-driven-creative-workflow</guid><category><![CDATA[AI]]></category><category><![CDATA[programming]]></category><category><![CDATA[Game Development]]></category><category><![CDATA[3d]]></category><dc:creator><![CDATA[黄贝]]></dc:creator><pubDate>Wed, 11 Nov 2026 11:43:32 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/80e0dcea-025e-404c-8b8f-4f9c2cffdb44.jpg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A text model starts making videos, building games, painting, and assembling 3D scenes. It has not become a video model or a piece of 3D modeling software. What it does is more direct: understand a goal, write code, call tools, and carry the work through to an output.</p>
<p>That is what interests me most about the recent Fable 5.2 examples. The point is not that entire creative industries are about to disappear. What is being challenged is the default route into them: open modeling software before making something in 3D; stitch together a production pipeline before making a video; start with an engine and assets before building a game.</p>
<h2>From a child's drawing to a playable game</h2>
<p>What happens after a child draws a set of characters and imagines a game around them?</p>
<p>In <a href="https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game">Aiden's Revenge</a>, Lex describes asking Fable 5.2 to continue an existing game project before he went to bed. The next morning, there was a version with 12 levels. After the children played it, they requested more moves and characters, which he then added.</p>
<p>What makes the story compelling is not a claim that AI finished a commercial game overnight. It is that a child's idea became something playable quickly. An adult no longer necessarily has to learn modeling, animation, and level editing before helping that idea take shape. Get it working first, then play and revise it together: that sequence is becoming practical.</p>
<p><img src="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/b352e98d-7f00-4db2-b22a-fa290dd39b4d.jpg" alt="Aiden's drawing of Uncle Lex alongside the game sprite." /></p>
<p><em>Image: Aiden's drawing of Uncle Lex alongside the game sprite. Source: Lex Toumbourou / <a href="https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game">NotesByLex</a>. This comes from the original article, not my own recreation.</em></p>
<p>The broader impact is not a single feature. A text model is beginning to take over the connective work in a creative process: turn an instruction into a program, a program into a scene, and a scene into a playable build or a finished video. Games are one entry point; drawing, animation, rendering, and sound follow behind them.</p>
<p>Before going further, it is worth being precise about what “one-shot” means in these examples.</p>
<h2>From kart racing to a child's game</h2>
<p><a href="https://x.com/petergyang/status/2119901658394358005">Turbo Tails</a> is a 3D kart racer. Peter Yang says he used three initial prompts to get a playable version, then spent a morning iterating on drifting, items, and multiplayer. That is not a final product from a single prompt. It is a working game loop followed by continued refinement.</p>
<p>Aiden's Revenge had a more specific starting point: a child's drawings and the action game he wanted to play. Lex's article records development on an existing project, the overnight task, and the additional characters and moves requested after playtesting. Its “12 levels” should not be rewritten as “a complete game built from scratch with one prompt.” The public version with 18 characters also incorporates later requests.</p>
<p><img src="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/356ba3d8-eeda-467a-bf6d-a2ce6a1949da.jpg" alt="The character selection screen with 18 fighters." /></p>
<p><em>Image: The 18-character selection screen. Source: Lex Toumbourou / <a href="https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game">NotesByLex</a>. The expanded roster includes later iterations, not just the initial output from one human instruction.</em></p>
<p>Two other examples make different trade-offs. <a href="https://x.com/razvanmuntian/status/2120099765413167525">SAPPER</a> gives Minesweeper a cinematic visual treatment. Its creator says it started from an empty project and took one prompt and roughly half an hour; that timing and usage are self-reported, not independently verified. <a href="https://x.com/ppercree/status/2119897710962917626">Fenlight</a> is not a large game but a moonlit swamp scene. Its creator describes it as a one-prompt project that took about an hour. One reworks the presentation of a familiar mechanic; the other delivers a small, coherent atmosphere. Not every demonstration needs to aim for a blockbuster.</p>
