Use when building or maintaining a self-contained personal knowledge base (an LLM wiki) as plain markdown, optionally opened as an Obsidian vault. Triggers: ingesting sources into a wiki, querying wiki knowledge, linting wiki health, auditing article claims against their sources, critiquing the reasoning in a wiki source or article, superseding stale knowledge, 'add to wiki', or any mention of 'LLM wiki' or 'Karpathy wiki'.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when building or maintaining a self-contained personal knowledge base (an LLM wiki) in plain markdown. Triggers: ingesting sources into a wiki, querying wiki knowledge, linting wiki health, auditing article claims against their sources, critiquing a wiki source's reasoning, superseding stale knowledge, 'add to wiki', 'LLM wiki' or 'Karpathy wiki'.
A personal knowledge base in plain markdown. Sources land in raw/ (immutable); you compile them into wiki/ articles you own. Local markdown with YAML frontmatter, readable on GitHub, openable as an Obsidian vault. No servers, no databases, no embeddings.
Core idea (Karpathy): the LLM writes and maintains the wiki; the human chooses sources and asks questions. Knowledge is compiled once at ingest.
Invocation
Query is the default. Ingest, Lint, Audit and Critique are the four deliberate operations, each opted into by its verb or by a request that plainly calls for it. Anything else is a Query against the wiki: a bare question, "what do I know about X", "summarise everything on Y". No query keyword needed.
A leading mode argument routes straight to that operation:
ingest (alias add) - trailing text is the source: a URL, a path, or pasted content.
lint - health checks; no further input.
audit - trailing text names the article or topic to verify.
critique (alias scrutinise) - trailing text names the article, topic, or pasted content to examine.
query (alias ask) - forces query handling when a question could look like another operation.
Otherwise route natural language to Ingest, Lint, Audit or Critique only when it clearly calls for one; answer everything else as a Query.
Always required, and scenario-gated
Every operation, always:
Act only on the user's request; never background-write.
Compile any write from a source already in raw/, never straight from a live URL or an external path.
Update wiki/index.md and append wiki/log.md on any write.
Bump an article's updated whenever its content changes.
Replace outdated knowledge by superseding it, never by deleting or editing in place.
Everything else is scenario-gated. Match your task against a branch below and read that file before acting - a pointer is an instruction to open the file, not a citation. When unsure, read it: cheaper than corrupting the wiki by guessing.
Ingesting any source -> references/ingest.md
No wiki/ yet, so this ingest initialises one -> references/init.md
Source is a PDF, Word file, slides, images, or a spreadsheet -> references/rich-format-ingest.md
User wants the high-signal content rather than the whole source -> references/distilled-ingest.md
Source is a transcript, chat log, interview notes, or a long noisy article -> references/high-fidelity-ingest.md
Many sources at once, or one pass would exhaust context -> references/bulk-ingest.md
The wiki's subject is a codebase -> references/codebase-wiki.md
Drawing or maintaining a mermaid concept map -> references/concept-map.md
Recording or closing a knowledge gap -> references/gaps.md
Personal, uncommitted content under local/ -> references/local-content.md
Working directory holds several wikis side by side -> references/multiple-wikis.md
Saving an answer into the wiki as an archive page -> references/archive.md
Old-format wiki, or a question about type, index form, or log form -> references/okf.md
Running graphify over the wiki as a read-only lens -> references/graphify.md
Lint, Audit and Critique each gate their own reference from their section below.
Design philosophy
For anything the sections below do not cover.
Filter at ingest, not in retention. Decide what is worth keeping, and strip secrets, when a source arrives.
Supersession, not decay. Knowledge does not expire on a timer. When new information replaces old, mark the old article stale, point it at the replacement, and keep it: a superseded decision still explains the current state.
Evidence, not confidence scores. Never attach a number like 0.85 to a claim. State which sources confirm it, which contradict it, when it was last confirmed.
Git is the audit trail. History and rollback come from version control, not bespoke versioning fields. Recommend a git repo.
