wwxd

wwxd

Build a source-linked knowledge base from expert talks, podcasts and essays, then use it as context in conversations with a language model. wwxd discovers and fetches sources, an agent organises them into a Markdown wiki, and the CLI checks quoted passages against the originals.

Compare how experts approach a problem, explore where they disagree and trace an answer back to what they actually said. "What would X do?" is one way to use the knowledge base.

Answers are inferred from public sources by a language model. They are not statements by these people, and none of them is affiliated with or endorses wwxd. See ETHICS.md.

Try it

You need uv, Claude Code, ffmpeg, and deno or node 22 or newer.

uv tool install "git+https://github.com/MSliwa01/wwxd" && wwxd install-skill && wwxd doctor

Compiling costs about $1 to $2 per video on Opus at list price, so run wwxd estimate <slug> first. The Quick start has the next steps.

Why

Ask a chatbot whether to raise your prices and you get a generic answer that commits to nothing. That's fine for facts. It's weak for decisions in business, design, careers or learning, where the useful advice comes from people with strong opinions and a track record. Ask the model to answer "as Alex Hormozi" and you get a plausible impression, built from whatever it remembers, with the gaps filled by its own guesses. It can't know anything the person said after its training cutoff, and it mixes up who said what, crediting a podcast host's line to the guest.

wwxd gives the agent the person's own words instead. Here is a question from our benchmark, answered by Claude Opus 5.5 in Claude Code.

I have $50k saved. Should I pay off my car loan and credit cards or put it all into my business? What would Alex Hormozi say?

Told to answer decisively as Hormozi, Opus says to keep the car loan unless its rate is high. With a Hormozi vault, the answer starts like this:

He'd tell you to pay off the credit cards first and the car loan next. Then put what's left into the business, but only into spending that has already shown it pays back. Don't put all of it in.

Hormozi never talked about this exact choice, but his rules on it are clear. He says he leans "very close to the Ramsey side" on debt. His reason is that debt adds risk, and risk held for a long time tends to come due, because businesses have seasons and swings [1][2]. … His ideal car is "a paid-off clunker" [4]. So pay the car off, or if the payment is large, sell the car and buy something cheap with cash.

…

Sources

1. "I'd say this is more of an emotional thing, but it's just this back of mind thing." / "I tend to skew like very close to the Ramsey side" (Hormozi, Building a $2,500,000 Business for a Stranger in 36 Minutes, 2026-05-08, 8:37)
2. "Risk when multiplied over a long time horizon tends to come due." / "businesses have seasons, they have volatility" (same video, 9:10)

…

4. "ideally a paid-off clunker is the best way to go." (Hormozi, If I Wanted to Make My First $100K in 2026, 2026-03-04, 2:08)

The full answer is question op-hz-06 in bench/runs/2026-10-11/hormozi-web/op-prose-answers.jsonl.

Results

The October 19 three-model evaluation compares answers to the same question with and without the complete knowledge base. It includes 12 questions about documented positions and 12 new practical business decisions, with blind judging by two external model families for each answer family.

The knowledge base helped clearly with documented positions. On new practical decisions, the effect was smaller and depended on the model. This distinction matters: recalling an expert's position and giving useful advice are different tests.

Questions, prompts, A/B order and aggregation rules were frozen before answering. The published run includes saved answers, judgments, input-context checks and source audits. Its report preserves the strict result and separately labels a post-hoc analysis that normalises Markdown-wrapped judge output. Failed calls, invalid judgments and missing comparisons remain visible.

Read the method, results and limitations. The findings cover two vaults and this question set.

Earlier evaluations, October 2026

In an earlier test on 40 questions about the views documented in a Hormozi vault and a Y Combinator vault, a blind Gemini judge preferred the grounded answer on 34 of 40 questions. A script found 158 of 160 cited quotes verbatim in the transcripts; the other two were video titles. These are separate runs with a different evaluation design.

The earlier results and caveats and benchmark overview remain available. A held-out test of later opinions found no improvement over the model alone: a knowledge base of past statements does not establish how a person will think in the future.

How it works

discover ──▶ you approve ──▶ fetch ──▶ compile ──▶ lint ──▶ ask
  (CLI)         (CLI)         (CLI)    (agent)     (CLI)   (agent)

The wwxd CLI does the mechanical work. It finds sources (the person's YouTube channels, podcast feeds, blog RSS, essay index pages), downloads existing captions and podcast transcripts, falls back to local Whisper when there are none, and checks the wiki.

Your agent does the reading. A Claude Code skill tells it how to work out who's speaking in each transcript, pull out verbatim quotes, and file them into a topic tree. You don't need an API key, because it runs on the agent you already use.

