Ornith-1.0: the open-source coding LLM that matches Claude Opus 4.7 — and runs on your laptop. This open-weights agentic-coding model hits 77.5 on Terminal-Bench and 82.4 on SWE-bench Verified, beating Claude Opus on both, then ships a 9B size you can `ollama run` locally. MIT licensed, four sizes, and it taught itself the harness.

Opened up a new area for me: The dreaded social posts and keeping track of what is done and what is a good idea to do in the future. My copy paste from the log Me: Are you following our ICM approach for the skill for the social media folders too? Claude Cowork: Thought process Thought process Executing now — including the receptionist file so it finally gets tracked. Loaded tools Loaded tools Good catch — let me not commit until I've answered this properly, because it changes what I build. Short version: no, and here's the important distinction. ICM Three-Lane (Lane → Domain → Skill) governs skills, not content artifacts. What I've been doing is creating content (content/social-media/...), which is an output location, not a skill — so it isn't required to sit in the lane tree. But if you want a repeatable, agent-routable social media skill, that absolutely must follow ICM. Let me verify the exact conventions rather than guess.

Focus is becoming one of the rarest assets in business. Everyone is distracted. Notifications. Meetings. Emails. Messages. Constant context switching. The organizations that learn how to protect attention will outperform those that don't. Focus = momentum.

Fixing a project done by DeepSeek this week in May. It seems like I just wasn't getting anywhere. Claude was asking if it should check in stuff. Didn't create PR's, no tests being built.. Silly questions all around, and worst of all, it started breaking things that were working, like it was not listening to me. Then i figured it out, found where the process was dropped and fixed what was wrong.. 😁Claude's final comment is what brings a smile to my face. 😁 Claude: "That closes out the whole task — MeetingRecord is now a fully valid ICM project (routing, 5 stack-appropriate dev skills, 4 personas, closeout contract, knowledge layer, all merged to main), and its merge workflow matches the rest of the portfolio." Thank you @Jake Van Clief 😇

Most people drive AI like a slot machine. Type a vague wish, pull the lever, hope. Wrong output, pull again. Three hours later: forty messages, half right code, no idea which version was good. That is not building. That is gambling with extra steps. The fix is not a cleverer prompt. It is a plan. Here is the exact habit, start to finish. The one idea Stop prompting. Start defining outcomes. A prompt is a wish. An outcome is a result you can check. AI accelerates, it does not generate. Point it at a clear target and it closes the gap fast. Point it at a vague one and it just gets lost faster. ———————————————————————— The chain (this is the whole method) 1. Brainstorm first. Do not open a chat and start barking instructions. Open a conversation whose only job is to decide what you are building. What does done look like. What must be true. What is out of scope. I use `/brainstorm` and answer one question at a time until the fog clears. 2. Freeze it in a spec. A spec is the brainstorm written down and made hard to argue with. Three parts: - Objective, in one sentence you could prove true or false. - Constraints, the rules that bound the solution. - Definition of done, where every item can be checked with a single action. If you cannot write the objective sentence, you are not ready to build. Better to learn that now, for free. 3. Break it into a phased plan. Turn the spec into ordered, bite sized phases. The rule that matters most: each phase must be small enough to finish in one sitting, with room to spare. 4. Hand each phase over as a typed brief. This is the part that does the heavy lifting. Read the next section. ———————————————————————— Think in PRDs A typed brief is a product requirements document, scaled to whatever you are making. It is the difference between "build me a settings page" and a short doc with a goal, constraints, acceptance criteria, and phases. It feels like bureaucracy. It is the opposite. The brief is a forcing function: - You cannot write acceptance criteria for a feature you do not understand.