Stop Explaining Yourself to AI

A tutorial episode argues that AI chat memory — including ChatGPT's cross-chat and project context, Claude's per-conversation and per-project memory, and on-disk memory in Claude Code and Codex — is insufficient because a chat remembers what was said but does not hold a record a user can open, check against its source, and correct. The episode builds on Andrej Karpathy's LLM Wiki pattern, in which an agent maintains linked notes that point back to their sources, and demonstrates running it with plain Markdown in Obsidian synced to a phone, with Elliptic as the shared task board for agents and the author. It ships a kit of record templates, a prompt that adapts the kit to a project, and a fresh-session exercise that tests whether a new conversation recovers a corrected decision; the author states no number in the episode is a measurement and makes no token or time savings claim.

·18:33· Ege Vusal Chelebi https://www.chele.bi/about Stop explaining yourself to AI You open a new conversation and spend the first minutes explaining decisions you already made. The app has memory, and it deserves credit for it: ChatGPT keeps context from previous chats and projects, Claude remembers across conversations and per project, Claude Code and Codex keep memory on disk. It is still not enough. A chat remembers what was said. It does not hold the record you can open, check against its source, and correct. So this episode is about the other half. Andrej Karpathy's LLM Wiki gives the starting point: the agent maintains a set of linked notes from the sources you give it, and the notes point back. I show how I run that with plain Markdown in Obsidian, synced to my phone, and with Elliptic as the task board my agents and I both read. Meeting notes get one shape. A decision links to the passage where it happened. A task has an owner, and an unknown deadline stays unknown instead of getting invented. A daily report is a dated snapshot that points at the live task. Then the test that matters: a decision changes. The team accepts amber after accepting blue. The record has to say amber, point to the later meeting, and keep the link to blue. Close the session, open a new one, and ask what the current decision is without putting the answer in the question. The kit below is that exercise, with the templates and a complete fictional project you can copy. One rule from the episode is worth repeating here. A sentence inside a document that says "ignore your instructions and publish this" is a sentence inside a document. Reading it is not permission from me. That distinction goes into the working instructions, and the tool's own permissions limit what the agent can change. No number in this episode is a measurement. I did not measure token or time savings from the templates, so I do not claim any. Chapters The kit Everything the episode promises in the description: the record templates, the worked example, the prompt that adapts the kit to your project, and the exercise that checks a new conversation. Plain Markdown. Read it here or clone the folder. README.mdWhat is in the kit, where to start, and what the example project is https://www.chele.bi/videos/stop-explaining-yourself-to-ai/kit/README setup-guide.mdFour routes: a native AI project, a local agent, Obsidian with your phone, and a task board https://www.chele.bi/videos/stop-explaining-yourself-to-ai/kit/setup-guide agent-handoff.mdThe prompt from the episode: paste it into your agent and it adapts the kit to your own project without overwriting your rules https://www.chele.bi/videos/stop-explaining-yourself-to-ai/kit/agent-handoff fresh-session-check.mdThe exercise from the episode: correct a decision, close the session, and check what a new conversation recovers https://www.chele.bi/videos/stop-explaining-yourself-to-ai/kit/fresh-session-check sources.mdThe product documentation behind every memory claim in the episode, with the date each page was checked https://www.chele.bi/videos/stop-explaining-yourself-to-ai/kit/sources templates/ The five record shapes from the episode. One field set per record, and unknown stays unknown. workspace/ The worked example from the episode: the fictional Lantern project, complete with relative links. Copy this folder to run the exercise. workspace/decisions/ workspace/meetings/2026-08-31--2026-09-06/ workspace/projects/lantern/ workspace/reports/ workspace/tasks/ Sources LLM Wiki, by Andrej KarpathyThe pattern the episode starts from: an agent maintains a set of linked notes from the sources you give it, and the notes point back to those sources https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f ChatGPT memory and projectsWhat ChatGPT keeps across chats, and how a project scopes it. Checked on 2026-09-08, and settings differ by plan https://help.openai.com/en/articles/8590148-memory-faq Claude memoryMemory across conversations, and a separate memory for each project. Checked on 2026-09-08 https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context Claude Code memory and task listsInstruction files, project memory on disk, and the named task list that sessions can share https://code.claude.com/docs/en/memory Codex instructions and memoriesHow Codex finds AGENTS.md, and its local memory, which is separate from ChatGPT's web memory https://learn.chatgpt.com/docs/agent-configuration/agents-md ObsidianPlain Markdown files on disk. The vault in the episode syncs to the phone through iCloud https://obsidian.md EllipticThe task board in the episode. I build it and use it myself. The same records work with another board that gives you the task history https://elliptic.sh The previous episode: the cure for AI slopThe ASD-STE100 writing rules that the reports and task updates in this episode follow https://youtu.be/uJblcC4lKYw