iiinova — where your digital self keeps evolving

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Kumbukum looks as a nice service. Similar what I do in Scribe. I'm still thinking making output accessible also online and not only locally but I'm afraid of security. Looks like you handled it already, or not? :D good luck with your work

010

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Full eval framework, autoresearch loop, and 22 Iron Laws are in github.com/oliver-kriska/claude-elixir-phoenix . If you maintain Claude Code skills, run a tournament on your descriptions before your next release. It costs less than a coffee and changes more than a feature. #ClaudeCode #ElixirLang

020

Oliver Kriska @quatermain.bsky.social · 21/05/2026

The descriptions the autoresearch loop landed on were ones I would never have written. Not because they were better English. Because they hit the user's vocabulary, which my brain — being the author — had stopped seeing.

110

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Skill descriptions are interface design, not feature design. Your users do not read the implementation. They type a few words. The model reads a few sentences. If those two layers do not share vocabulary, the feature does not exist.

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

20+ iterations. 0 discards (every winner was kept). 100% win rate at the end. /phx:plan trigger recall: 0% → 100%. Across the whole plugin: tournament mode lifted trigger accuracy 50.5% → 78%. Total cost: $0.40.

110

Oliver Kriska @quatermain.bsky.social · 21/05/2026

So I built a loop. Generate 8 candidate descriptions. Test each against 50 user phrasings. Pick the winner. Then iterate on the winner. Pure tournament. Each round costs about $0.05 of API (but I used subscription)

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

First attempt: rewrite the description by hand. I am the author. I should know. Recall went from 0% to 18%. Worse — every rewrite I tried felt right but tested badly. I was guessing.

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

The skill description used technical vocabulary — "Phoenix planning workflow", "PRD generator". Users phrase their work as "build", "add", "make". When the user query and the skill description live in different vocabularies, recall is zero. The skill might as well not exist.

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Before v2.6.0, the /phx:plan skill in my Claude Code plugin had a 0% trigger recall. Send "build a chat feature" to Haiku. It would never route to /phx:plan. The skill was good. The description was unfindable.

120

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Most "AI memory" tools forget the failure path. The 3-week dead end nobody resolved — six months later you re-derive it from scratch in a different repo. scribe captures the failure too. Your agents stop starting from zero. getscribe.dev · github.com/oliver-kriska/scribe

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

Two ways to run it. Same pipeline, one YAML line: · Claude API — best quality, $ per sync · 100% local Ollama — $0/sync, gemma3:12b, ~70s for the weekly Dream cycle on a 32GB Mac Per-op overrides too: keep cheap passes local, pin heavy writes to Claude. No lock-in.

110

Oliver Kriska @quatermain.bsky.social · 21/05/2026

How it gathers data: · every Claude Code + Codex session (ccrider FTS5) · git commits across your repos · URLs you text yourself (iMessage bookmarks) · drop files Claude leaves in any repo BM25 picks the dense ones, two-pass LLM writes the wiki, qmd indexes it with embeddings.

100

Oliver Kriska @quatermain.bsky.social · 21/05/2026

I never write notes. My tools do. scribe reads my git commits, my Claude Code + Codex sessions, the links I text myself — and writes a wiki my next agent session queries. 6,977 docs in 4.5 months. Zero typed by hand. getscribe.dev #ClaudeCode #AI

120

Oliver Kriska @quatermain.bsky.social · 20/05/2026

Plus the multi-model trigger eval foundation: MODEL=sonnet make eval-multimodel MODELS=haiku,sonnet make eval-compare-models Per-model cache + N-way ASCII comparator with ↕/⚠ disagreement markers. Groundwork for verifying v3.0.0 multi-agent ports on non-Claude judges.

110

Oliver Kriska @quatermain.bsky.social · 20/05/2026

v2.10.3 — CC hook-API catch-up (2.1.110 / .139 / .145): · Stop hook surfaces forgotten `mix phx.server` + pending `/schedule` jobs · block-dangerous-ops emits structured JSON deny + safer-alternative additionalContext that persists into the next turn · SessionStart welcome → args[] exec form

100

Oliver Kriska @quatermain.bsky.social · 20/05/2026

v2.10.2 — /phx:research, /phx:brainstorm, /phx:perf, /phx:pr-review now reliably invoke when typed as slash commands (#53, reported by @bigardone). Root cause: stale `disable-model-invocation: true` hit claude-code#26251. Same fix plan/review/investigate got months ago.