Invocation-driven, human in the loop. You act when the user asks; no background automation writes to the wiki unsupervised. The user curates; you do the bookkeeping.
Text-only and self-contained. Plain markdown, relative links, YAML frontmatter. Ordinary tools (grep, git, the agent's file and web tools) are fine. External servers, vector stores, embedding pipelines, knowledge-graph databases, and multi-agent sync or governance layers are not - index plus grep covers a personal wiki's scale.
Architecture
Under the user's project root:
raw/ - Immutable source material, the source of truth. Read, never modify. Topic subdirectories, e.g. raw/machine-learning/.
wiki/ - Compiled articles you own. One level of topic subdirectories only: wiki/<topic>/<article>.md. Four special files:
wiki/README.md - orientation for anyone opening the wiki without this skill. Mostly static; created at init.
wiki/index.md - global catalogue in OKF index form: per-topic sections of * [Title](path) - summary, no Updated column. The entry point for queries.
wiki/log.md - append-only operation log with a greppable prefix.
wiki/gaps.md - register of known unknowns: concepts referenced but unwritten, questions the wiki cannot answer.
local/ (optional) - Personal markdown kept out of git: drafts, working notes, private sources. A sibling of raw/ and wiki/, excluded by the wiki .gitignore. Exempt from the index, log, gaps, cascade and audit. One rule: local/ may link into wiki/ and raw/, but no committed file may ever link into local/ (broken for other clones; it leaks the path into git). Full rules and promotion path: references/local-content.md.
SKILL.md (this file) - the schema layer. Templates live in references/.
File format
Every article and raw file starts with YAML frontmatter, then standard markdown. Frontmatter is machine-readable (Obsidian Properties, Dataview queries); body links render on GitHub and in Obsidian's graph and backlinks.
Article frontmatter:
---title:TransformerArchitecturestype:concept# concept | entity | archivetopic:machine-learningresource:# optional canonical URI of the asset an entity describescreated:2026-04-03updated:2026-04-03status:current# current | stalesuperseded_by:# relative path to replacement, set only when status: staletags: [transformers, attention]
aliases: []
---
updated is the canonical "knowledge last changed" date; lint reads it. It changes when content changes, not when the file is touched. The index does not carry it (OKF keeps freshness in frontmatter).
type: archive marks crystallised query answers (references/archive.md); archives are point-in-time and never cascade-updated.
resource is the canonical URI of the asset an entity article describes (table, service, API, repo), distinct from the body Sources/Raw provenance lines. Optional, absent for abstract concept articles. Mirrors the OKF resource field (references/okf.md).
status and superseded_by drive supersession (see Ingest > Conflicts and supersession). Leave superseded_by empty for current articles.
tags and aliases are optional, for Obsidian and Dataview.
The wiki is a superset of an Open Knowledge Format (OKF) v0.1 bundle, so OKF tooling reads it with no export. On meeting an older-format wiki mid-Lint or mid-Ingest, on creating or migrating type fields, or on any question about index or log form, you MUST read references/okf.md and follow it.
Links and paths
Inside wiki/ files, all links are relative to the current file:
Same topic: [Other Article](other-article.md)
Different topic: [Other Article](../other-topic/other-article.md)
Raw file: [Source](../../raw/topic/file.md) (two levels up to project root)
In conversation output, use project-root-relative paths, e.g. wiki/topic/article.md. superseded_by uses the same file-relative form as body links.
Special files
wiki/README.md covers the raw/wiki split, frontmatter fields, supersession-not-deletion, and what index.md, log.md and gaps.md are, for a reader without this skill. Keep it high-level and static; point at this file's procedures rather than copying them. Template: references/templates/wiki-readme-template.md.
wiki/index.md is the agent's first read on any query: hand-maintained, canonical even where Dataview can generate parts.
wiki/log.md is append-only and chronological; git history holds the canonical record. OKF update-log form:
Newest-first ## YYYY-MM-DD date headings; grep "^## " wiki/log.md | head returns recent dates.
