# How my LLM wiki works Adam Seed — adamjseed.co.uk — September 2026 Answers to the LLM wiki exchange prompt, describing structure and conventions only. The pattern is Andrej Karpathy's LLM wiki: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f ## 1. Shape Markdown in a private git repository, edited by an agent and readable in Obsidian. Three top-level folders: immutable raw sources, agent-maintained pages, and generated outputs. Page types are people, organisations, sources, projects and meals. Every page carries frontmatter with created and updated dates and a confidence rating. Git is doing more work here than it looks. Every change is a commit with the reasoning in the message, so a rewrite is never actually destructive — the previous version is always recoverable. That single property is what makes aggressive tidying safe rather than frightening. ## 2. Rules that are actually written down - Raw sources are never edited. If the original is not plain text, keep the original file AND a literal text mirror. No reformatting, no tidying. - Source pages must not be lossy. Anything short enough is reproduced in full, because queries read the wiki layer and never the raw layer. - Derived documents regenerate from the raw original, never from a summary. - Write each fact once, on the page that owns it. The log records what happened and links out. If a fact would need editing in two places when it changes, it is in the wrong shape. - If nothing owns a fact, create the page — people especially. Added after the rule above quietly lost things with no home. - Keep genuinely log-shaped detail in the log: what was tried and rejected, what failed and in what order. That has no current-state page to live on. - The log is append-only; completed months are archived, not edited in place. ## 3. What it does, not just holds I use mine for a mix of things — some running regularly, others still being refined. - Inbox triage — works through what an automated classifier has flagged, groups it, drafts replies, clears the rest. - Renewals and life admin — finds policies, subscriptions and contracts buried in email, files them with their dates, and surfaces what is expiring. - Meal planning — a week of evening meals from a library of what the household actually eats, accounting for who is in and what they will not eat. - Networking and opportunities — keeps a tailored CV and cover letter ready from a standing version, tracks people worth staying in touch with, and builds prep when a conversation turns into something worth taking further. - Remembering people — birthdays, anniversaries, gift ideas, and a nudge when I have not spoken to someone in months, built from what is already on their page. - A daily briefing — what is on today, what needs a reply, what is expiring, published to a private page I can open on my phone rather than something I have to go and ask for. - A researcher — watching a set of blogs, feeds and releases, and folding what actually matters into that briefing rather than creating another inbox to read. - Calendar — reading what is coming up, and putting things into it. - School and household admin — term dates, permissions and the steady drip of paperwork that comes with children. - Holidays — a trip as one thread: flights, accommodation, who is going, what is still to book. - Home projects — quotes, contractors, decisions and status for anything being done to the house. - Money — capital, rates and scenarios against a target, with the decisions and the reasoning kept alongside them. - This website — the posts and resources on it are drafted from what the wiki already holds. - Getting documents in — a watched folder for things arriving, not just publishing things out. - Capturing what I actually did this week — it knows my CV, not my week, and closing that gap would make everything else better. - Writing in my voice — the style rules applied to anything drafted on my behalf. The division that works: a cheap deterministic classifier handles volume, and the LLM only sees the ambiguous remainder. It never replaces that classifier, it works on top of its output. Standing rule: surface, never act. Anything needing judgement gets flagged rather than handled — an unsigned contract, a failed payment, an expiring policy. Nothing runs unattended. ## 4. What I changed my mind about ### Folders are a filing system; this is a linking system Sources started in a folder per project, mirroring how I had organised things before. Wrong twice, a layer apart — first for raw sources, then again for project pages, where one folder held exactly one file. The general point: a folder forces every item to have exactly one parent, so you have to choose one. A source relating to three projects does not have one parent, and a link costs nothing. Anything that can be expressed as a link should not be expressed as a folder. ### Summarising the source material The first source page was a tidy summary of a CV. That is wrong in a specific, non-obvious way: queries read the wiki layer and never touch the raw files, so a summarised source page means every future answer is built on a lossy copy, while the exact original sits one folder away, unread. Reversed within a day. ### Deleting old log entries to save space The log had grown to 128KB in nineteen days, and the plan was to rewrite the old entries short. Before deleting anything I checked what existed only in the log — and found real things, including two former colleagues with no pages anywhere. Deleting would have destroyed the only record of them. So the order changed: migrate what is at risk onto real pages first, then rewrite. The check is the transferable part, not the rewrite. ## 5. What broke - Everything was written twice — the log and the pages narrating the same events at the same depth. 128KB in nineteen days. - Eight people and organisations existed only in log entries, including someone from a significant professional conversation the day before. - A heading that lied: "open questions" holding twelve resolved items, one of them labelled "live". - Parallel sessions duplicated work — two sessions independently produced the same tailored document because a branch was never merged. - Confidently wrong once: a tailored CV silently widened a claim from one role to a whole career. Caught on review. The fix became a hard rule — tailor by reordering and selecting only, never by rewriting sentences. ## 6. Keeping it honest The part almost nobody builds, and the part that decides whether this is still useful in a year. - Automated integrity checks: every page indexed, no orphans, no broken links, no missing frontmatter, every raw source referenced by something. - Pins: a register of corrections made by hand, so a later ingest cannot silently reassert the old version. It protects the edits the agent has no other way of knowing about. - Confidence ratings and explicit unknowns. Guessing is the failure mode; "not confirmed" is a valid recorded state. - Ask before losing detail. If a tidy-up might drop something and it is unclear whether it exists elsewhere, stop and ask. - Unsolved: scale. The index is read in full on every query. Fine at eighty-five pages. Not fine at eight hundred, and I do not yet have an answer for it. ## 7. What I would do differently Decide where each fact lives before writing anything into it. Nearly every problem above traces to the same root: no rule about page versus log, so everything went in both, and anything without an obvious home went in neither. And on day one: put it in git. Not for collaboration — for the ability to rewrite aggressively later without risking the loss of something only recorded in one place.