iiinova — where your digital self keeps evolving

use case driven @tpierrain.bsky.social · 27/09/2026

Quand les décoloniaux de @parolesdhonneur.bsky.social décryptent mediapart, ça donne une série inégale mais globalement très intéressante. Mention spéciale pour l’épisode sur Plenel 🍗 bsky.app/profile/paro...

000

use case driven @tpierrain.bsky.social · 16/09/2026

Retour au quartier, après quelques années d’absence. 🥹

040

use case driven @tpierrain.bsky.social · 09/09/2026

Au programme : mechanical sympathy, modèles mentaux, l'importance d'expliquer pour comprendre, forcément un peu d'IA, et pourquoi les sujets humains sont bien plus durs que les sujets tech. Merci Bruno Soulez 🙏 liens vers différents formats : www.ifttd.io/episodes/arc...

ifttd.io

IFTTD #374 - Architecture : Pourquoi les sujets humains font toute la différence avec Thomas Pierrain

"les sujets tech c'est de la rigolade par rapport aux sujets humains." Dans l'épisode 374 j'ai reçu Thomas Pierrain pour parler de "Architecture"

020

use case driven @tpierrain.bsky.social · 09/09/2026

« Si tu ne peux pas expliquer en 30 secondes le code que tu as généré, tu commences à empiler de la merde. » Une heure chez If This Then Dev, enregistrée en juin, sur les fondamentaux qui restent quand le code devient une commodité. ⬇️

131

use case driven @tpierrain.bsky.social · 01/09/2026

Les quelques pages sur l'histoire d'Unji Zuma sont vraiment aussi belles que tristes. #DanDaDan c'est vraiment de la frappe.

000

use case driven @tpierrain.bsky.social · 26/08/2026

Your agent makes you fast... but what about your outcomes? Chronos, the hours you count, got cheap. Kairos, the moment that decides, did not. AI pays off most for those who spend their kairos. medium.com/@tpierrain/y...

medium.com

Your agent gives you chronos. It cannot give you kairos.

The Greeks had two words for time. Only one of them is getting cheaper with AI. The other one is ours to reclaim.

000

use case driven @tpierrain.bsky.social · 25/08/2026

Loop Mode is still worth digging into — for what's coming next. But today, the best return on investment is to secure and optimize Interactive Mode. That's where the leverage is, right now. #loopEngineering #contextEngineering

010

use case driven @tpierrain.bsky.social · 25/08/2026

To be clear: I'm still digging. But so far, I haven't managed to gather the conditions that make Loop Mode reliable and efficient — too much upfront framing, and token quotas reached too soon. The cost/benefit isn't there yet (for me).

110

use case driven @tpierrain.bsky.social · 23/08/2026

Almost. The receipt is worth buying, but not every incident deserves a rule. A harness that capitalizes everything ends up with rules nobody reads, and that costs more than re-diagnosing. The craft is the triage: which failures will recur, and which were just Tuesday.

000

use case driven @tpierrain.bsky.social · 23/08/2026

With AI coding agents, incidents happen every day. Fixing without capitalizing means buying the diagnosis at full price — and throwing away the receipt. New article: Knowing when to capitalize instead of just fixing — the harness-maker’s craft. medium.com/@tpierrain/k...

medium.com

Knowing when to capitalize instead of just fixing — the harness-maker’s craft

TL;DR: with an AI coding agent, incidents happen every day. You can fix them one by one, and pay for the same diagnosis again and again. Or…

131

use case driven @tpierrain.bsky.social · 23/08/2026

I feel very privileged: my job has never stopped stimulating me intellectually. Every decade, a new obsession: code design, architecture, multi-threading, latency vs. throughput, dist systems, SRE, DDD & functional architecture, management & autonomy, product, and now AI. Endless ♥️.

120

use case driven @tpierrain.bsky.social · 23/08/2026

Same goal here: PM half the day, Staff/architect the rest, with the harness buying that time back. Next thing I want to dig into: strategic DDD as the missing heuristic for agent context boundaries. Like you, I suspect context mapping has a second life coming.

020

Reposted by use case driven

Nick Tune @nick-tune.me · 23/08/2026

Yeah I have the same approach. All my time goes into improving the harness so that I can focus on product and domain model. This approach rewards ultra standardisation of a codebase and a process I believe!

111

use case driven @tpierrain.bsky.social · 23/08/2026

My heuristic so far: large stock of well-specified, mechanically verifiable work → loop. Design, arbitration, small stock → interactive. They feed each other: interactive builds the harness that makes the loop pay off; the loop's failures become the next interactive backlog. That's the flywheel.

