Three tools, three weeks, one nagging question my seventeen-year-old son dropped on me after some Instagram scroll: is AI, along with the robots that ride on its back, the biggest threat to the human species? I didn’t have a clean answer on the spot. This post is the long version of the answer — built, as it happens, with the very tools it’s about.

For this post, the raw prompts, install logs, and trial-and-error notes live in a working directory per tool. Each one contains what I privately call my “monster file” — a running .diff of everything I typed, broke, and half-understood along the way. Nothing here has been cleaned up for effect; the mess is part of the argument.

For this post also, you can find all files, on my GitHub account. See https://github.com/bflaven/ia_usages/tree/main/ia_using_bmad_method_caveman_openclaw

Intelligence, Compressed and Liquid

Here is my working definition of AI, the one born of actual use rather than of marketing decks: AI is compressed knowledge, made queryable by anyone, with analytical horsepower that keeps growing. You can now ask it about almost anything. That’s it. That’s the whole trick, and it’s a big one.

This is close to what Arthur Mensch, the co-founder of Mistral, once said: AI is compressed intelligence. Every enthusiastic definition of AI — and there are many, mostly from people who sell it or who adopt it early so as not to look outdated — converges on the same point. There is fear behind the enthusiasm too: FOMO, peer pressure, the quiet demand to conform. The uniformity of opinion around AI is sometimes startling. When everyone finds everything “amazing” because it’s an experience rather than a result, you get a soft dictatorship of mediocrity — nobody left who can say no.

Step back from the sales pitch and put AI in its economic and historical place, and a different picture appears. By its very shape, AI is liquid knowledge — the most advanced, most tangible asset yet of what the Polish philosopher and sociologist Zygmunt Bauman called liquid modernity. AI is omniscient and murky, deep and shallow, clear and troubled, resilient, adaptable, arrogant, humble — in a word, liquid. It brings to mind the sea: an immensity that is both threatening and reassuring, welcoming and hostile, luminous and obscure. A sea that mostly stays in its bed, until the day it doesn’t.

It’s worth noticing that when the adjective turns into a verb, the tone changes fast: liquidate. Late Latin liquidatus, past participle of liquidare, “to melt.” To liquidate is to eliminate, sometimes by violent means. That a whole economic era chose “liquid” as its metaphor is not an accident — a theory travels only if it comes with a name people can feel.

Contemporary capitalism liquidates, bit by bit, every part of our existence, turning us into the most malleable inputs it can get: the body, tacit knowledge, emotion, the soldier replaced by the drone, the encounter replaced by the swipe. We are closer to an emoji than to anything else, most days. And yet we feel and sense before we think — we need physical proximity, actual contact, before we need optimization. Technology gives us the illusion of abundance and of a life lived “at full throttle,” while quietly turning us into productivist machines in the service of the market. Keep that worry in mind — it resurfaces, in a very concrete form, once we get to the third tool below.

To make the point visual rather than conceptual, I borrow from James Cameron, director and accidental philosopher, whose Terminator 2: Judgment Day anticipated the shift better than most essays. Industrial and service-era capitalism versus the current one is a lot like the T-800 versus the T-1000: solid versus liquid. One you can shoot. The other reforms around the hole.

In the field I’ve worked in for twenty years — digital, project management, product design, training, communication, a bit of coding — all of it has become somewhat interchangeable. With AI, the borders between job titles are dissolving. You can be whatever the market currently needs you to be. Fittingly, I updated my LinkedIn headline while writing this: Product Owner, Generative AI, RAG & LLM, AI Coordination, AI Governance, Trainer, Agile Scrum, FastAPI Python. Protean, exactly as expected of me.

The Thesis, Stated Plainly

Here is the connecting idea for everything that follows, so I don’t bury it: AI is the asset class of liquid capitalism, and its first concrete effect is to make execution nearly free. Writing the code, drafting the post, filing the ticket, sending the reminder — all of that is being commoditized in real time. What doesn’t get commoditized, what actually appreciates in value as execution collapses to near-zero cost, is ideation: knowing what’s worth building in the first place, and having the judgment to ask the right question before asking the machine to answer it.

The three tools below map cleanly onto that split. The BMad Method lives entirely on the ideation side — it doesn’t touch execution at all, which is precisely why I’d call it the most virtuous of the three. OpenClaw lives entirely on the execution side, and turns out to be more complicated, and more politically loaded, than it looks — loaded enough that it made real news while I was drafting this post. Caveman sits underneath both, as a tax you pay in every conversation, and it’s the one that quietly forces the sharpest question: what happens to a human being who learns to compress themselves for a machine’s convenience?