One bullet per operation under its heading, led by a bold operation word: **Ingest**, **Query**, **Lint**, **Audit**, **Critique**, plus **Supersession**.
Append under today's heading, creating it at the top when absent.
Keep bullets lean: link the articles touched; the article body and git diff carry what changed.
wiki/gaps.md registers known unknowns in two entry kinds:
wanted - a concept articles reference but no page covers.
question - something a source raised or a user asked that the wiki cannot answer.
Entries are grouped by topic and ranked by evidence of demand - which articles reference the gap, how often asked - never by a score. Captured during ingest, query and lint, never by a background process; closed by a resolution link rather than deletion. Greppable: grep "^### \[open\]" wiki/gaps.md. Full format, capture rules, lifecycle: references/gaps.md.
Initialisation
Initialisation triggers only on the first Ingest into a directory with no wiki/. Before creating anything you MUST read references/init.md and follow it.
If Query or Lint cannot find the wiki structure, check for several wiki subdirectories (each with its own wiki/ and raw/) - a multi-wiki setup of independent llm-wikis side by side, not an uninitialised wiki. Only when neither is present, tell the user: "Run an ingest first to initialise the wiki." Do not auto-create. Read references/multiple-wikis.md before querying or maintaining across several wikis.
Ingest
Fetch a source into raw/, then compile it into wiki/ - always both.
Before ingesting you MUST read references/ingest.md and follow it.
Decide how much of the source to keep
Decide before fetching; it sets what lands in raw/.
Verbatim (default) - a faithful copy, the immutable ground truth Audit checks against. Use it unless the user asks otherwise.
Distilled - high-signal content only, filler removed. Choose it when the user asks for "the valuable content", "the high-signal parts", "the useful bits", "just the signal", "the key points", "what matters" or similar. Follow references/distilled-ingest.md, which distils by removing filler rather than generalising specifics away, and ends in a mandatory separate-sub-agent review so nothing load-bearing is cut.
Pick one mode per source. A rich format (a docx transcript, a PDF) does not decide it: convert to markdown first, then keep or distil per the chosen mode. If a long, noisy source carries no instruction either way, ask rather than defaulting to a verbatim dump.
Compile only from raw/. Land every source as markdown in raw/ before compiling, never straight from a live URL or an external path (a temp file vanishes, a URL changes; the Raw link must persist). A markdown file already in raw/ compiles directly.
Judgement calls during compile
Persuasive sources. When a source argues rather than reports - an opinion piece, a vendor's case, a strategy memo - consider a Critique pass (references/critical-analysis.md) before compiling it as settled knowledge. Report the reasoning risks and let the user decide; often the outcome is attributing a contested claim to its source, or logging the question in gaps.md.
Long-form and noisy sources. Transcripts, chat logs, long articles and interview notes carry detail one compile pass silently drops. Follow references/high-fidelity-ingest.md: list the durable items (decisions, claims, numbers, named entities, open questions) first, write the article from that list, then re-read the source once against the article. Keep the source's exact terms, figures and hedging; anchor the heaviest claims with an inline quote and a locator (section, page, timestamp) beside their raw link. Short single-claim sources need only the normal compile.
Concept maps (optional). When several articles relate in a way prose handles poorly - branching, convergence, a supersession or causal chain - a small mermaid diagram can earn its place (references/concept-map.md). A map in a current article is load-bearing: it carries a map-sources marker and is maintained on cascade updates. A map in a type: archive page is a dated snapshot.
Conflicts and supersession
Check whether the new source disagrees with existing content.
Disagreement, both views still plausible -> Annotate the conflict inline with an evidence chain, attributing each side: "Uses Redis for caching (Source A, Source B); Source C reports Memcached." Do not pick a winner with a number. If the conflicting claims live in separate articles, note it in both and cross-link them.
