Understanding Jev, the AI Built for Decisions (Sprint Day 2): Two Second-Brain Systems - a Leadership Self-Review Built Problem-First, and a Tag-Classifying Knowledge Base Where a Cheap Decision Model Replaces a Chat Model

Dilip began a second brain without a purpose and abandoned it. The second attempt started from a pain: as one of the older leaders in a young startup, nobody gave him coaching or feedback, and colleagues are all firefighting. He searched for 'a system that can coach me' and it turned out to be a form of second brain. His lesson: people take a solution and attach terms to it ('now I have a second brain, what do I use it for?'); instead say what you want to achieve, build the most brute-force version and refine. Sometimes a second brain is not needed at all. He mocks the tool-choice trap (Claude or Codex? Box or Dropbox? Obsidian or Hermes?) - decision paralysis costs more than a wrong pick, so pick one and begin; his own mistake was to start with Hermes and Telegram (cumbersome) before moving to a plain Claude project where his muscle memory already is. Guardrails he used: a 90-minute timer for research, no token optimisation at the start ('the feedback is more valuable than the tokens'), connect a system only when you feel the pain it solves, and treat effectiveness as self-assessed. He labels these Day 1 and Day 2 talks 'BTS' sessions - behind the scenes of how he thought, not a finished recipe - and says they only fit sprints, not the main programme.

Version 1: a Claude project. Quarterly goals sit in the project instructions; six principles of leadership (reverse-engineered from executive-coaching writing, inspired by Ethan Evans) sit there too; each week he dictates his activities from the phone and once a week asks 'based on what I did, coach me: which activities moved me toward the goals, which away, which had nothing to do with them'. Fix 1: goals change each quarter, so they move out of the instructions into a markdown file the project is told to read; principles stay in the instructions. Fix 2: mobile chats and laptop memory did not sync, so every dictation is saved as a weekly markdown file (week -> month -> quarter syntheses). Google Drive was tried and dropped because the Drive connector hit API limits and read big files only partially, so the files moved to Box, an enterprise store described as AI-native (5 GB free was ample) with a Claude connector; a self-review folder syncs between the Box cloud and his desktop and a new folder is created each week. He added the connector only after the simple version worked. Fix 3: while travelling he stopped dictating for 30 days; since he lives in Slack, email and Granola, a Slack and email connector now pulls his activities, commitments and loop-closing every three days, with voice notes for nuance (Whisper Flow on the phone). Fix 4: the six principles felt too simplistic, so he adopted the book Multipliers (the company gifted it to staff), converted the Kindle book to chapter-by-chapter markdown with ChatGPT/Claude and added it as the project's knowledge. Every conversation is stored as markdown, an optional index.md tracks patterns, and the system warns when he repeats a mistake. He stresses AI is not plug-and-play: it took iteration, and he asked the AI to be brutal, not kind.

A second brain is a personal knowledge management system: store, retrieve, query and organise. Pre-AI, people dropped links and notes by date and manually linked related ones into a web (a knowledge graph) where related topics sit close together and serendipitous links connect distant ones. Evernote, OneNote and Roam Research served that, Obsidian made it local, and Notion began as knowledge management and became a work operating system. Tiago Forte coined 'second brain'; Nate Jones and Cole Medin are named as good follow-ups; Andrej Karpathy's thread on LLM-built wikis made it viral. With AI the human dumps and the AI observes, connects and arranges. Participants named Obsidian, Notion, Otter, Mem and NotebookLM-style retrieval systems they use. He built a reference site (shared in chat) of 18 approaches to build a second brain and 33 real builds, plus where the Jev API fits; he does not teach it, it is for reading when ready.

His reasoning: matching an article to one of nine tags is classification, so the output is a tag name. A frontier chat model (he reads OpenRouter prices as spoken: Opus about 4 USD per million input and 10 per million output, Sonnet about 2, Haiku about 0.1) would work but wastes money on generation. Jev is presented as a reinforcement-trained 'decision' model from a company he calls TypeShift AI (founder 'Diego Almeda', described as an ex-OpenAI GPT-4 builder; names as heard) built to make fast, structured decisions software can use directly: you send content plus a list of choices and get back a typed answer with probabilities. Pricing as spoken: input about 0.042 USD per million tokens, output tokens free because the answer is a decision, which is an order of magnitude below chat models; a million tokens is roughly the size of the Harry Potter books. Limits he stresses: Jev cannot create files, rewrite a CV or draft a lesson - it can score whether a CV matches a job description. He confirmed with Codex that the documentation supports choosing from a supplied list and returning probabilities, and ran two live test calls through OpenRouter (an MCP article mapped to the MCP and connectors tag; an unrelated gardening sample mapped to Misc) for a total cost of about 0.0000117 USD. A local LLM is free but weaker. He says the sprint billing of Jev as the topic was deliberately kept organic: introduce it only where it earns a place.

Problem statement: dump different sources of information and have the system tie each to the sessions he teaches so he can upgrade a session with the new tactic. Steps in the room: a new ChatGPT/Codex project on a fresh folder; dictation with Whisper Flow (switching to Willow Voice when the daily limit ran out); ask Codex to check his Granola meeting notes for his AI-accelerator sessions and propose tags - it returned nine (context engineering, personal AI setup, MCP and connectors, skills, workflow automation, agent systems, local and open models, problem framing and one more) which he fixed as static tags, anything else going to Misc. Then the design questions to the class: how to match a tag to an article (slugs? an LLM?), which led to Jev. Final workflow: paste an article URL; FireCrawl extracts the markdown and keeps the source link (with ScrapeCreators as fallback if FireCrawl fails); two independent steps - Jev selects one of the nine tags, Claude Haiku via OpenRouter writes a separate summary; save an article file, a summary file, and update the tag's markdown page with links to both; no lesson drafting, no agents. Folder layout: articles, summaries, tags, and a second-brain app folder (a small Python back end with a terminal-style HTML front end designed from a Matrix terminal prompt copied from superdesign.dev), kept separate so it can go to GitHub. He told Codex to choose the stack and build the MVP; an insufficient-credit error on a provider halted the first run and he said he would fix it and show it next sprint. Chat adds free alternatives (a free FireCrawl-like scraper and a free Jev-like option) and he plans a Chrome extension later.

As announced by Dilip and repeated by Kartik: until now every Outskill programme (accelerator, fellowship, Catalyst) was for generalists, with engineering fellowship and engineering accelerator as the technical tracks. Community feedback asked for depth in a function, so two paid programmes are being tested through a priority waitlist form: AI for Marketing Leaders (campaign design, marketing planning, creative strategy, running a team) and AI for Founders and Business (business systems and running teams; complements Catalyst, which is for people without a business yet). About six weeks each, alpha cohorts that will take a fraction of applicants as co-builders, pricing and curriculum to follow; people who fill the form get a masterclass explaining each. The Reboot cohort is separate: seven days, open to everyone from the generalist, accelerator and Catalyst cohorts, tentatively starting around 7 December (subject to change), no waitlist, covering what changed in AI and which principles still hold. Housekeeping: certificates by link (a first link failed and was fixed) or by email, social-post links for tagging Outskill and the mentors, and the LMS route to every sprint and master-class recording (AI master classes and sprints space, Resources and Recording).