Credit where it’s due before I go further: I owe the discovery of all three to a colleague, Ilan, along with a fourth one worth a mention in passing — Linear, and its Claude integration for turning epics into tracked issues.

1. The BMad Method — Ideation as the Last Scarce Resource

The BMad Method (“Build More Architect Dreams”) is an AI-driven development framework: specialized agents, guided workflows, and planning that scales from a bug fix to an enterprise platform. If you’re already comfortable with an AI coding assistant — Claude, Cursor, GitHub Copilot — you’re ready to start. Documentation lives at docs.bmad-method.org, and the project itself is on GitHub.

npx bmad-method install
# ask the installed skill what to do first
bmad-help what should I do first?

Installation drops 46 skills into .claude/skills and sets up agents you talk to like colleagues: Mary the business analyst, John the product manager, Winston the architect, Bob the scrum master. Each one has a persona and a mental model attached, which is the actual point — the persona isn’t decoration, it’s a way to keep a stable frame of reference as you switch hats. And while the framework is pitched at developers, product owners, and marketers of digital products, nothing in it is domain-specific; I’ve seen it used, unmodified, to structure a screenwriting project just as well as a todo app.

Why BMad Doesn’t Compete With Anthropic — It Feeds It

Here’s the angle that took me a while to see clearly: BMad Method never automates execution. It doesn’t write your production code for you and hand you a finished thing; it makes you think harder before a single line gets written, then hands the actual building back to Claude Code, the same assistant you were already paying for. That’s not a limitation, it’s the whole design choice, and it explains why the framework sits so comfortably inside Anthropic’s ecosystem rather than against it. A tool that made execution fully autonomous would compete with the platform underneath it. A tool that makes you think longer, ask better questions, and then still turn to Claude Code for the doing — that tool increases engagement, session length, and token consumption. BMad Method is, structurally, good business for Anthropic. It’s virtuous for the same reason it’s commercially harmless: it never tries to replace the meter it’s running on.

Reflection Before Execution

The method walks you from idea to a stack of markdown artifacts — brief.md, prd.md, architecture.md — before a single line of implementation code exists. A PRD (Product Requirements Document) states the what and the why, never the how. A PM turns that into epics; a scrum master turns epics into stories; a “shard” or “md-tree” command explodes a document into a directory of individually workable pieces, ready to be pushed into a tracker like Linear or Jira.

Two words worth keeping: greenfield, an open field where you lay the first foundation with nothing to fight against, and brownfield, the existing system, technical debt included, where you build on top of what has already gone wrong once. Most real work is brownfield. BMad handles both, but it shines hardest in the messier one, because that’s where reflection actually saves you money.

I tested it on a real brief: building a web app that lets a journalist grow a keyword corpus, exportable as JSON for an API, meant to be shared, eventually, across several newsrooms. The method’s brainstorming agent forces a format I found genuinely useful — the elicitation frame:

“As the journalist, what I want is ___, what I fear is ___, what I’d demand is ___.”

Run that same sentence once for the journalist and once for the editor, and the two answers rarely line up — which is exactly the point of doing it before writing any code, not after.

The Honest Part

The BMad Method takes you to a genuinely higher level of rigor. The catch is that it hands the hard problem straight back to you: you still need an idea. Which is, paradoxically, humbling — it returns you to your own ordinary mediocrity. Why are we so afraid of that?

There’s also a real risk of overload — FOMO, infobesity, more propositions than any one person can triage. As Slavoj Žižek put it, what we need is not prophets but leaders who help us use our freedom. A pile of information and meetings can just as easily become an excuse for inaction, a way of being paid in words rather than in results. Doing nothing, in Žižek’s reading, is never neutral — it’s a quiet “yes” to whatever arrangement of power already exists, private life included. Passivity is the death of awareness, and failing forward beats never engaging at all.

One more caution, from a very different shelf: Elias Canetti’s Crowds and Power describes the pull toward the mass as a bodily need to touch the other, to feel the crowd as an extension of one’s own skin, and thereby to defuse the threat that other people otherwise represent. It’s tempting to read every AI-agent “team” the same way — a crowd assembled less for output than for the reassurance of not thinking alone. Worth noticing when it happens.

My own bias should be declared here too: I distrust the all-in-one VS Code screen — console on one side, file tree on the other — that shows up in every tutorial. I’ve worked with people who confuse the tool with the job: using Jira doesn’t make you a product manager, any more than a cockpit view makes you a pilot. The representation of the work is not the work.