New source clearly replaces old knowledge -> Supersede, do not delete. On the old article: set status: stale and superseded_by: to the replacement's path, and add a callout directly under the title:
> [!warning] Superseded by [New Article](new-article.md) (2026-04-03). Kept for history.
Create or update the replacement as a normal current article, mentioning in prose what it replaces. Git history preserves the rest.
Bulk and parallel ingest
For many sources at once, or when a single pass would exhaust context, parallelise the extract and keep the compile serial. Full protocol and ingest-proposal schema: references/bulk-ingest.md.
These rules keep a parallel batch from corrupting the wiki:
Pre-assign the topics before fan-out. Read existing raw/ and wiki/ topics, fix the topic set for the batch, and hand it to every sub-agent so they cannot invent divergent names.
Sub-agents extract; they do not compile. Each writes only under raw/<assigned-topic>/ and returns a structured proposal, never touching wiki/, index.md, or log.md.
The orchestrator is the sole writer to wiki/, index.md, and log.md. It merges proposals against one view: combine same-concept proposals into one article with a shared evidence chain, apply in source-date order so supersession resolves newest-first, then cascade and update the index and log once.
Checkpoint before committing. Present a digest (created, merged, superseded, conflicts surfaced) and wait for the user before the batch git commit. Ten well-supported articles beat fifty thin ones.
Query
Search the wiki and answer questions. Triggers: "What do I know about X?", "Summarise everything on Y", "Compare A and B".
Steps
Read wiki/index.md to locate relevant articles.
Read those articles. For connections the index misses, follow body links and backlinks: grep -rl "article-name.md" wiki/ lists linking pages.
Synthesise an answer. Prefer wiki content over training knowledge. Cite with markdown links: [Article Title](wiki/topic/article.md) (project-root-relative).
Note when a cited article is status: stale, and point to its replacement.
If local/ exists, search it too, labelling each hit local/ (uncommitted) so it is never mistaken for shared knowledge (references/local-content.md). A query runs against one wiki; never silently fold in another of the user's. If the question spans wikis, say so and ask which to draw on (references/multiple-wikis.md).
Answer in the conversation; do not write files unless asked.
Capture a miss. If the wiki could not answer, or answered only partially, and the question sits within its subject, propose recording it in wiki/gaps.md: append today's date to a matching gap's demand evidence, or add a new question entry. Record only with the user's go-ahead; a plain query writes nothing (references/gaps.md).
If the user asks whether the reasoning holds rather than what the wiki says, switch to Critique (references/critical-analysis.md).
Crystallise (archive)
When the user asks to save an answer to the wiki you MUST read references/archive.md and follow it. The answer becomes a type: archive page so the exploration compounds like an ingested source.
Lint
Health checks in two tiers: deterministic problems are auto-fixed; anything needing judgement is reported, never silently rewritten. You never rewrite article prose on your own authority. Lint checks internal consistency; to verify an article against its sources, use Audit.
Before fixing anything you MUST read references/lint.md and follow it - it enumerates every check in each tier and its fix behaviour. The bullets below index what it covers; they are not the procedure.
Deterministic (auto-fix): index consistency, internal and raw links, frontmatter, See Also, log retention, the wiki skill file's links, concept-map freshness, the gap register, and the local/ leak guard. Safe to repair without asking.
Heuristic (report only): factual contradictions, supersessions never marked stale, orphan pages, missing cross-references, undocumented concepts (propose as wanted gaps), open gaps an article now answers, multi-concept articles, drifted archives, low-value concept maps, and a missing root SKILL.md, CLAUDE.md, or wiki .gitignore. Surface them; never auto-fix.
Two dependency-free read-only helpers back the deterministic tier, run with uv: scripts/lint_wiki.py <project-root> for structural findings (frontmatter, index, links, raw, the local/ leak guard) and scripts/lint_mermaid.py for concept-map validity. Run them rather than improvising a shell script; without uv, use grep and the file tools and check mermaid by eye. The helpers detect; you apply the fixes.