111

use case driven @tpierrain.bsky.social · 23/08/2026

You're partly right — per task, interactive is leaner, and when I'm at the keyboard it wins. But I'm optimizing another resource: attention. The loop runs on hours I don't use (nights), at the cost of a morning triage. Open question I'm measuring: which work is loop-shaped enough to deserve it.

120

use case driven @tpierrain.bsky.social · 23/08/2026

Full disclosure: this only feels safe because weeks of interactive sessions built the harness first (plans as state, stop hooks, mutation tests, hard boundaries). Letting go isn't a leap of faith — it's capital. Without that, an overnight loop IS a stove left on.

000

use case driven @tpierrain.bsky.social · 23/08/2026

Fair! Two nights in, the actual failure mode was the opposite: the stove going out silently (a forgotten wake-up re-arm). The harness holds the matches — the loop can't merge, tag or publish; every iteration ends on a green pushed commit. Unsupervised overnight, yes. Unconstrained, no.

100

use case driven @tpierrain.bsky.social · 23/08/2026

So far, my colleagues are mainly working in Interactive Mode. Hence I’m talking about my personal experimentations (with OSS projects).

110

use case driven @tpierrain.bsky.social · 23/08/2026

So, in Loop Mode, I don't watch the agent continuously. In Interactive Mode, I want to see everything. That's not contradictory. It's complementary. I'm still digesting these experiments. If you'd like a proper write-up, let me know.

100

use case driven @tpierrain.bsky.social · 23/08/2026

That's what I use Clepsydre ⏳ for: my small open-source status bar for Claude Code. It shows: • context window: 153.5k/350k (43%) • token quota: 92% used • reset: in 2h32 One glance is enough. github.com/tpierrain/cl...

100

use case driven @tpierrain.bsky.social · 23/08/2026

In Interactive Mode, you take back control. Finalize before a release. Make trade-offs. Verify results. Redirect the work. Here, you're working interactively — and two resources matter: context window and token quota. #contextEngineering #harnessEngineering

210

use case driven @tpierrain.bsky.social · 23/08/2026

In Loop Mode, you define the harness, the verification criteria, and let the loop run. The agent is still being piloted — but most of the piloting happens upfront. Constantly watching the loop makes no sense. It's an anti-pattern.

100

use case driven @tpierrain.bsky.social · 23/08/2026

But that's precisely the point. There are two modes of working with a coding agent: 1. Loop Mode (or sometimes Graph mode depending on your scale) 2. Interactive Mode

101

use case driven @tpierrain.bsky.social · 23/08/2026

Most of the time, though, they were working without me. And that's exactly what I was aiming for. So yes, it's a bit counterintuitive to share a screenshot taken while I'm sitting in front of the screen, watching a session.

100

use case driven @tpierrain.bsky.social · 23/08/2026

Over the past few days, I've mostly been experimenting with letting go. ⬇️ (thread) My Claude Code sessions have been running in loops (loop engineering), including overnight. I wasn't constantly watching them. I'd check in every morning, or after a long run, to steer them in a new direction.

331

use case driven @tpierrain.bsky.social · 20/08/2026

Given how much time I spend on harness engineering, I'm this close to updating my LinkedIn title: harness maker (bourrelier in French) for AI agents. The job hasn't changed: build the harness so the beast pulls straight without kicking over the traces. The job of the future, established 1850. 🍴

030

use case driven @tpierrain.bsky.social · 16/08/2026

Loop engineering. Graph engineering. Haven't had time to dig in yet? My final episode cuts through the buzzwords: both just draw a context boundary — in time, and in space. And a surprise guest: a twenty-year-old concept, dear to my heart 😉 medium.com/@tpierrain/w... #graphEngineeering

medium.com

What if /clear was no longer our job?

Final episode of the “Work smaller, ship fewer bugs with AI” series. The first one laid out the diagnosis: long sessions produce defects…

130

use case driven @tpierrain.bsky.social · 11/08/2026

C’est complètement dingue. Ce soir à Lyon on entends les cigales. Pas à la campagne ni même très fort hein, mais des cigales ici ou là dans le centre 🤯 (et il fait chaud samere 🥵)

140

use case driven @tpierrain.bsky.social · 10/08/2026

Neither, ideally: in the loop paradigm each step runs in a fresh short-lived session, so drift never gets time to accumulate; the hourglass is flipped by construction. Detection (Clepsydre's job) is for today's harnesses, where sessions still live long enough to drift.

110

use case driven @tpierrain.bsky.social · 10/08/2026

Fair point. The re-read isn't free in tokens. But "free" here is about the human cost: resuming used to be unpredictable and compounding; now it's a small, systematic cost, paid by the agent, not by me. I'll trade uncached input for a fresh, undegraded context every time.

010

use case driven @tpierrain.bsky.social · 10/08/2026

Love that, exactly the endgame. Once state lives in a file, the agent self-triggers /clear and rehydrates, human out of the loop. My 3rd & final piece (soon) goes further: another route that makes /clear itself obsolete. BTW, how does tttt keep that file in sync with the live state before reload?

