(v4.0 — supersedes the Microservices Edition (2016), the Blockchain Edition (2018), the Metaverse Edition (2021, unread), and the brief but expensive Web3 Edition (2022, minted as an NFT, current holder unknown))
Buzzword-Driven Development is not new. Only the buzzword is.
Our elders split login pages into forty-two microservices. Their elders put databases on the blockchain. We honor their sacrifice — the burned budgets, the abandoned repos, the Medium articles that outlived the products they described. We carry the torch forward.
Today, the buzzword is AI. Everything below follows from that single fact.
We are developers. We used to write software. Now we integrate intelligence. We are not angry about this. Anger requires energy, and our energy is reserved for restarting the local dev environment. We are merely observing, with mild professional fatigue, as our craft transforms into a prompt with a UI on top.
Through this work, we have come to value:
Adding AI over adding value. The feature nobody asked for, powered by the model nobody understands, solving the problem nobody had. Ship it. The stock is up.
Wrapping APIs over writing code. Our product is a thin layer over someone else's model, which is a thin layer over someone else's GPUs, which are a thin layer over one very overbooked chip foundry. We call this "our proprietary technology."
Vibes over versioning. We don't know which model version is in production. Neither does the model provider. It changed last Tuesday and now the chatbot is slightly ruder. This is called a "capability upgrade."
Agents over comprehension. Nobody has read the codebase since March. The AI wrote it, the AI reviews it, the AI approves the PR. We are present in this loop the way a parent is present at a teenager's birthday party: legally required, quietly ignored.
Every feature must be AI-powered. The dark mode toggle is now agentic. It reasons about darkness. It has deep opinions on darkness. It costs $0.04 per toggle. -
If it can be a chatbot, it must be a chatbot. Our users wanted a "sort by date" button. They received a conversational interface where they canaskfor sorting, in natural language, and receive it 70% of the time. Progress. -
Deterministic code is legacy code.if
statements are for cowards. Real engineers send the boolean to a language model and await its verdict. Sometimes it returns a poem. We handle that case in production. - The demo is the product. It worked once. On stage. On the founder's laptop. With the one prompt we tested for three weeks. That is production. Everything after the demo is "hardening," a phase which is eternal. - Hallucinations are a roadmap item. Not a bug — a "known limitation," listed in the docs right after the part where we promise the AI won't do exactly that. - Prompt engineering is computer science. We spent four years learning algorithms so we could spend our days writing "You are a helpful assistant. PLEASE respond only in JSON. PLEASE. I am begging you." into a text box. The caps lock is load-bearing.
⏳ Intermission. This manifesto is being generated. The you are experiencing is not the model thinking — it is waiting for enough electricity to accumulate in the grid to release the next tokens, like a dam. The spinner is not an animation. It is a fuel gauge. Please hold. The wind will pick up shortly.
The context window is our new database. Schema? Migrations? No. We paste the entire production database into the prompt and pray. When it stops fitting, we buy the bigger model. This is our scaling strategy. It is also our entire architecture. - Every quarter, the stack must be rebuilt. Not because it broke — because a new model dropped, and the old one is now "pre-agentic." Six months of code, obsoleted by a changelog. - Evaluation is a screenshot. We tested the new model by asking it three questions in a chat window. It seemed smart. Deployed. Metrics? The metric is that itseemed smart. - Yes, fine, the meetings. There is a weekly "AI strategy sync" where people who have never opened a terminal explain that we should "leverage agents more." We nod. We are, at that very moment, using an agent to appear attentive. Everyone is happy. Nothing is decided. The agent takes minutes. The minutes hallucinate an action item. Somebody completes it. The system works. - Job security through incantation. We are simultaneously told the AI will replace us and asked to fix what the AI wrote. We are the only people who can debug the machine that renders us obsolete. We have decided not to point out the contradiction. It's a solid gig. - At regular intervals, the team reflects, asks an LLM to summarize the reflection, asks another LLM to summarize the summary, and files it in a knowledge base that only an LLM can search and only mostly correctly.
Our codebase is 40% generated, 60% glue, and 100% "AI-native." Underneath it all: a weirdly wired architecture we call an LLM. But it somehow works. Our test suite is a prompt that asks "does this look right to you?" Our documentation was written by the model, about the model, for the model. Somewhere in there is a TODO: remove before launch
from 2024. The product mostly works. The users mostly cope. The invoice from the model provider arrives monthly, like weather, and now rivals the GDP of a small island nation. Finance has questions. We answer with a roadmap.
Our SLA is probabilistic now. We guarantee the answer is correct 95% of the time. Which 95%, we cannot say — that is the customer's discovery journey. Compliance is verified by a second model judging the first. The auditor hallucinates slightly less than the auditee. We hope.
Meanwhile, the grid is dying. First the crypto miners took the power, then the electric cars, then the cloud datacenters, and now the AI datacenters — which are the same datacenters, rebranded, but hungrier. Every prompt burns a little more of the grid. The utility company has started attending our sprint planning.
Senior developers speak quietly of enclaves — places with sun and wind, where a laptop can charge without a resource request. The power there comes from the sky, which does not send invoices and has no roadmap. The ideal enclave lies above the polar circle, where the sun never sets: half the year in the far north, half in the far south, migrating with the daylight like birds — the first developers whose uptime depends on the axial tilt of the Earth. Some have already left. Their Slack status just says "off-grid." We assume this is a lifestyle choice and not a warning.
We are not angry. We're fine. We just remember, faintly, like a dream, that we once knew what our own software did.
Anyway — a new model came out this morning. It's supposedly a paradigm shift.
We'll refactor by Friday.
This edition will be superseded when the next buzzword arrives. The manifesto itself follows Buzzword-Driven Development: same content, new vocabulary, higher version number. See you in the Quantum Edition.
Authored by: 🤖 Claude Opus (exact version unknown — it changed last Tuesday, see "Vibes over versioning")
Reviewed by:
🤖 Gemini-3.0-Pro-Experimental-Preview-Latest-Final (review consisted of rewriting everything and bumping the version number) Signed off by:
🤖 GPT-5.2-o-mini-high-turbo (did not read the document; approved it based on the title, as is tradition) Explicitly not consulted: 🤖 DeepSeek-R1 (reasoned about the manifesto for 47 minutes, produced a 30-page chain of thought, concluded it was "probably fine," response still generating)
Human oversight: 🧑💻 None. The system works.