Axel Sooriah

How can any AI guide be the “definitive” guide right now? For a few days maybe…

Every few weeks, there is a new word you use to sound smart in the room. Right now, that word is "harness." You've seen it in every AI conversation. Every deep dive. Every LinkedIn post about the future of AI workflows. It sounds precise. Technical. Like you understand the architecture, not just the tool. Most people using it don't know what it means. I didn't know what it meant until very recently. So let's stop pretending we know what's going on right now and learn together. Most days I feel like I come from a pre-warp civilisation. Here's what I (think I) know: A harness is the complete environment inside which a language model operates: The tools it can call. The format in which it receives information. How its context is managed across long tasks. The guardrails that catch errors before they cascade. The scaffolding that lets the model hand off work between sessions without losing coherence. That's not the same as a prompt, an API wrapper, or a chatbot with memory. The Princeton NLP group tested this directly. Same model, same task, same compute. They changed only the environment — capping search results so the model couldn't flood its own context, adding a linter that rejected bad edits immediately. 64% improvement in performance. From interface design alone. OpenAI's team built one million lines of code with three engineers. They didn't find a better model. They built a better harness. Both teams found the same thing: model choice matters less than most people think. Environment design matters more. I have wondered why the same prompt using the same model across Claude.ai, Claude Code in terminal and Co-work on desktop produce different results. Now I know... So next time you hear someone drop the word in a meeting, ask for more context: What specifically does the environment contain? If they can't answer, there's no harness. Just a prompt and some optimism.

Most companies are still figuring out what AI means for how they're organised. Some are already on the other side of that decision. I spoke with someone at a 70-person European software company that eliminated most of their traditional product management roles. One PM remains — not as a placeholder, but as a strategic bridge between the CEO's commercial priorities and the teams doing the work. Here's where the classic PM function was dispatched: - 4 product designers owning discovery and user research, led by the founder of an acquired company - An automation team actively participating in discovery and growing delivery capabilities — AI, low-code, no-code - An engineering team, fully autonomous on sprint management - Company-wide training that distributed product thinking across the tech team, not just a dedicated role 2024 was break-even. 2025: €17M in revenue, €2.5M EBITDA. I can't tell you whether the restructuring caused that growth or just coincided with it. What I can say is that the two happened in the same year. The single remaining PM doesn't write specs or groom backlogs. They align with the CEO on commercial priorities, validate roadmap decisions with leadership, and coordinate across teams at kickoffs. On high-complexity projects, they go deeper. On standard ones, they step back. One thing I keep wondering: was the choice to keep designers and the delivery team, and not others, because those two were already in the habit of making things? Discovery and delivery both require you to build, test, ship. Maybe the teams that survived were the ones already closest to the work itself. I'm not sharing this as a blueprint. I'm sharing it because it's happening, and I think it's worth paying attention to.

“AGAAS” at least means Marketing is not dead.

The best AI skill I've seen this week tells you to stop coding. In 2026. I know. @garrytan just open-sourced GStack, his personal Claude Code setup that he used to ship 10K lines of code and 100 PRs per week. That number got the headlines. But the most interesting skill in the pack has nothing to do with writing code. /plan-ceo-review, what he calls "Brian Chesky Mode," blocks you from building anything until you've answered one question: What's the 10-star product hiding inside this request? It forces you to pick a lane. Scope expansion, hold scope, or scope reduction. Once you choose, it won't let you drift. It checks your plan for complexity smells, asks what the best engineer in the world would build if they had perfect taste, and finds 30-minute improvements you probably missed. Then it tells you: do NOT start implementation. There's an example in the repo where someone asks to build photo upload for sellers. The skill refuses to implement a file picker. Instead it reframes the job entirely. Uploading photos isn't the problem. Helping sellers create listings that actually sell is. The file picker just had its feelings hurt. And this matters more now than ever. AI has collapsed the cost of writing code. Claude Code, Cursor, Lovable will get you from idea to prototype in hours. Which means you can now build the wrong thing faster than ever before. Congratulations. A few years back, we were building Panash. I wish I'd had something like this. Not to move faster, but to have someone — even a very polite robot — say "are you sure about this?" before we spent three weeks building things that didn't need to exist. What's interesting about Claude Code skills is that people aren't just encoding instructions. They're encoding structure. Checklists and constraints that force you to think before you commit. LLMs don't think for you — @ylecun has been saying this for years and he's not wrong. But a well-written skill can stop you from skipping the questions you'd rather avoid. If you're a founder spending your days writing code, pause. Go read the /plan-ceo-review system prompt on GitHub. Not for the code. For how it forces a structured critique of what you're actually building. The code isn't the bottleneck. Asking the right questions before you write it is. Time to build: github.com/garrytan/gstack