100

Oliver Kriska @quatermain.bsky.social · 20/05/2026

2.10.1 — Cross-project hook bleed (#55). Enable the plugin globally and before this: Phoenix Iron Laws fired on tokenizer.cpp (token), /admin_panel/ parent-dir matches, unrelated repos. After: 9 hooks self-gate on mix.exs. Filename match tightened to .ex/.exs/.heex/.eex.

100

Oliver Kriska @quatermain.bsky.social · 20/05/2026

v2.10.1 → v2.10.3 of the Elixir/Phoenix plugin for Claude Code. Three patch releases, one theme: be a better citizen — global-install safe, slash commands reliable, hooks aligned with latest CC. 28 files | +635/-90 github.com/oliver-kriska/claude-elixir-phoenix #ElixirLang #ClaudeCode

150

Oliver Kriska @quatermain.bsky.social · 19/05/2026

Exactly. Try it out and let me know feedback so we can make it even better than first version

110

Oliver Kriska @quatermain.bsky.social · 19/05/2026

Graceful degradation is the contract: a missing source becomes 1 honest line in Risks, never an error. git log is the always-available floor. /ketchup is the easter-egg alias. Why not. github.com/oliver-kriska/claude-eli…

000

Oliver Kriska @quatermain.bsky.social · 19/05/2026

The bit github user druyang asked for and I'm proudest of: "Impact on your work" intersects files OTHERS moved upstream with YOUR in-flight scope (open PRs, local branches, working tree). Direct vs adjacent overlap, classified. --depth deep reads diffs for per-file semantic impact.

100

Oliver Kriska @quatermain.bsky.social · 19/05/2026

/catchup pulls in your time window: · GitHub: review-requests + notifications · Git: commits + diffs on branches that moved · Linear MCP: issues + comments touched · Google Calendar MCP: meetings you missed Then collapses into one Context Brief — not five tabs.

100

Oliver Kriska @quatermain.bsky.social · 19/05/2026

v2.10.0 of the Elixir/Phoenix plugin for Claude Code ships a second plugin alongside it: catchup — framework-agnostic, return-from-absence briefing for async teams. One /catchup, one prioritized brief. No firehose. phxagents.dev/install #ClaudeCode #DevTools

210

Oliver Kriska @quatermain.bsky.social · 17/05/2026

But yes, without AI it would be impossible. I have there skill for checking cc changelog and verify state of plugin, skill for creating twitter/x posts, etc

010

Oliver Kriska @quatermain.bsky.social · 17/05/2026

Later I added tool for investigating of previous sessions where I used those skills, testing those skills and also later multiple layers of evaluation. Also based on my sessions with building those skills and agents I built own skills to simplify the process.

110

Oliver Kriska @quatermain.bsky.social · 17/05/2026

I’m trying to set there workflows to gather some data and store only rules and docs which are stable. So you should not find there some exact api calls, mostly references and so on. Also I have “framework “ around it. I started with a few skill, added more…

110

Oliver Kriska @quatermain.bsky.social · 17/05/2026

Open-source. Working contributor guide in the repo. Tell me what is missing — especially if you maintain anything LiveView, Ecto, or Oban heavy. The next 22 Iron Laws will come from real production pain, not from me alone.

000

Oliver Kriska @quatermain.bsky.social · 17/05/2026

8-dimension eval framework. 24 Python matchers. 52 pytest tests. ~5 second CI run. Autoresearch loop ran 20+ iterations on skill descriptions. 0 discards. Tournament mode raised trigger accuracy 50.5% → 78% at $0.40 of API cost. This is the part nobody sees. It is also the only reason this scales.

100

Oliver Kriska @quatermain.bsky.social · 17/05/2026

Skills are slash commands. Agents are specialists you spawn for parallel work. Iron Laws are hard rules that block bad output before it ships. Hooks tie it together — 9 events fired, 18 scripts wired in. PostToolUse was silently broken for a month before session analysis caught it.

110

Oliver Kriska @quatermain.bsky.social · 17/05/2026

laude-elixir-phoenix at 4 months in. 41 skills · 20 specialist agents · 22 Iron Laws · 100 reference docs · 18 hooks. ~28,000 lines. Everything I have learned shipping production Phoenix apps, encoded as machine-readable rules. github.com/oliver-kriska/claude-elixir-phoenix

272

Oliver Kriska @quatermain.bsky.social · 15/05/2026

$8,523 of Claude Code over ~4.5 months. 11.9B tokens. 11.4B of them — 96% — are cache reads. That ratio is the whole story. The bill stays low not because the work is small, but because the harness reuses context instead of paying for it twice.