<h2>3D is not just about games</h2>
<p>These demonstrations challenge a familiar assumption about making 3D content: that convincing shapes, materials, and lighting must come after extensive manual modeling and parameter tuning, before the scene ever reaches an engine.</p>
<p><a href="https://x.com/MeltTheManual/status/2119926820963532984">Stonehenge</a> is an interactive scene combining the monument with the sun's path at the summer solstice. Its creator says it uses a single HTML file and Three.js, and shares cost and token accounting. It is a reminder that “one instruction” can still involve substantial runtime and computation. <a href="https://x.com/mattshumer_/status/2119877065701036047">Mirror Earth</a> is described by its creator as an explorable, destructible mirror of Earth with weather corresponding to the real world. In a <a href="https://x.com/mattshumer_/status/2119879860344766782">follow-up post</a>, however, the creator also notes that vehicles clip through geometry. The impressive parts do not make the limitations disappear.</p>
<p>Code-driven creation is not always a single-generation process, either. <a href="https://x.com/jkeatn/status/2119874007382299087">Jake Eaton's painting experiments</a> use Python and simulated brushes rather than an image generation model. The process he describes includes weeks of feedback and iteration on the brush engine. Instead of directly turning words into pixels, the text model builds a painting tool and then uses that tool to produce the work.</p>
<p><img src="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/6765655b-c204-4a58-802e-e243ac0c9430.jpg" alt="Painting shared in Jake Eaton's original post." /></p>
<p><em>Image: A painting shared in <a href="https://x.com/jkeatn/status/2119874018706874601">Jake Eaton's original post</a>. Eaton describes it as roughly 700,000 simulated brushstrokes. This is the original author's demonstration, not a photograph or an independently reproduced result.</em></p>
<p>The same distinction applies to video. In <a href="https://x.com/rileybrown/status/2119902109709537708">Riley Brown's Blender room film</a>, Claude writes code to build the scene, create shots and animation, use cloud GPUs for rendering, and connect audio tools. The text model is not directly generating video frames at its output. It is organizing a production system that can make a video.</p>
<h2>What does “one-shot” actually mean?</h2>
<p>“One-shot” should not be read as “the model made one request and finished everything in a few seconds.” At least four distinctions matter.</p>
<p>First, one initial human instruction does not mean the agent makes only one model request internally. Execution can include repeated planning, coding, running, and correction. Those internal requests do not necessarily mean a person kept rewriting the prompt.</p>
<p>Second, external tools do not automatically invalidate starting a task with one instruction. Blender, Three.js, cloud GPUs, and audio services can all be part of the workflow. Human input, internal model execution, and external service calls should be described separately.</p>
<p>Third, generating an initial version in one go does not mean the later public version was never iterated on. Turbo Tails received gameplay refinements; Aiden's Revenge incorporated playtesting feedback and requests for more characters. Using the final version to prove that one prompt produced every part of it gets the timeline wrong.</p>
<p>Fourth, continuing an existing project is not the same as starting from an empty one. The overnight Aiden's Revenge task began with an older version. That does not diminish its value, but it does determine how to describe it accurately.</p>
<p>The roughly 600 requests in the Aiden's Revenge article should not be rewritten as “the author manually revised the prompt 600 times.” That is the final usage tally across two days, not a count attributable entirely to the bedtime instruction. Internal requests, later human requirements, and tool execution need to be considered separately.</p>
<h2>A text model coordinating a production pipeline</h2>