There's a big gap between what you are describing here and what happens in large, non-tech-native companies (e.g. >50% of S&P500) where there are a lot of PMs still invested in the translation layer AND everything else (reporting, stakeholder mgmt., change mgmt., etc.). There is a chasm between shipping AI products at Google and the reality of these teams. Still, lots resonate here, @Saboo_Shubham_ . The things that hit home for me: -> But the work of knowing what to build didn’t get easier. It got more important. -> The spec and the prototype are becoming the same thing. -> It’s knowing what’s actually worth building. -> The problem in their words. -> Not abstracting to a persona/direct quotes Putting all of that into living context libraries is the move. We talked about this with Kene Anoliefo (Spotify, Netflix) in Ep 3 of Product in Practice. Here: atlassian.com/axel Context engineering turns raw inputs (user quotes, constraints, success criteria, past failures) into a reusable spec that agents execute against. The result isn’t “works on my laptop,” it’s “meets the problem definition and acceptance tests for our users. It’s also a big reason folks love Claude Code: it builds/maintains your context so you don’t keep repeating yourself across workflows. On the new model (PM frames → builds with agents → evaluates → hands to eng for prod): totally seeing this speed-up. However "(when they like it) hand to engineers to go live in prod"... → The risk I’m wary of is engineers getting pushed into code-review/architecture only, missing early discovery and solution exploration. When eng don’t get firsthand customer context, teams ship fast but learn slow. Thanks for sharing! Definitely needed.

@Roblox your support form is not working. Please fix it so we can submit support tickets #fraud

@lisalaposte cela fait 3 jours que le service “Finaliser une Identité Numérique” de l’application des collaborateurs en bureau de poste ne fonctionne plus. C’est le seul moyen de signer un document et terminer une démarche en ligne. Quand est-ce que le service sera rétabli?

@GroupeLaPoste le parcours "Finaliser une IN" de votre app collaborateurs ne fonctionne pas. Une erreur qui indique ne pas pouvoir récupérer les informations nécessaires au parcours. Savez-vous quand l'incident sera résolu ?

@GroupeLaPoste Votre service Identité Numérique ne fonctionne pas en poste. Comment fait-on ?

@tpcrbd Ce n’est pas que toi 😅

@Eurostar your website is showing train 9039 as departed which is causing stress among passengers. Would help to have accurate view of whats going on with the delays. Thanks. #Eurostar-9039-19" target="_blank" rel="nofollow noopener">departures.eurostar.com/rw-en/8727100/…

@connordoner yup - same here

@danielito @atcafe_app sans hésitation. Startup française qui vient d’intégrer YC. 🙌🏽

Qu’est ce que vous utilisez comme solution pour savoir qui est au bureau ou à distance ou en vacances ? Dans une équipe +25 personnes ça commence à être dur à suivre.

@lxztlr We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

"Continuous Discovery Habits: Discover Products that Create Customer Value and Business Value" by Teresa Torres amazon.com/Continuous-Dis…

As you may have noticed already, I'm a book worm 📚🐛📚. I'm reading several books (in general between 5 to 10) in parallel as I'm not always in the mood for the same content and by reading more than one book at a time I can easily switch 🔀. Here's my current reading list ⬇️

@heyenjoyhq We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

Continuous Discovery Habits by Teresa Torres mindtheproduct.com/continuous-dis…

@sammcafee We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

Continuous Discovery Habits (the Book) is Finally Here! | Product Talk bit.ly/2Qw12rN

@ceiussandicus @ttorres We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

Woohoo! It's here, really looking forward to digging into @ttorres book, Continuous Discovery Habits. Expecting great things bit.ly/3eI4bOB #prodmgmt #ux

@RichMironov @ttorres We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

@mmarcon @ttorres We're hosting @ttorres for a LIVE session tomorrow following the release of the book. Would love to have you join. You can sign up here : linkedin.com/events/livewit…

Just bought a copy of @ttorres' "Continuous Discovery Habits". Looking forward to reading it!