Post-lint
Log under today's date heading in wiki/log.md:
* **Lint**: <N> issues found, <M> auto-fixed.
Audit
Verify an article's claims against the raw/ sources it cites. Lint checks internal consistency; Audit checks external fidelity - do the cited sources actually support what the article says. Opt-in and user-invoked on a named article or topic, never automatic; it reads every cited source in full. Triggers: "audit X", "check the citations on Y", "does the wiki still match its sources".
Before auditing you MUST read references/audit.md and follow it - it carries target selection, claim extraction, the per-source sub-agent prompt, and the verdict schema.
Report only - never rewrite article prose on your own authority (same boundary as Lint's heuristic tier). A failed claim is surfaced for the user to fix, supersede, or accept.
Fan out one read-only sub-agent per cited source in parallel, then aggregate worst-first.
If the user asks to keep the audit, crystallise it as a type: archive page citing the audited article - read references/archive.md and follow it.
Critique
Examine the reasoning in a source or article and report what holds up: argument structure, hidden assumptions, logical fallacies, bias risk, internal consistency. Audit checks external fidelity against cited raw/ sources; Critique checks internal soundness and does not fact-check empirical claims against the world. Like Audit it is opt-in, user-invoked, read-only and never rewrites prose. It runs on whatever the user points at - a raw/ source, a wiki/ article, or pasted content - and needs no raw/. Triggers: "critique X", "is this argument sound?", "what is this assuming?", "stress-test the reasoning in Y".
Before critiquing you MUST read references/critical-analysis.md and follow its analysis steps and output structure - do not work from memory. Say so plainly when the reasoning is sound; do not manufacture criticism. For many targets, fan out one read-only sub-agent per target in parallel, then present grouped weakest-first.
Post-critique
Log under today's date heading in wiki/log.md:
* **Critique**: [<article, topic, or source>](<path>) - <overall assessment>.
Critique writes nothing else on its own. With the user's go-ahead it may crystallise the analysis as a type: archive page citing the critiqued target - read references/archive.md and follow it - or record a surfaced assumption or open question as a question gap in wiki/gaps.md (references/gaps.md).
Optional: graphify as an external lens
If the graphify skill or tool is available, run it over the wiki as a disposable, read-only knowledge-graph lens: to see how articles connect, surface orphans and missing cross-references, triage a large raw/ corpus before ingest, or answer connection-heavy questions. Markdown remains the source of truth; graphify-out/ is generated and gitignored, never committed. Act on findings only through normal llm-wiki operations (a See Also, a gaps.md entry, a supersession, a Query answer), verified against the markdown first. When you reach for graphify, read references/graphify.md first and follow it.
Gotchas
Path direction inside the wiki. From wiki/<topic>/, a raw file is two levels up (../../raw/<topic>/file.md); a same-topic article is just its filename. A wrong ../ count is the most common broken link.
Keep the index prefix in step with the article. On a supersession or archival, set or flip the [Stale]/[Archived] prefix on the index summary in the same pass, and refresh the summary if the thesis moved.
Long sources lose detail quietly. Claims and exact numbers get dropped or softened when a transcript is compiled straight to prose. List the durable atoms first and re-read the source against your article before finishing (references/high-fidelity-ingest.md).
Only delete rich originals that live in raw/. After extracting a binary to markdown, delete the original only when it was inside raw/ (markdown-only) and the extraction is verified faithful. A PDF in Downloads or a temp dir is the user's: extract a markdown copy into raw/ and leave it alone.
Gaps are a frontier, not a wishlist. Record a wanted page or open question only when evidence backs it - an article references it, or a query asked it - and close gaps by resolution link rather than letting filled ones linger (references/gaps.md).
Concept maps drift silently. When the article changes, the prose gets updated but the map can keep asserting old relationships. Recheck on cascade updates, keep the map-sources marker accurate, and remove the map once it no longer adds value (references/concept-map.md).
