Coffee Talks · TechPulse Trenčín
What works for me
A map of all the decisions you can make to better use LLMs.
August 6 2026
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20261 / 18
A running example
Split a text into sentences, with stable ids, across 45 languages
What do we need?
One id per sentence, identical wherever the text appears, whatever the language.
How can we do it cheaply?
The cheapest model that reliably splits sentences in every language.
How can we trust the result?
Deterministic ids and checks, so a bad split is caught, not shipped.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20262 / 18
Hello
Robin Cussol
Staff Engineer at Kiwi.com. (Used to be?) Frontend platform lead, and lately shipping a backend service with an LLM at its core.
robincussol.com · @RobinCsl
Regular attendee of the TechPulse meetups. Talked here two years ago: 3 Principles for Effective (Platform) Teams.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20263 / 18
Why you should listen to me
What I've been up to
GitHub Copilot → Cursor → Claude Code → OpenCode CLI → Pi
Played with Ralph Loops before they were cool
Frontend platform lead → backend service with an LLM at its core
Vibe coded a PWA to replace my meditation app
Vibe coded a PWA to drill Slovak into my brain
Used a /teach skill to learn some FreeCAD and Slovak
I've been testing LLMs regularly to extend my Maths PhD thesis result, it's been fun
Timer + session tracking · Termius → Raspberry Pi → Amp, built on my phone.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20264 / 18
The pitch
A lot of decisions in front of you
When you want to accomplish something requiring intelligence, there are a lot of decisions in front of you, unless you have unlimited tokens, never get rate-limited, and don't care about privacy.
Using Auto mode in Cursor is fine. It just doesn't teach you anything.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20265 / 18
The map
Category of decisions
01
Models
which model, provider, settings
02
Harness
agentic vs workflow, caching, sandboxing
03
Context window
what you let the model see
04
Review & building
trusting output, shipping it
05
Cost & experiments
cheap tokens, fun projects
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20266 / 18
01 · Models
The model, the provider, the settings
Model
SOTA vs open source · dense vs MoE · input/output price · modalities.
Provider
Inputs used for training or not · reliable or not · long-term support · run it locally?
Settings
Temperature · thinking vs no-thinking.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20267 / 18
01 · Models
Good enough often wins
A less intelligent model might require more turns to accomplish a task, ending up more expensive than a pricier per-token model.
Router models: some folks route each request to the most apt, cost-effective model for the task.
Most underrated
DeepSeek v4 Flash
Most fun I've tried
Gemini 3.1 Flash Preview LiveGemini TTS is crazy good
just today: probable price increase
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20268 / 18
02 · Harness
The harness you drive it with
Agentic vs Workflow
Loops of tool calls, or a fixed pipeline. Pick per task.
Coding Agent
The default driver for software work. Keep it as simple as possible. I recommend pi, it's simple and works with all kinds of models.
Stable prompts
Variable inputs at the end → stable prefix → input caching. A cache miss on a long conversation, or a misconfigured pi, bills full price every turn.
Security
Sandbox the agent, or don't.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/20269 / 18
03 · Context window
Manage what the model sees
Right context
Pi /tree goes back to a previous state of the coding session, from the LLM payload point of view. Went on a side quest? Backtrack without polluting the context.
The dumb zone
The noisy middle that dilutes attention. Cut it before it cuts quality.
Stable prefix
Keep the prompt head constant across turns: implicit or explicit input caching.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202610 / 18
04 · Code review
Trust the review, keep your judgment
The review agent says it's all good, so just ship?
Not blindly. Exercise your judgment, and stay knowledgeable enough about your codebase and domain.
remove all the slop in the last commit
in a fresh session, clean context, sharpest review
Actively observe what the agent is doing and course-correct if need be. If that happens all the time, consider adding it to your instructions.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202611 / 18
04 · Building
From idea to shipped
PRD → Tasks
Write the plan as a PRD, then break it into tasks the agent can execute.
Matt Pocock's skills
Reusable, composable skills: teach the agent once, use it everywhere.
Keep the loop tight
Small batches, quick reviews. Momentum beats one big perfect run.
PRD → tasks → review → ship
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202612 / 18
05 · Cost
Cheap or free tokens
OpenCode Go
$10/month for $60 of usage — scan to get started.
opencode.ai/go
referral — we each get $5 ;-)
OpenRouter
Free models, one API.
Google AI Studio
Free tier, perfect for experiments.
Run locally
The cheapest tokens are the ones you don't use.
Open models run on consumer hardware: Qwen, Gemma (Gaming GPU), DeepSeek v4 Flash on a 128 GB MacBook Pro.
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202613 / 18
05 · Experiments
Small models, big ideas
On your phone
LFM or Gemma models can run on phones, small enough to live on-device, no cloud, no token bill.
Learn to run models locally with Unsloth.ai, or on GPU with llama.cpp.
Small models = very fast and good enough for summarizing and classifying tasks
Example: A HackerNews watcher, summarizes new posts, filter to only the ones you care about
Why bother
The decisions above don't just save money, they make the cheap experiments possible.
vs
I use Fable 5 for everything $5 to read your current directory
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202614 / 18
Back to the example
The sentence job, done cheaply
First attempt: a script with thinking enabled, burning tokens for no reason
Gemini 3.1 Flash Lite: more than enough intelligence, great at languages
Not sure DeepSeek v4 Flash would have done as well on niche European languages
Just a prompt and an API call: enough, no coding agent needed, just a script
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202615 / 18
Further reading
References
earendil.com/postsfrom the folks behind pi
ghuntley.comRalph Loop creator, lots of useful ideas
github.com/mattpocock/skillsreusable skills for AI-assisted dev
youtube.com/@aiDotEngineerAI Engineer Conference channel
youtu.be/xUnRQ9vLXxoTheo Browne: "Everything we knew about software has changed"
AI Engineer Conference on YouTube
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202616 / 18
Parting words
Be more ambitious
Theo Browne asks it of us
Revisit the practices, tools and approaches you take for granted, especially as intelligence gets cheaper, even while token prices climb.
The only constant is change
None of this existed a year ago. It's currently the worst it's going to be.
Subsidized tokens are ending
Don't lock yourself into one provider. Bring your harness of choice. Open models run on consumer hardware. earendil.com/posts/session-portability
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202617 / 18
KISS KISS GOODBYE
Keep things simple
and learn by doing
robincussol.com · @RobinCsl
Robin Cussol · Coffee Talks by TechPulse Trenčín · 06/08/202618 / 18