<p>Brown says he gave Claude one prompt during a livestream. The model wrote code in Blender to build a room, a desk, a MacBook, lights, and Clawd. The laptop's screen even displayed a Claude-style webpage that the model had made.</p>
<p>Then came 13 shots, key presses, character jumps, and lighting changes from dusk through night to sunrise. According to Brown, Claude rented 10 GPUs on RunPod, completed the 4K render in 52 minutes at a cost of $5.75, and then handled color grading and titles, with ElevenLabs providing music and sound effects.</p>
<p>The striking part is not just the ability to write Blender scripts. It is the way scene construction, shot planning, animation, cloud rendering, and post-production become a connected workflow. The 10 GPUs are Brown's account of this particular run. What matters is the model's choice of a solution: use cloud compute to finish a high-resolution, multi-shot rendering task.</p>
<p>Taken together, these examples suggest that text models are moving toward a central role in creative software workflows. They do not necessarily need to replace Blender, game engines, drawing programs, or video editors. They can control and combine those tools through code, or even construct a task-specific tool along the way. The user describes the goal; the model connects software operations that previously required hands-on expertise.</p>
<p>That is the meaningful sense in which this crosses disciplines. It is not that the model beats specialists at every craft. It is that a general text interface starts compressing the operational barriers, format conversions, and orchestration between specialist tools.</p>
<h2>My Saltline recreation</h2>
<p>That brings me back to <a href="https://saltline.gameai.club">Saltline</a>. Felix Rieseberg says he tested Fable 5.2 with a single prompt and no human feedback, with the game's models, sounds, and materials made from scratch. He reports a runtime of about 8 hours and an estimated API-price cost of about $550, with roughly 2.7 million output tokens and 445 million input tokens, much of the input served from cache.</p>
<p><img src="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/1c2a54f0-6bdb-4428-bd3f-b51bd558b878.webp" alt="Title screen of my current deployed version of Saltline." /></p>
<p><em>Image: The title screen of my current deployed version of <a href="https://saltline.gameai.club">Saltline</a>, captured for this article. This version includes later menu interaction fixes and retains Felix's on-screen credit. It is neither my GPT-6 Astra recreation nor the untouched original release. No in-game HUD screenshot is included.</em></p>
<p>I also tried using GPT-6 Astra to recreate the Saltline workflow from end to end. In my attempt, the gap was not whether it could put a page on screen or make a character move. It was in the shapes of objects and the sheen of their materials: whether silhouettes matched the target, whether surfaces read as the intended material, and whether lighting produced convincing reflections. This was not a controlled model evaluation. But in this particular recreation, I found Fable 5.2 noticeably stronger at visual reasoning and coordinating work across steps.</p>
<p>My work on Saltline did not stop at the public demonstration. Preparing it as a deployable website also meant dealing with obscured buttons, layout at different window sizes, and sluggish click responses. The original demo and the later modifications need to be distinguished. The online version includes subsequent fixes; those fixes should not be credited to the original build as problems it had already solved.</p>
<p>That is the distance between a demonstration and a deliverable. A demo shows that a model can turn a goal into an impressive first version. Deployment asks whether a button can actually be clicked, whether content crowds together in a narrow window, and whether an interaction responds promptly. One case makes those tasks visible. It does not establish a success rate for AI projects in general.</p>
<h2>The assumption being challenged may be tool worship</h2>
<p>Saltline still uses Three.js, as does Stonehenge. Fable 5.2 has not eliminated the 3D engine. What it challenges is the belief built around engines and manual modeling: that only someone who has mastered specialist software and crafted each model by hand can create a convincing 3D world.</p>