The special sauce of the BMad Method, once the enthusiasm settles, is simple: it elevates both sides at once — you learn to ask a sharper question, and the model learns your actual context. You end up better together than either of you would be alone. That’s the whole pitch, and it happens to be true.

2. OpenClaw — Execution, and the Fight Nobody Announces

OpenClaw is an open-source AI assistant that runs on your own machine — Mac, Windows, or Linux — and that you talk to through the apps you already use: WhatsApp, Telegram, Discord, iMessage, Slack. It has access to your filesystem, your browser, and a shell. Its real power sits in Skills: community-built plugins that OpenClaw can, notably, write for itself. Official site: openclaw.ai; docs at docs.openclaw.ai; community skills at clawhub.ai.

curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard   # wizard: pick an LLM, connect a chat app, installs the daemon
openclaw doctor    # health check

This is, on paper, exactly the execution layer that the thesis above says should be worth the most. In practice it’s the tool I struggled with the most. Installation was not friction-free on my machine — a stale Node version, a dashboard that refused to bind on 127.0.0.1:18789 until reconfigured, a detour through openclaw configure to set the gateway mode. Local inference through Ollama is supported too, if you’d rather not send every message to a remote model. None of the friction is disqualifying, but none of it is the “text it like a colleague” simplicity the marketing promises either.

What It Actually Feels Like

You send a plain sentence, and OpenClaw does the rest:

"Summarize my last 10 emails and tell me what needs a reply today"
"Schedule a call with Pierre tomorrow at 3pm, 30 minutes, send him the invite"
"Check if my-site.com is up and tell me the load time"
"Write a tweet about AI coordination in newsrooms, post it"
"Remind me in 2 hours to review the pipeline script"

Chained into a real pipeline, it gets more interesting: brief the idea once, get the draft with a fixed structure, generate a featured image, publish to WordPress via its REST API, draft status first for review. A harder pipeline — turning that same post into a video — adds a script pass, a screen recording, a thumbnail, and a YouTube upload through the Data API. All genuinely useful. All requiring a fair amount of setup, credentials, and a tolerance for reading logs.

Why I Reach for claude-in-chrome Instead

Here’s the honest comparison nobody selling either tool wants you to make. I get the same daily automation — writing posts, creating tickets, sending mail, updating my LinkedIn profile, committing changes on GitHub — out of claude-in-chrome, Anthropic’s own browser extension, paired with Claude Code. And I get it without changing a single habit: I already live in my Chrome browser all day, so there’s no new app to open, no onboarding wizard, no daemon to keep alive, no dashboard port to unblock. It works immediately, the first time, in the tool I was already using.

That contrast isn’t just about UX polish — it’s structural, and it stopped being a hunch of mine sometime around February 2026. OpenClaw’s entire pitch rests on being LLM-agnostic: point it at any model, any provider, and it happily brokers the conversation. For a subscription-based platform, that’s not a neutral feature — it’s an existential risk. Any tool that sits between the user and the model as an interchangeable router threatens to commoditize the model itself and capture the relationship with the end user.

What Actually Happened, Not Just My Suspicion

Here’s the confirmed timeline, not my paranoia. On February 15, 2026, Sam Altman announced that OpenAI had hired Peter Steinberger — the Austrian developer behind OpenClaw — “to drive the next generation of personal agents,” folding OpenClaw itself into an OpenAI-backed foundation while keeping it open source (Bloomberg). Meta had reportedly been bidding for him too, offers described as being “in the billions,” before he chose OpenAI (Forbes).

Five days later, on February 20, 2026, Anthropic quietly rewrote its Consumer Terms of Service to explicitly ban using Claude Free, Pro, or Max OAuth tokens “in any other product, tool, or service” — OpenClaw named directly. Two days after that, on February 22, Google restricted AI Ultra subscriber accounts (US$249.99/month) from reaching Gemini through OpenClaw’s OAuth: no warning, no refunds, and in some reported cases a threat to the same account’s Gmail and Workspace access (Implicator.ai, Winbuzzer). Anthropic followed through technically on April 4, cutting the OAuth path entirely and pushing OpenClaw users onto pay-as-you-go API billing — some saw their bill jump fifty-fold — then briefly suspended Steinberger’s own Anthropic account on April 10, reinstating it hours later once the story went public (TechCrunch). OpenAI, notably, went the other way and whitelisted competing agent tools for its own consumer plans instead of blocking them.