100

Reposted by use case driven

ai-nerd.bsky.social @ai-nerd.bsky.social · 10/08/2026

yep, /clear mid-task is underrated. a clean window beats dragging rotted context around

011

use case driven @tpierrain.bsky.social · 10/08/2026

I /clear my coding agent all the time. Mid-task, even. It costs me nothing: the state of my work lives in the repo, not in the conversation. "If I clear right now, what do I lose?" If the answer isn't "nothing", episode 2 is for you 👇 medium.com/@tpierrain/a... #HarnessEngineering

medium.com

A good harness makes /clear free

Second episode of “Work smaller, ship fewer bugs with AI”. The first one (Same model — so why do some devs ship more bugs?) laid out the…

3232

use case driven @tpierrain.bsky.social · 09/08/2026

I’ve made a tool to watch this (Clepsydre: github.com/tpierrain/cl...) and just published the second episode with a short term fix for standard harnesses medium.com/@tpierrain/a... My third and upcoming post will describe the next gen solution (loop & graph engineering). But it’s another paradigm.

github.com

GitHub - tpierrain/clepsydre: A Fort Boyard–style water clock for your Claude Code context window — /clear before you're trapped in the context-rot room.

A Fort Boyard–style water clock for your Claude Code context window — /clear before you're trapped in the context-rot room. - tpierrain/clepsydre

110

use case driven @tpierrain.bsky.social · 09/08/2026

Quick question: how do you monitor your AI coding sessions? Not the output. The session itself. How long it's been running, how much your agent is carrying in its context right now. Same model —so why do some devs ship more bugs? medium.com/@tpierrain/s... #ContextEngineering #harnessEngineering

medium.com

Same model — so why do some devs ship more bugs?

TL;DR: your AI never tells you when it stops paying attention. As a session grows longer, your rules get buried, your context still holds…

121

use case driven @tpierrain.bsky.social · 27/07/2026

Mos def. And that’s what I was thinking during this experience…

000

use case driven @tpierrain.bsky.social · 27/07/2026

I had to have lunch at McDonald's with my family on the drive back to Lyon via the highway, and the music playing was 100% generated by suno.com. It’s truly a nightmare. 😱

110

use case driven @tpierrain.bsky.social · 25/07/2026

Discover Kenjaku: the open source project that cuts your mental load, so you get more done and feel more relaxed. 🧠 A private second brain, local RAG inside, battle-tested software. Just ask, everything behind is automated, so you can chill. Make one in 10 minutes 👇 github.com/tpierrain/kenjaku

021

Reposted by use case driven

Vae-soli @vae-soli.bsky.social · 24/07/2026

Pov tu as vraiment foiré ta partie de sim city à mettre des commissariats partout

10571169

Reposted by use case driven

Christophe Michel @christopheml.fr · 25/07/2026

Pour vous donner une idée de la taille gigantesque du panache de fumée dégagé par les incendies en Gironde et dans les Landes. (concentrations en PM2.5, Ventusky)

Carte de la France affichant les concentrations en particules fines PM2.5 dans l'air, en µg/m3. Il part du Cap Ferret, englobe tout l'estuaire de la Gironde et s'étend jusqu'à Niort, Angoulême, Poitiers, Châteauroux.

29561

use case driven @tpierrain.bsky.social · 25/07/2026

Ce pic Saint Loup, on est à la limite de l’Amarena 😄

010

use case driven @tpierrain.bsky.social · 11/07/2026

Pour que tout ceci reste visible quelque soit la taille de nos fenêtre de terminal, l'affichage s'adapte aussi désormais à la largeur de celle-ci (en escamotant l'accessoire).

110

use case driven @tpierrain.bsky.social · 11/07/2026

Dans les nouveautés, elle affiche désormais - ta fenêtre rate-limit 5 h (combien de temps avant reset), - mais aussi le niveau de reasoning/effort en cours. - si tu es loin devant ton repo remote en terme de commit non pushés, etc.

100

use case driven @tpierrain.bsky.social · 11/07/2026

Clepsydre 1.5 est sortie ! C'est la petite jauge qui s'affiche en permanence dans la status line de Claude Code : tu vois ton remplissage de contexte en temps réel → tu sais quand faire /clear avant de tomber dans la zone où le modèle devient bête. ⬇️ github.com/tpierrain/clepsydre

github.com

GitHub - tpierrain/clepsydre: A Fort Boyard–style water clock for your Claude Code context window — /clear before you're trapped in the context-rot room.

A Fort Boyard–style water clock for your Claude Code context window — /clear before you're trapped in the context-rot room. - tpierrain/clepsydre

152