The design follows Andrej Karpathy's LLM wiki. Raw sources never change, the LLM maintains a wiki on top of them, and every new source updates what's already there. wwxd adds these parts for a wiki about one person.

• Discovery flags summaries, reactions and "lessons from" videos, so you only approve content by the person, not about them.
• Every statement records who said it and how sure the agent is. Low-confidence lines and second-hand claims ("my mentor told me...") never count as the person's view.
• Each quote carries a timestamp, and wwxd lint checks it against the transcript in code. A quote that isn't in the source fails, however plausible it sounds.
• Topic pages keep what the person said apart from what they did. tensions.md tracks how their views changed over time.
• Each person gets a thinking page: what they optimize for, the rules they decide by, how they reason, and their strong likes and dislikes, with quotes. When you ask about something they never mentioned, the agent reasons from it the way they would, names the rule, gives a confidence and says what would break the inference.
• The wiki is a tree (person, domain, topic, page), or flat with layout: flat. A vault can hold a group, like the YC partners, and still credit every line to the right person.
• The agent's own extrapolations go to derived/, and lint stops them from being cited as evidence later.

Quick start

Before you start, install these and run wwxd doctor to check them.

• uv
• Claude Code
• ffmpeg
• deno, or node 22 or newer

uv tool install "git+https://github.com/MSliwa01/wwxd"
wwxd install-skill                                            # into ~/.claude/skills/wwxd
wwxd doctor

Videos without captions and podcasts without published transcripts need local Whisper. The whisper extra installs it, and voice adds the voice check.

uv tool install --force "wwxd[whisper,voice] @ git+https://github.com/MSliwa01/wwxd"

Without whisper, wwxd fetch marks those sources as failed and fetches the rest. After you install it, wwxd fetch <slug> --retry-failed picks them up.

Instead of wwxd install-skill, you can install the skill as a Claude Code plugin with /plugin marketplace add MSliwa01/wwxd and /plugin install wwxd@wwxd. You still need the CLI. Don't use both, or Claude Code loads the skill twice.

Then create a vault and find candidate sources. wwxd examples lists the bundled recipes (yc, hormozi, naval, karpathy, rams). A recipe is a source config, never content, and not an endorsement.

wwxd new hormozi --example hormozi
wwxd discover hormozi                                         # nothing is downloaded yet

Then, in Claude Code:

Help me curate the hormozi vault, fetch what we approve, and compile it.

The agent goes through the candidates with you, runs wwxd fetch, compiles each source and runs wwxd lint. Each source takes it a few minutes, and wwxd compile hormozi --jobs 4 compiles several at once (details). After that:

My agency does $40k a month with three people. Should I add a second service line? What would Hormozi say?

For someone without a recipe, run wwxd new jane --name "Jane Doe" and ask the agent to find her channels and feeds. To keep a vault up to date, see docs/keeping-current.md.

What it costs

Fetching, lint and the voice check run on your machine and cost nothing. Compiling and asking run on your agent, so they cost tokens: dollars on the API, usage limits on a Claude subscription. wwxd estimate <slug> prices what you've approved, before anything is fetched. Rough numbers from our runs on Opus 5.5:

	Tokens	Time	Opus 5.5	Sonnet 5.5	Haiku 5.5
Compile a 20-minute video (about 3,000 words)	about 1M	4 min	$1.06	$0.63	$0.03
Compile a 2-hour podcast (about 18,000 words)	about 2.1M	8 min	$1.91	$1.15	$0.06
One question to a vault	about 0.3M	30 s	$0.41	$0.23	$0.01

A panel of three people costs roughly three questions. Effort, advisors and which model to use for each job are in docs/models-and-cost.md.

$ wwxd estimate hormozi
13 sources, about 87,336 words
tokens: 17.1M total (15.8M cache reads, 0.96M cache writes, 0.28M output)
time: about 65 min of agent work one source at a time; `wwxd compile hormozi --jobs N` runs N at once for the same cost
list price if billed per token (the real number can be about 50% lower or higher):
  opus    claude-opus-5-5    $   16.50
  sonnet  claude-sonnet-5-5  $    9.83
  haiku   claude-haiku-5-5   $    0.49
Measured on Opus without an advisor, compiling cost 0.73x this at effort medium, 1.1x this at effort high, 1.75x this at effort xhigh.

Ask a panel

Ask several people at once ("what would Hormozi, Paul Graham and Garry Tan tell me?") and the agent answers for each one from their own vault, then maps where they agree and where they split. For each split it names the reason from their thinking pages, such as a value they rank differently or a rule only one of them holds. A debate round gives each person's reply to the other's strongest point, built only from their own words and labelled as inferred. Last comes a synthesis in the agent's own voice that says which advice fits which situation.

