It's the end of the world as we know it
(and I feel fine)
How I'm surviving & thriving in the tokenpocalypse.
Olavs Rāciņš · AI Founders Hub Riga · 29.06.26
Same usage. 23× the bill.
The subsidised phase of AI is ending. Usage costs real money, and the bill is yours.
So stop playing around with AI. Start using it in a targeted, efficient way, with clear ROI.
caveman: brain still big, mouth small
A Claude Code skill (also Codex, Gemini, Cursor, 30+ agents) that cuts 65 to 75% of output tokens at the same technical accuracy.
github.com/juliusbrussee/caveman · Side note: a Mar 2026 paper found constraining models to brief answers improved accuracy 26 points on some benchmarks. Verbose isn't always better.
Before · 69 tokens
I'll go ahead and carefully read through the configuration file to fully understand the current setup before making any changes.
After · 19 tokens
read config. understand setup. then change.
ponytail: write less code
"The best code is the code you never wrote."
Folders instead of frameworks
Interpretable Context Methodology (ICM) / Model Workspace Protocol: numbered folders are pipeline stages, plain markdown files carry the prompts & context. No framework code.
Replaces multi-agent orchestration (CrewAI, LangChain, AutoGen) with filesystem structure, for sequential, human-reviewed workflows.
Jake Van Clief & David McDermott, Eduba / University of Edinburgh · arxiv.org/abs/2603.16021
I'm already running this
Odysseus
Self-hosted AI workspace: chat, agents, research, docs, email, notes, calendar. Docker Compose, single-binary install, full ownership of your data.
OpenCode
Open-source coding agent. Any model from any provider, no vendor lock-in. 160K GitHub stars, 7.5M monthly devs. Terminal, IDE & desktop app.
Headroom: compression as a layer
A context-compression layer: library, proxy, MCP server, 6 algorithms. 60 to 95% fewer tokens on real workloads, accuracy preserved on standard benchmarks (GSM8K, TruthfulQA, SQuAD v2, BFCL).
Also shapes output (verbosity steering, effort routing), so it complements caveman rather than competing with it.
Drop-in, no code changes · github.com/headroomlabs-ai/headroom
SRE incident debug
65,694
5,118
tokens · 92% saved
If you remember four things
01 · Compress Context
Why use much context when less do trick
02 · Folders > frameworks
Filesystem structure can replace orchestration code for sequential, human-reviewed workflows.
03 · Avoid lock-in
Self-hosted + provider-agnostic keep optionality as pricing shifts.
04 · Don't use AI
You don't have a token optimisation problem, if you think
Thank you. Do you have any questions?
Build with me, or connect
Olavs Rāciņš
delulu.solutions