<p>In my GPT-6 Astra recreation, shape and material were where I felt the difference most directly. Brown's example highlights another capability: not only constructing space, but planning shots, choosing rendering resources, and carrying the scene through a production pipeline. Different examples do not make a rigorous leaderboard. Together, though, they show the depth and range of Fable 5.2's approach to these tasks.</p>
<p>None of this makes modelers, technical artists, or engine engineers unimportant. Production content still needs to be controllable, editable, and consistent, and must pass performance and quality checks. The earlier change is more likely to happen in prototyping: someone with an idea can get a visible or playable version sooner, then decide what deserves further work.</p>
<p>The shift goes beyond the barrier to hand-built 3D models. We have often treated knowing a particular application as if it were creative ability itself. Now a model may write code to build a scene, use Blender for animation, rent GPUs to render it, and call an audio tool for a soundtrack. The tools remain, but a person may no longer have to open and operate each one individually.</p>
<p>What is being lowered is the workflow wall between video, games, drawing, and 3D. Professional creators do not lose their value. But making a first watchable or playable piece may no longer require becoming proficient in four different applications first.</p>
<p>To explore an example, try <a href="https://saltline.gameai.club">Saltline</a>. For the original accounts, see the <a href="https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game">Aiden's Revenge article</a>, <a href="https://x.com/petergyang/status/2119901658394358005">Turbo Tails post</a>, <a href="https://x.com/jkeatn/status/2119874007382299087">Jake Eaton's code-driven painting</a>, and <a href="https://x.com/rileybrown/status/2119902109709537708">Riley Brown's Blender production notes</a>. Times, costs, and production workflows cited here are the creators' own accounts, not independently verified measurements.</p>
]]></content:encoded></item><item><title><![CDATA[AI Boosts Productivity Tenfold. Why Do We Feel More Exhausted?]]></title><description><![CDATA[Introduction I recently watched Fu Peng's interview AI Boosts Productivity Tenfold, which discussed why AI short dramas lose money. One point resonated…]]></description><link>https://bayswong.hashnode.dev/ai-boosts-productivity-tenfold-why-do-we-feel-more-exhausted</link><guid isPermaLink="true">https://bayswong.hashnode.dev/ai-boosts-productivity-tenfold-why-do-we-feel-more-exhausted</guid><dc:creator><![CDATA[黄贝]]></dc:creator><pubDate>Mon, 09 Nov 2026 14:00:56 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/3c333e86-4211-4c06-8c41-357f5802f963.jpg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Introduction</strong></p>
<p>I recently watched Fu Peng's interview <em>AI Boosts Productivity Tenfold</em>, which discussed why AI short dramas lose money. One point resonated with me:</p>
<p><strong>AI greatly raises production efficiency, but under old relations of production, simply raising efficiency does not bring enterprises more profit. Instead, because higher efficiency lowers costs, surplus means of production are poured further into the war for traffic acquisition, leading to still more spending.</strong></p>
<p>This is not a problem with AI. It is a problem with old relations of production.</p>
<hr />
<p><strong>I. The Efficiency Paradox: Why Does Higher Efficiency Bring More Exhaustion?</strong></p>
<p>At the beginning of the year, I was still optimistic about this question. I thought rising productivity would quickly change relations of production. Looking at it now, the process is much slower than I imagined.</p>
<p>Let us use a simple logical chain to understand this paradox:</p>
<ol>
<li>AI tools raise the production efficiency of content/products by 10x</li>
<li>Unit production costs fall sharply</li>
<li>A team that once needed 10 people can now be handled by 1 person</li>
<li>But - where did the saved costs go?</li>
</ol>
<p><strong>The answer is: the fight for attention.</strong></p>
<p>When everyone can produce efficiently, supply explodes, and competition becomes a fight for traffic and attention. The labor cost of the 9 people you saved ultimately goes into ads, promotion, and customer acquisition.</p>
<p>AI short dramas are a typical case. Production costs did fall, but user acquisition costs rose instead. End result: the higher the efficiency, the more exhausted people become, and the harder money is to make.</p>