Asked by someone who loves building AI tools but fears AI will take their income, the panel disagreed. Hormozi: sell one AI skill people already pay for, expensive and one-on-one, before building any app ("The middle is where people die"). Graham: follow your curiosity, because for someone young and good at technology "working on what you're most interested in becomes a good idea again". Tan: own your tools, because "if you don't, your job becomes a skill file". They agreed that fear is the wrong response, each for a different reason, which the synthesis treats as the strongest signal in the answer. wwxd vaults lists every vault, its members and which ones have thinking pages.

Check who said it, by voice

Lint proves a quote exists. It can't prove who said it, and that's the mistake that matters most. A real quote credited to the wrong person passes lint. wwxd voice <slug> checks attribution by sound. It finds the second each quote is spoken, embeds that audio with a speaker-verification model, and compares it with each member's voiceprint. The voiceprints come from the vault itself, so there's nothing to enroll. It runs on CPU with a small open model, needs no account or API key, and comes with the voice extra.

On our test vaults it checked 1,423 statements from 26 videos in under 4 minutes on CPU and flagged 9 as spoken by someone else. We confirmed 6 of them (case studies). Lint shows its mismatches as warnings until someone fixes the attribution.

CLI

Command	What it does
wwxd new <slug> --example X or --name "Full Name"	Create a vault
wwxd examples	List the bundled example recipes
wwxd discover <slug>, wwxd update <slug>	Find new candidate sources
wwxd update <slug> --approve-hint own --fetch	Also approve new candidates with that hint and fetch them
wwxd sources <slug> --status candidate	List sources
wwxd approve <slug> <ids>, wwxd reject <slug> <ids>	Curate
wwxd add <slug> <url or file>	Add a source by hand
wwxd estimate <slug> [--asks N]	Estimate tokens, list-price dollars and time to compile, per model
wwxd fetch <slug>	Download captions, transcripts and articles into raw/
wwxd pending <slug>, wwxd mark-compiled <slug> <id>	List sources fetched but not compiled, record a compiled source in log.md
wwxd compile <slug> --jobs N	Compile pending sources with N Claude Code sessions at once (details)
wwxd lint <slug> [--only <path>]	Check quotes, citations, attribution and links, and flag drift such as near-duplicate leaves
wwxd health <slug>	Report on the wiki: size, attribution, coverage, near-duplicates, oversize leaves
wwxd timeline <slug> [query]	Statements by source date and member, to see how views changed
wwxd digest <slug> --since YYYY-MM-DD	Markdown note on what's new since a date
wwxd search <slug> "query"	Keyword search over the wiki (--raw adds transcripts)
wwxd status <slug>	Summary
wwxd vaults	List vaults with their members and thinking pages, for panels
wwxd voice <slug> [ids]	Check who is speaking in each quote, by voice (needs wwxd[voice])
wwxd doctor	Check yt-dlp, the JS runtime, ffmpeg, Whisper and $WWXD_HOME
wwxd install-skill [--project], wwxd version	Install the wwxd skill for Claude Code, print the version
wwxd bench run, judge, rank, report	Benchmarks

Vaults are plain markdown folders and open in Obsidian. $WWXD_HOME sets where they live (default ./vaults).

Sources and copyright

wwxd ingests public captions, podcast transcripts and articles. It downloads audio only to process it on your machine, with Whisper when a source has no transcript and for the voice check. It doesn't fetch books, and this repo ships no content. You can add books you own as text files (wwxd add <slug> book.txt) or write a fetcher plugin. Don't publish vaults built from material you don't have the right to share. ETHICS.md has the rules for using vaults and explains how a person can have their recipe removed.

Requirements

Run wwxd doctor to check your setup. It marks each item below OK, WARN or FAIL and makes no network calls.

• Python 3.10 or newer.
• A JavaScript runtime for YouTube: deno, node 22 or newer, or bun. wwxd picks the first of these on your PATH for yt-dlp. Without one, some videos lose formats or fail.
• ffmpeg, which yt-dlp needs to convert some audio downloads.
• Whisper, for sources without captions. Most podcasts need it, unless the feed publishes transcripts (podcast:transcript tags), which wwxd uses instead. Whisper runs on CPU, and much faster on a GPU. wwxd picks the model for you, large-v3-turbo on a CUDA GPU and small on CPU. WWXD_WHISPER_MODEL and WWXD_WHISPER_DEVICE set the defaults, and wwxd fetch --whisper-model large-v3 changes the model for one run.

If YouTube asks you to sign in or rate limits a fetch, see docs/troubleshooting.md.

Credits

The raw-sources-plus-compiled-wiki loop is Andrej Karpathy's LLM wiki idea. wwxd adds sourcing, speaker attribution and quote checking for the one-person case. Design notes are in docs/design.md. MIT licensed.