Agentic engineering, coding taste, and research insights from building Command Code.
Jev is a general-purpose decision model. taste-1 is a meta neuro-symbolic decision model for one domain, coding taste, running at 100T-token scale this year. Neither is an LLM. Architecture, pipeline, and the differences, in diagrams.
Full-stack engineer Sivaram Pandariganthan burned through 527M tokens on DeepSeek V4 Flash for 20 cents using Command Code's tool call repairs, cache-friendly sessions, and vision tool.
Evan uses Command Code for agents, evals, and data work. The taste feature learned to distinguish between his engineering process and production data lake workstreams, even where they overlap.
YouTuber AICodeKing tested Command Code's harness against open models like GLM 5.2 and concluded the gap between open and closed models is smaller than the gap between good and bad harnesses.
Set one objective and walk away. /goal lets Command Code work toward a goal across many turns, and only calls it done after an independent check verifies it from real evidence.
JeffersonNunn.eth ran 6.9B tokens across 83,022 agent runs on Command Code for about $156 in three weeks, keeping 16 windows going every day, and still had spend left.
vel (@Velcrafting) ran 4.8B tokens in May across a $281 AI stack. Command Code was the $1 line item, tagged "Insane ROI" next to $200 of OpenAI and $40 of grok.
Yzn (@yazanbruv) ran 500M+ tokens on DeepSeek V4 through Command Code's $1 plan and shipped a fully functional Android app, now in review, for less than a dollar.
Austin Ginder, a heavy $200/mo Claude Code user, ran 21.5M tokens through DeepSeek Pro on Command Code for 75 cents — and started moving his Opus 4.7 work to it.
Use your Command Code plan with OpenAI and Anthropic compatible clients using your Command Code API key.
The skill that turns Command Code into your design agent. Eighteen modes for color, type, layout, motion, states, and accessibility. Built against glassmorphism by default
Rules files and skills decay over time. The real solution is coding taste, patterns learned from how you actually write code.
Ship faster with OpenAI GPT models in Command Code, with a purpose-built coding harness, taste conditioned on your coding patterns, and cross-model PR review.
Ran the same 3 prompts through Command Code and Claude Code. Claude wrote correct code, but Command Code wrote *my* code every single time. The difference is worth seeing.
Run npx taste pull and Command Code starts building UI in the style of Apple, Vercel, Shopify, or seven others. Learn how design taste files work.
Command Code learns from your feedback to generate code that matches your coding taste. How we built a system that adapts to how you actually write.
We shipped hooks today, a programmable API to the Command Code lifecycle. Run your own code, deterministically, at the boundary where the agent crosses from deciding to acting.
Why static configuration decays, why tutorials only go so far, and why continuous learning from developer behavior is the missing layer.
Shift from scattered PR feedback to confident shipping decisions with agentic review that understands your codebase, your workflow, and how you think about code.
A deep dive into the full stack of terminals — what makes a terminal, how shells work, what TTY and PTY mean, POSIX standards, ANSI escape codes, and more.
Taste files solve a personal problem. But engineering is a team sport. Here's how we built the infrastructure to make accumulated engineering judgement portable across teams and open source projects.
When long coding sessions hit their limits, most agents break by losing context. Here’s why compaction fails and how a better approach fixes it.
AI coding agents write correct code, not your code. Teams that build and share coding taste (learned preferences from accept/reject/edit signals) compound a 5-10x productivity advantage. Here's the information-theoretic argument for why taste is the next frontier of AI-assisted development.
Learn how our PMM team built a creator-discovery pipeline as a Command Code skill that finds dev creators better than the SaaS tools like Modash and Upfluence. Runs in 10 minutes, and costs about 3 cents per run.