<hr />
<p><strong>II. The Real Battlefield: Paid Traffic, App Store Listing, and the Rise and Fall of Coze</strong></p>
<p>This is not theoretical deduction. It is what I have personally felt over the past few months.</p>
<p><strong>The paid-traffic dilemma</strong></p>
<p>I recently ran a series of paid acquisition campaigns and clearly sensed:</p>
<ul>
<li>Greater investment in competing for users</li>
<li>Lower conversion rates</li>
<li>Continuously rising cost per acquired customer</li>
</ul>
<p>We were forced to open new channels, such as AI-driven acquisition. But reality is harsh - its effect on paying consumers is limited. This is a forced choice, not an optimal one.</p>
<p><strong>The App Store time sink</strong></p>
<p>App Store listing has now become a waiting game measured in months. First-time listing? Be ready to wait.</p>
<p>This too is an expression of old relations of production: the barrier to production has fallen, but the capacity of distribution channels has not expanded. Supply explodes, channels saturate, and waiting times naturally stretch.</p>
<p><strong>The Coze warning: even big companies cannot rest on past wins</strong></p>
<p>Coze has announced a phased wind-down of its enterprise plans, with new purchases set to pause at the end of November 2026. From its rise and fall you can see that, in this era of intense change, both big companies and small teams find themselves reacting to change.</p>
<p>Resting on past achievements? Unrealistic.</p>
<p>This is destined to be an era where one act leaves the stage and another takes it. Yesterday's winner may become irrelevant today.</p>
<hr />
<p><strong>III. Survival Strategies for Individuals and Small Teams: Diversification and Knowing When to Stop</strong></p>
<p>Facing this environment, I reached several conclusions:</p>
<p><strong>1. Multiple income sources, a portfolio of products</strong></p>
<p>Do not bind your fate to a single product.</p>
<p>The production efficiency AI raises should let you use those productivity gains across multiple products, rather than going all in on one direction.</p>
<p><strong>2. Income should come from multiple sources</strong></p>
<p>Do not expect one product to carry everything. Look for contributions from several products. Each product contributes a share; together they form a stable cash flow.</p>
<p><strong>3. The craft of knowing when to stop</strong></p>
<p>In this era of exploding productivity, <strong>knowing when a product is good enough</strong> challenges every product decision-maker.</p>
<ul>
<li>Do not over-polish</li>
<li>Do not optimize without limit</li>
<li>Validate fast, iterate fast, let go fast</li>
</ul>
<p>MVP is not a philosophy. It is a survival necessity.</p>
<p><strong>4. Small and focused can beat broad and comprehensive</strong></p>
<p>Reach the customers you can realistically serve, and serve the core user group well. That is more valuable than blind expansion.</p>
<hr />
<p><strong>IV. The Truth of Enterprise AI Adoption: Reflections from Failure to Success</strong></p>
<p>This year I had two sharply different experiences with enterprise AI adoption, which gave me a deeper understanding of "relations of production."</p>
<p><strong>Failed consulting: the old SaaS mindset</strong></p>
<p>At the beginning of the year I did two consulting projects. The idea was "start with consulting, then deliver software products." They failed.</p>
<p>Why? Because we were still in the old SaaS mindset: <strong>trying to use AI to improve existing processes</strong>.</p>
<p>That approach does not essentially change relations of production. It only optimizes efficiency inside the old frame. The enterprise's org structure, power structure, and how benefits are distributed stay unchanged. AI is merely a faster tool.</p>
<p><strong>Successful workshop: an empowerment mindset</strong></p>
<p>Mid-year I ran a workshop whose goal was <strong>to help enterprise employees learn to build Agents and solve real problems</strong>. It was very successful.</p>
<p>Where was the key to this success?</p>
<p><strong>Not optimizing processes, but changing relations of production.</strong></p>
<p>Specifically:</p>
<ul>
<li>Empower enterprise employees and unlock their productivity and creativity</li>
<li>Use productivity gains to break through the enterprise's areas where responsibilities are unclear</li>
<li>Break the so-called "departmental silos"</li>
</ul>
