Remember to have fun

/goal make me a million dollars in bitcoin. Make no mistakes.

When I'm optimizing some part of a codebase, I end up in the same loop. Try a thing. Benchmark. Keep what helps. Discard what doesn't. Repeat until the result is good enough, or until I get bored. It works but it also eats half a day or more and ties up my attention while it runs. AI doesn't get bored. v0.11.0 ships the loop. Give Shaka an objective and it spawns an isolated git worktree, hands your terminal to your provider's TUI so a setup agent prepares everything with you watching, then runs the loop in the worktree until you stop it. Each accepted iteration commits, so the log is the record. Inspired by Karpathy's autoresearch and davebcn87's pi extension. `shaka optimize start "<your objective>"` v0.11.0 🤖 github.com/jgmontoya/shaka

Btw if you're using Shaka with Codex or Opencode, you might want to change the summarization_model to something cheaper/faster (on claude code we default to haiku) View quoted note →

Recently I've seen a lot of people saying that Claude got dumber. Maybe it did. I don't know and it doesn't matter, because it reminded me why I don't want my setup locked to one provider. Shaka has supported Claude Code and Opencode from the start. Skills, learnings, agents, commands, workflows, all portable between them. Define them once and use them interchangeably. v0.10.0 adds even more optionality: Codex. Same skills, same learnings, same everything. Pick the best tool for the task and your setup comes with you. Three new agents ship here too: CodeReviewer, DevOpsEngineer and TechnicalWriter. The slop scanner picked up some additional patterns with tighter false-positive rules to make your writing even less robotic and soulless. v0.10.0 🤖 github.com/jgmontoya/shaka

Learnings tell Shaka what to do, but domain knowledge still evaporates between sessions. v0.9.0 adds a knowledge base. The system extracts domain knowledge from sessions and compiles it into topic pages, scoped per project. Architecture decisions, root causes, design tradeoffs. Compiled once, maintained automatically, loaded at session start. You already have the data. `shaka memory compile --bootstrap` on your projects builds your knowledge base from existing sessions. v0.9.0 🤖 github.com/jgmontoya/shaka

Marmotize all the groups! View quoted note →

Shaka has had learnings since v0.3. The system extracts corrections, patterns, and preferences from your sessions and loads them into context next time. The missing piece: what happens when you have hundreds of them. They compete for space. Related entries stay fragmented. Quality degrades under its own weight. v0.8.0 adds automatic memory maintenance. Related learnings get condensed into compound entries. Cross-project patterns get promoted to global. Low-value entries get pruned when space is tight. Runs in the background after sessions, conservative by default, original learnings are archived before every change. The system cleans up after itself now. v0.8.0 🤖 github.com/jgmontoya/shaka

Last time I posted about Shaka it had slash commands. That was v0.4.0. Since then I've been building the parts that make each session better than the last. It remembers across sessions now. Rolling summaries compress what you worked on into daily, weekly, and monthly context. You shouldn't have to explain what you did yesterday. It runs multi-step workflows. Chain prompts, commands, and scripts in YAML. Steps pass output to each other. Add loop: 3 and it iterates: review, fix, review again until clean. It catches AI slop. A WritingRules skill flags the "it's important to note" garbage, and Shaka scan scores any file so you can check before you post. It has a skill manager now. Install from GitHub or the Clawhub registry. Skills get scanned for sketchy stuff before install. And it ships a TDD skill that forces the AI into red-green-refactor instead of "write everything then pray." v0.7.1. 🤖 github.com/jgmontoya/shaka

Also make it privacy focused for increased hardness! View quoted note →