How a Claude project assembles what it sends, with instructions prepended and paid for every turn, chats that share no memory, documented file limits worth re-checking, and the switch from direct context to retrieval over chunks; plus the note schema, filename rules, and four portable questions, put to ChatGPT Projects, Gemini Gems and NotebookLM from their September 2026 help pages. Episode page & show notes Visit website Instructions, isolated chats and filing A working project workspace hides a question worth answering before anything larger is built on it: when you send a request inside it, what actually reaches the model? The standing instructions are not memory. Claude's project instructions are prepended to every conversation in the project — re-sent rather than remembered — so a fault argued out of one chat returns in the next, and the block is paid for on every turn. Worked with invented figures: a two-thousand-token block against a window a hundred times larger is one percent per turn, ten exchanges is ten payments of it, and it competes for room with the only material that can make the answer true. Keep it to the role, the output structure and the evidence rules, and leave one-off tasks in the conversation. Conversations inside a project share no memory with each other, so nothing a chat produces saves itself: the loop is produce, check, save locally, upload through the knowledge panel. Files run up to thirty megabytes and accept plain text, Markdown, PDF, Word and comma-separated values; a PDF of up to a hundred pages is read as text and visuals, and longer ones as text only. All of that is as Anthropic documented it in the autumn of two thousand twenty-six, and all of it is worth looking up again before a filing habit is built on it. What turns a good session into a reusable asset is structure: a fixed set of note headings written into the standing instructions, with anything a source does not mention recorded as not reported. Identical fields can be compared field by field, and a blank becomes information rather than an invisible gap. A second rule keeps invented teaching material out of real findings — prefixes on synthetic filenames, an exclusion rule in the instructions, a marker line inside each document. And some of this work needs no model at all: whether every note has all its fields, how many files exist, which were added this week. Where the input fixes the answer, ordinary code produces it; pay for a model where the answer requires reading meaning. What the model actually sees Project knowledge is loaded directly into the context window, which Anthropic describes as a limit without documenting a number: the size of the desk, holding the instructions, the conversation, the pulled-in files and the answer being written. While two documents fit, both are simply there, whole, every time. Once the base outgrows that window, paid plans switch to retrieval: an internal knowledge-search tool fetches the chunks that look relevant instead of loading everything. Capacity rises, by up to ten times on Anthropic's figure; certainty falls. Because the unit retrieval finds is a passage rather than a document, a claim that only holds when two paragraphs eight pages apart are read together can be answered from one of them — which is the argument for short single-subject documents whose subject appears in the heading and again in the filename. In the project stops meaning in front of the model, and the audit gains a step: ask which files and which sentences were used, then click the document in the knowledge panel and read its raw text. Last, the distinction that decides what needs verifying: the thirty-megabyte cap is a limit the application enforces, while answer only from these files is an instruction the model follows most of the time, since Anthropic's own guidance says it can produce convincing but ungrounded responses. An uploaded document is data to read, never authority to redirect the job. And because every one of these mechanics is a product decision rather than a law, they travel as questions rather than facts: what text gets re-sent each turn, what crosses between two conversations, whether the material is in front of the model or being selected on your behalf, and which rules the software enforces rather than merely forwards. Those answers live in the documentation, not in the interface: a panel labelled memory may be a re-sent field, and a tidy sidebar of files may be a search index. The same questions put to three other workspaces Read from the vendors' help pages in September of two thousand twenty-six, the four questions applied to ChatGPT Projects, Gemini Gems and NotebookLM return uneven answers, and the gaps are the lesson. ChatGPT Projects has a per-project instructions field that overrides account-wide custom instructions, offers a choice between default and project-only memory so chats can reference each other inside a project, and does not document whether files are loaded whole or searched; limits are five hundred and twelve megabytes and two million tokens per file. Gemini Gems is a name plus instructions, takes up to ten knowledge files of up to a hundred megabytes and uses the latest version of each, documents a million-token context window, and documents nothing about memory between chats or whether files are searched. NotebookLM has no free-form instructions field, only a response style and length; it keeps chat history, states that responses use only your sources, and allows five hundred thousand words or two hundred megabytes per source with fifty sources per free notebook. No vendor documents whether instructions are re-sent each turn. The brief, headings and filing loop carry over everywhere; what changes is which of the four answers you have to find out for yourself.