<p><strong>Let the people who hear the gunfire not only call for artillery, but also hand-build an Italian cannon.</strong></p>
<p>This is the real value of enterprise AI adoption. Inside an enterprise, externally purchased high-end systems are inferior to hand-mixed black powder.</p>
<p>Why? Because the former is tool substitution; the latter is decentralization of power and reconstruction of capability.</p>
<hr />
<p><strong>V. Core Insight: Changes in Relations of Production Are Slower Than We Imagine</strong></p>
<p>I had already thought about Fu Peng's point at the beginning of the year. I was optimistic then, and believed rising productivity would quickly change relations of production.</p>
<p><strong>But reality is: changes in relations of production are far slower than we imagine.</strong></p>
<p>Why?</p>
<ol>
<li><strong>Organizational inertia</strong>: Most enterprises' power structures, mechanisms for distributing benefits, and KPI systems were designed for old modes of production</li>
<li><strong>Cognitive lag</strong>: Management still understands new tools with old thinking</li>
<li><strong>Path dependence</strong>: Vested interests are unwilling to let go</li>
</ol>
<p>So you see a strange phenomenon:</p>
<ul>
<li>AI raises content production efficiency by 10x</li>
<li>But enterprise org structure does not change</li>
<li>Power distribution does not change</li>
<li>Interest distribution does not change</li>
</ul>
<p>The result: <strong>productivity explodes, but relations of production stay rigid, and the tension between the two keeps growing.</strong></p>
<p>And the cost of that tension is ultimately borne by front-line employees, founders, and small teams - you are more exhausted.</p>
<hr />
<p><strong>Closing: What Kind of Change Do We Need?</strong></p>
<p>This article is not meant to sell anxiety. It is meant to understand the present and find a way out.</p>
<p><strong>For individuals and small teams:</strong></p>
<ul>
<li>Accept that changes in relations of production are slow</li>
<li>Use diversification to hedge single-product risk</li>
<li>Use agility against scale advantages</li>
<li>Stay small and focused; know when to stop</li>
</ul>
<p><strong>For enterprises:</strong></p>
<ul>
<li>Do not treat AI only as a tool; think about organizational reconstruction</li>
<li>Real AI adoption empowers employees rather than replacing them</li>
<li>Decentralizing power matters more than optimizing processes</li>
</ul>
<p><strong>For this era:</strong></p>
<p>We are living through a productivity revolution, but the revolution in relations of production is far from arriving.</p>
<p>This transition will be painful. Many people will feel lost and exhausted. But remember one thing:</p>
<p><strong>Change has already begun. You just have not seen the results yet.</strong></p>
<p>Until then, stay alive, diversify, stay agile, and wait for the day when new relations of production truly arrive.</p>
<hr />
<p><strong>What about you? What have you felt in this efficiency revolution? Feel free to leave a comment and share.</strong></p>
]]></content:encoded></item><item><title><![CDATA[How to Build a Multi-Model Marketing Workflow That Doesn't Lose Context]]></title><description><![CDATA[If you have ever copy-pasted a campaign brief into ChatGPT, gotten a usable draft, then opened Claude to tighten it, then realized Claude has no idea what…]]></description><link>https://bayswong.hashnode.dev/how-to-build-a-multi-model-marketing-workflow-that-doesn-t-lose-context</link><guid isPermaLink="true">https://bayswong.hashnode.dev/how-to-build-a-multi-model-marketing-workflow-that-doesn-t-lose-context</guid><dc:creator><![CDATA[黄贝]]></dc:creator><pubDate>Mon, 29 Jun 2026 17:08:39 GMT</pubDate><content:encoded><![CDATA[<p>If you have ever copy-pasted a campaign brief into ChatGPT, gotten a usable draft, then opened Claude to tighten it, then realized Claude has no idea what the brief said — this guide is for you. The problem is not that any single model is bad. The problem is that the brief lives in your head, not in the tool.</p>
<p>This guide walks through how to wire up a Python script that runs ChatGPT, Claude, and Gemini against the same brief in parallel, keeps the conversation state in a small JSON file, and lets you compare outputs side-by-side without re-typing the brief each time. The pattern generalizes to any workspace that lets you call multiple model APIs.</p>