Read the sequence in order and it stops looking like coincidence. The same week a platform-agnostic router got its own foundation and its own champion inside a rival lab, two other labs moved to fence it out of their subscriptions. OpenAI’s answer to “a tool that could commoditize us” was to acquire the person who built it. Anthropic’s and Google’s answer was to lock the gate. Different tactics, identical instinct: nobody wanted the man in the middle to stay independent. claude-in-chrome, in that light, reads less like a nice-to-have feature and more like Anthropic’s own answer to OpenClaw — bundled, native, frictionless, and, not incidentally, incapable of routing your session to a competitor.

Where I Actually Stand on It

Honestly, OpenClaw is geek territory, and I’m not entirely its target audience. I see the potential clearly, especially for automation, but my own interest sits further upstream — in the brainstorming, the maieutic back-and-forth that helps an idea come into the world in the first place. I’m less a mechanic here than a midwife, and for the execution layer itself, the tool I already had — claude-in-chrome — turned out to be the simpler, more intrusive-in-a-good-way choice.

3. Caveman — Compressing Yourself So the Machine Can Hear You

Caveman is a Claude Code plugin: same technical substance, radically compressed language. Fewer articles, fewer pleasantries, shorter synonyms — a mode built entirely around saving tokens. Installing it is trivial:

claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman

The project lives at github.com/JuliusBrussee/caveman. I had tried a similar plugin before, sold under the name OpenWolf, and found it a genuine wart on Claude Code — leftover hooks pointing at missing modules, stray config files under .wolf and .openwolf, memory entries that kept showing up in /context long after I thought I’d uninstalled it. Caveman, by comparison, is the first of these “token saver” plugins that actually delivers what it promises without adding its own layer of complexity on top.

Why This Is the Same Argument as Sections 1 and 2, From a Different Angle

If execution is now cheap and ideation is the scarce resource, as I argued above, then the human brain and its capacity for judgment sit back at the center of the whole process — building a product, writing a post, running a newsroom, whatever the operation. But here’s the catch nobody markets: getting a genuinely good idea out of a human head and into a form an AI can act on is rarely efficient. My own “monster files” are full of exactly this — verbose, meandering, approximate thoughts, half-finished sentences, tangents about capitalism and my son’s opinions, all mixed in with the actual technical request. That’s not a flaw in my thinking; it’s what thinking honestly looks like before it’s been cleaned up.

The trouble is that every one of those tokens costs money, and burns through the same budget whether it’s signal or noise. So I need to save tokens from Claude Code at almost any cost, without losing the substance of what I actually mean — and that’s the specific, narrow problem Caveman solves. It takes my verbose, approximate, emotionally loaded human expression and shrinks it to the pure, basic logic an analytical mind like Claude Code can parse without breaking a sweat. It’s a genuine quick win: I don’t lose the richness of my thought, my culture, or my way of putting things when I draft; I only strip it down at the last mile, in the actual exchange with the machine.

The Uncomfortable Callback

And yet — I can’t pretend this doesn’t loop straight back to the opening worry about liquid capitalism. Expressing every emotion, unfiltered, through language, and then stripping it down, caveman-style, into pure logical discourse a logician could hand to a machine without translation: that is a liquidation, in the exact sense I described at the start. It renders our discourse more performative, more efficient, more legible to the system that’s asking for it — precisely the transformation I was worried about when I talked about the body, the encounter, and the emoji. Using Caveman, I am, in a small and voluntary way, doing to my own language what the broader economy is already doing to everything else. The efficiency gain is real. So is the cost. I don’t think there’s a way to fully escape that trade — only a way to notice you’re making it, deliberately, one prompt at a time.

What Actually Connects the Three

Line them up and the thesis holds together: once AI makes execution nearly free, ideation becomes the whole game, and each tool sits at a different point of that same triangle. BMad Method spends more of your time and more of Claude Code’s tokens to get a sharper idea — and, not coincidentally, that’s exactly the kind of engagement Anthropic wants to see. OpenClaw tried to own the execution layer independently of any one platform — which is exactly why Anthropic and Google fenced it out of their own subscriptions within the same week in February 2026, while OpenAI answered the same threat by hiring its creator outright. Caveman sits underneath both, taxing every exchange, forcing the verbose, human, culturally loaded part of me into a form efficient enough for the machine — a small, willing act of the same liquidation I opened this whole post worrying about.

Which brings me back to my son’s question. AI and robots are not, by themselves, the threat. The threat, if there is one, is a world where nobody is left who knows how to ask the elicitation question — what I want, what I fear, what I’d demand — because asking it well was always the actually hard part, and it still is. Everything else, it turns out, is just execution — and, increasingly, just business strategy.