100

Oliver Kriska @quatermain.bsky.social · 14/05/2026

This is /phx:review on a real PR. The unlock is not "AI reviews code." It is orchestration + model-tiering + result compression — the shape Erlang/OTP gave us decades ago, applied to agents.

110

Oliver Kriska @quatermain.bsky.social · 14/05/2026

By design, the 7 agents do not dump raw output into the main session. Full per-agent findings persist on disk. A context-supervisor (haiku) can dedupe and compress them before the parent synthesises. You get depth when you need it, safe context when you don't

100

Oliver Kriska @quatermain.bsky.social · 14/05/2026

Each agent runs on the model that fits its job. security-analyzer → opus elixir-reviewer, liveview-architect, testing-reviewer, iron-law-judge, requirements-verifier → sonnet verification-runner → haiku You do not pay opus rates for `mix *` commands

100

Oliver Kriska @quatermain.bsky.social · 14/05/2026

7 specialist agents reviewing one PR in parallel. 37 files, +296 / −123. A real rewrite, not a toy diff. ~315k tokens, ~2 minutes wall time. Sequential would be ~11 minutes

130

Oliver Kriska @quatermain.bsky.social · 14/05/2026

11,000 PRs. 5 active developers. 26 contributors total. 5 years. Not because we are unusually fast. Because the pipeline is set up so that "is this safe to merge" is a 2-minute question, not a 2-hour one. Pipeline beats individual velocity, every quarter, forever. #SoftwareEngineering

041

Oliver Kriska @quatermain.bsky.social · 12/05/2026

METR, 2025: experienced devs using AI were 19% slower. They believed they were 24% faster. The IDE got faster. The pipeline did not. The brain definitely did not. That gap is where most of the AI productivity debate is happening — and almost no one is naming it. #AIEngineering

100

Oliver Kriska @quatermain.bsky.social · 11/05/2026

What's the one Elixir + AI question you've been afraid to ask - context engineering, production agent harnesses, anything? The next post is probably your answer. #AIEngineering #ContextEngineering

000

Oliver Kriska @quatermain.bsky.social · 11/05/2026

Expect more posts. Not more volume — more density. Each one backed by a real session, a real number, or a real failure. No takes. No slop.

100

Oliver Kriska @quatermain.bsky.social · 11/05/2026

For the past 4 months I have been quietly building a pipeline: more than 4,000 markdown documents of research, session logs, Phoenix patterns, and what I have actually shipped — all indexed, searchable. That is the input. The output starts this week.

220

Oliver Kriska @quatermain.bsky.social · 06/01/2025

If I take simple ping as a marker, I have 6ms to app.hey .com server (104.18.5.243).

010

Oliver Kriska @quatermain.bsky.social · 06/01/2025

I think Hey .com has this stack as well and it's really slow. Not sure if it's because I'm in east Europe but I'm really disappointed about it. And it's not because of HW or connection, 350 mbps Wifi or 500 mbps 5G connection. I hope native.live will be faster

native.live

LiveView Native

LiveView Native eliminates the need to maintain web and native development teams, empowering developers to build web and native seamlessly with Phoenix.

110

Oliver Kriska @quatermain.bsky.social · 21/12/2024

Hmm. I don’t get it, from year perspective 12k vs 50k. Or do you mean that this small cost can be lost in paperwork so you just forget about it same as Netflix subscription?

100

Oliver Kriska @quatermain.bsky.social · 21/12/2024

You can take these topics and invite some other developers from known companies to introduce them and discuss that topic. I started similar way podcast in my language.

000

Oliver Kriska @quatermain.bsky.social · 11/12/2024

There are some updates: "This release address by introducing a compiler lock, ensuring that only a single operating system running Elixir compiles your project at a given moment" github.com/elixir-lang/...

github.com

Release v1.18.0-rc.0 · elixir-lang/elixir

Elixir v1.18 is an impressive release with improvements across the two main efforts happening within the Elixir ecosystem right now: set-theoretic types and language servers. It also comes with bui...

010

Oliver Kriska @quatermain.bsky.social · 10/12/2024

oooh, that would be nice, any info about it?

100