<img src="https://cdn.hashnode.com/uploads/covers/6a42922e77d8ecf0faca420f/323d3daf-34c7-4f84-83ef-386733e7fc15.png" alt="" style="display:block;margin:0 auto" />

<h2>Why single-model workflows break</h2>
<p>A typical marketing team using a single AI tool runs into three predictable failure modes:</p>
<ol>
<li><p><strong>Brief drift.</strong> The campaign brief lives in a Notion doc. The first draft is generated with the brief in mind. By draft three, the prompt has been retyped three times and the positioning has subtly shifted.</p>
</li>
<li><p><strong>Model lock-in.</strong> Some models write tighter long-form. Some are cheaper for short copy. Some have better non-English coverage. Picking one model means leaving those advantages on the table.</p>
</li>
<li><p><strong>Onboarding tax.</strong> A new teammate reads the brief in Notion, asks the same questions in a fresh chat, and spends a week rebuilding context the rest of the team already has.</p>
</li>
</ol>
<p>These are not problems with the models. They are problems with the workflow around the models.</p>
<h2>The pattern: treat the brief as a persistent object</h2>
<p>The fix is structural. Instead of moving briefs between tools, attach the brief to the workspace as a serializable object. Instead of picking one model, call them in parallel and keep the best phrasing. Instead of re-explaining the campaign to every new chat, store the campaign context once and reference it on every API call.</p>
<p>Concretely, the workflow has four steps.</p>
<h3>Step 1: Lock the brief as a JSON object</h3>
<p>Before you write any copy, write the brief once as data. It includes:</p>
<ul>
<li><p>Positioning statement (1–2 sentences)</p>
</li>
<li><p>Target audience (specific, not "everyone")</p>
</li>
<li><p>Voice and tone (with 2–3 example sentences)</p>
</li>
<li><p>Hard constraints (length, banned phrases, required mentions)</p>
</li>
<li><p>Success metric (what does "this worked" look like)</p>
</li>
</ul>
<p>A working brief file looks like this:</p>
<pre><code class="language-python"># brief.py
brief = {
    "campaign": "Q3 product launch — outline emails",
    "positioning": "AI workspace for marketing teams who refuse to pick one model.",
    "voice": [
        "Direct, no jargon.",
        "Specific over generic.",
        "Confident without marketing-speak."
    ],
    "constraints": {
        "max_words": 180,
        "banned_phrases": ["AI-powered", "revolutionize", "game-changing"]
    },
    "success_metric": "Reply rate &gt; 12% on first send."
}
</code></pre>
<p>Save this once. Every subsequent model call references it. The brief stops drifting because there is no place to retype it.</p>
<h3>Step 2: Run the brief against multiple APIs in parallel</h3>
<p>For most marketing deliverables, you do not want one model's output. You want a comparison. Call the same brief against all three APIs at the same time and read the differences. The first round is not about picking a winner — it is about seeing how each model interprets the brief.</p>
<pre><code class="language-python"># parallel_call.py
import asyncio
from openai import AsyncOpenAI
from anthropic import AsyncAnthropic
import google.generativeai as genai

async def run_brief_against_all(brief: dict, prompt: str):
    """Fire the same brief at three vendors concurrently."""
    oai = AsyncOpenAI()
    ant = AsyncAnthropic()
    genai.configure(api_key=os.environ["GEMINI_API_KEY"])

    async def openai_call():
        resp = await oai.chat.completions.create(
            model="gpt-4o",
            messages=[
                {"role": "system", "content": str(brief)},
                {"role": "user", "content": prompt},
            ],
        )
        return ("openai", resp.choices[0].message.content)

    async def claude_call():
        resp = await ant.messages.create(
            model="claude-sonnet-4",
            max_tokens=1024,
            system=str(brief),
            messages=[{"role": "user", "content": prompt}],
        )
        return ("claude", resp.content[0].text)

    async def gemini_call():
        model = genai.GenerativeModel("gemini-2.5-pro")
        resp = await asyncio.to_thread(
            model.generate_content, f"{brief}\n\n{prompt}"
        )
        return ("gemini", resp.text)

    return await asyncio.gather(openai_call(), claude_call(), gemini_call())



Three outputs, one brief, ~2 seconds wall-clock


results = asyncio.run(run_brief_against_all(brief, "Write a launch email subject line."))
</code></pre>
<p><code>   </code></p>
<p>Common patterns when reading the three outputs:</p>
<ul>
<li><p><strong>Claude</strong> tends to write tighter long-form and respect length constraints more reliably.</p>
</li>
<li><p><strong>ChatGPT</strong> is usually faster on first drafts and better at structured output (JSON, tables).</p>
</li>
<li><p><strong>Gemini</strong> often wins on price-per-call for routine tasks and has good non-English coverage.</p>
</li>
</ul>
<p>The point is not "which model is best" — the point is that you no longer have to pick one.</p>
<h3>Step 3: Synthesize, don't copy-paste</h3>
<p>Once you have the parallel outputs, the actual work begins. Read all three. Keep the phrasing that fits the brand voice. Cut the parts that drift from the brief. Add the angle that none of them saw.</p>
<p>This is the step that AI does not automate. It is also the step that distinguishes a marketing team from someone using ChatGPT as a typewriter.</p>
<h3>Step 4: Turn the recurring motion into a reusable workflow</h3>
<p>The first time you run this pattern, it is a project. The third time, it is overhead. The fifth time, it should be a saved workflow.</p>
<p>A reusable workflow includes:</p>
<ul>
<li><p>The locked brief template (the JSON file above)</p>
</li>
<li><p>The parallel-call script (parameterized on <code>model</code> list)</p>
</li>
<li><p>The synthesis checklist (which model tends to win for which deliverable)</p>
</li>
<li><p>The publishing checklist (where this output goes next)</p>
</li>
</ul>
<p>When the workflow is saved, onboarding a new teammate is a workflow handoff, not a 30-minute Loom.</p>
<h2>Benchmark: what the three models actually cost</h2>
<pre><code class="language-markdown">| Model | Avg tokens / call | Cost / 1k calls (USD) | Avg latency (s) |
|-------|------------------:|----------------------:|----------------:|
| `gpt-5.5`           | 850 | $4.25 | 1.4 |
| `claude-sonnet-4.7`  | 920 | $4.60 | 1.8 |
| `gemini-2.5-pro`   | 780 | $2.34 | 1.1 |
</code></pre>
<p>BYOK setups typically reduce per-token cost 30–60%, depending on which tier you negotiate with each vendor. If you mostly run short copy, Gemini wins on cost. If you mostly run long-form with tight constraints, Claude wins on consistency.</p>
<h2>Common pitfalls when adopting this pattern</h2>
<p>A few things that go wrong, and how to avoid them:</p>
<ul>
<li><p><strong>Locking the brief too late.</strong> If the brief is saved after the first draft, it is a justification, not a constraint. Lock it before any generation.</p>
</li>
<li><p><strong>Comparing model outputs without a rubric.</strong> "Which is better" without a checklist produces inconsistent picks. Decide the rubric once, then apply it.</p>
</li>
<li><p><strong>Skipping the synthesis step.</strong> Copy-pasting the best model's output verbatim is not a workflow — it is a single-model workflow with extra steps.</p>
</li>
<li><p><strong>Saving the workflow too early.</strong> Save it after the third or fourth run, when the shape is clear. Saving after the first run locks in the mistakes.</p>
</li>
<li><p><strong>Ignoring provider rate limits.</strong> All three vendors have separate rate-limit tiers. <code>asyncio.gather</code> is fine; <code>gather</code> of 1,000 concurrent calls against the same API key is not.</p>
</li>
</ul>
<h2>When this pattern is not the right fit</h2>
<p>If your marketing output is genuinely one-off — a single launch, a one-time campaign, a single article — the overhead of a multi-model workflow is not worth it. Use one model, ship the work.</p>
<p>If your team ships every week across multiple channels — this pattern is the difference between chaos and a system. The first month is overhead. The third month is leverage.</p>
<h2>Discussion</h2>
<p>Have you tried running multiple models in parallel? What does your brief-as-object look like — JSON, YAML, Markdown frontmatter, something else? Drop a comment with the structure that worked for you and what you would change.</p>
<p>I am especially curious whether anyone has tried this pattern with local models (Llama, Mistral, Qwen) alongside the API vendors. The cost profile changes a lot once one of the three legs is local.</p>
]]></content:encoded></item></channel></rss>