FoF David Deutsch projects — AI Agent Skills

Vector search over a corpus of interview transcripts (David Deutsch by default). Returns the questions closest to yours that were actually put to the speaker, each with the answer given, the source transcript, and a timestamped link.

This is retrieval only: no AI model writes anything, nothing is billed, and no account is needed.

When to use

• "What did Deutsch say about X?", "Find the quote where he talks about Y", "Did he ever discuss Z?"
• Checking a claim about his views against the record before repeating it.
• Gathering cited passages to reason over yourself. If the user wants a synthesized answer written by the QRAG service, use skills/qrag/qrag-ask/README.md instead.

Setup

skills/qrag/_shared/references/setup.md — check fof-qrag status once per session.

Procedure

1. Phrase the query as a question someone might have asked the speaker. The index is built from interview questions, so "Why are good explanations hard to vary?" retrieves far better than keywords like "explanations hard-to-vary".
2. Run the search:

   fof-qrag search "Why are good explanations hard to vary?" --json
   

   Options: -n 20 for more passages (2–50, default 10); --from 2015-01-01 --to 2020-12-31 to limit by interview date (give both); --corpus fda-town-halls for another corpus.
3. Read the result:
   • match.quality — good (the corpus contains essentially this question), partial, or none (nothing on point; sources are loosely related at best). match.max_similarity is the number behind it. Say so rather than stretching weak matches.
   • sources[] — question, answer, source (transcript filename), timestamp (markdown link that opens the recording at that moment), stars (editor rating 0–5 of how good the passage is), similarity.
4. Try one or two rephrasings if the best similarity is low; different wordings of a question surface different passages.
5. To read around a passage, pass its source to the transcripts skill: fof-qrag transcripts excerpt "<source>" (skills/qrag/qrag-transcripts/README.md).

Reporting results

• Quote the answer text as given and attach its timestamp link and source title, so the user can hear it in context.
• The answer text is a lightly cleaned rendering of speech, not a published essay. Attribute it as something said in that interview on that date.
• Distinguish clearly between what the passages say and your own summary or inference.
• Prefer passages with more stars when several say similar things.

Failure modes

code	Meaning	What to do
PRIVACY_CONSENT_REQUIRED	The user has not accepted the privacy notice	Ask the user; see setup
QUESTION_TOO_LONG	Over 500 characters	Shorten to the core question
BAD_DATE_RANGE	One date without the other, or wrong format	Give both as YYYY-MM-DD
SEARCH_FAILED / NETWORK_UNREACHABLE	Service or network problem	Retry once, then report the message

Full QRAG: search the corpus, then have the Focus on Foundations service write an answer grounded in the retrieved passages. Supports follow-up questions in a saved conversation.

Each answer is a paid AI call on Focus on Foundations' account. Use it when a synthesized answer is wanted; use skills/qrag/qrag-search/README.md when cited passages are enough.

When to use

• The user asks a question about the speaker's ideas and wants an explained answer, not just quotes.
• A follow-up that depends on the previous answer ("how does that relate to fallibilism?").
• Not for bulk or automated querying: one deliberate question at a time. For many lookups, search instead.

Setup

skills/qrag/_shared/references/setup.md — check fof-qrag status once per session.

Procedure

1. Ask:

   fof-qrag ask "Why does Deutsch say optimism is a duty?" --json
   

   The call usually takes 10–60 seconds; progress goes to stderr. Same options as search: -n, --from/--to, --corpus.
2. Read the result: ai_answer, sources[] (the passages the answer was grounded in — same fields as search), match, model, cost_usd, and conversation (conversation_id, title, exchanges).
3. Follow up in the same conversation so earlier turns are sent as context (the service uses the last 12 messages):

   fof-qrag ask "How does that connect to his view of problems?" -c last --json
   

   -c takes a conversation id, a unique id prefix, or last. Without -c every question starts a new conversation.
4. Use --no-save for a one-off the user does not want kept (nothing is stored locally or synced).
5. Model choice. Without --model the standard model answers — use it by default. An account with special access may have more (fof-qrag access status lists them): fof-qrag ask "..." --model gpt-6-astra. Signed-in AI calls draw on the person's shared lifetime website allowance, so use one only when the user asks for it, and never in a loop. Astra can take a few minutes.
6. Review or export earlier work: fof-qrag conversations list, fof-qrag conversations show last, fof-qrag conversations export last --out answer.md.

Reporting results

• Present the AI answer as the QRAG service's synthesis, and cite the sources with their timestamp links. The sources are the evidence; the AI answer is an interpretation of them.
• If match.quality is partial or none, the answer leans on general background rather than on-point passages — tell the user.
• Mention the cost only if the user asks or is making many calls.

When the AI answer is refused

Focus on Foundations controls who may make paid calls. A refusal exits with code 3 in text mode or 1 in JSON mode, prints the reason, and still returns the search results (search_results in MCP and CLI JSON, printed passages in CLI text mode) — use them.

code	Meaning	What to do
LLM_ACCESS_DENIED	AI answers need a signed-in account with access	Tell the user; fof-qrag login, then fof-qrag access status (skills/qrag/qrag-account/README.md)
LLM_QUOTA_EXCEEDED	This month's separate AI-request count limit is used up	Report it; continue with search
LLM_SPEND_CAP_REACHED	The shared lifetime website allowance is used up	Continue with search; do not retry as a guest or with another model
MODEL_NOT_ALLOWED	The account cannot use that model	Use a model from fof-qrag access status, or none
LLM_TIMEOUT / LLM_FAILED	Model or service problem	Retry once; then report
Do not try to work around a refusal (other environments, other accounts, repeated retries).

List the transcripts behind the corpus, read the public excerpts, and — for invited collaborators with special access — read full transcripts.

Two tiers

Tier	What it is	Who can read it
Excerpt	The top-rated passages of each transcript: verbatim passages (transcript) or extracted question-and-answer blocks (qa). The same material the website shows.	Anyone
Full	The complete transcript: vrb verbatim, read readability-edited, qa all extracted Q&A.	A signed-in account that Focus on Foundations has given transcript access (by email; there is no code to enter) and that has accepted the transcript terms — or anyone signed in, for a transcript whose rights holder has openly released it (release: open)
The service decides access. This skill never works around a refusal.

When to use

• "Which interviews are in the corpus?", "Find the 2018 conversation with …", "Show me more of that interview."
• Reading around a passage that skills/qrag/qrag-search/README.md found: pass the result's source filename straight in.

Setup

skills/qrag/_shared/references/setup.md — check fof-qrag status once per session.

Procedure

1. Find the transcript. Words from the title and a date range both filter:

   fof-qrag transcripts list popper --json
   fof-qrag transcripts list --from 2018-01-01 --to 2019-12-31 --json
   

   Each row has id, name, date, youtube, page. With --access (needs sign-in) rows also carry full: {release, available, formats}.
2. Read the public excerpt. The reference can be an id, a unique fragment of the title, or a source filename from a search result:

   fof-qrag transcripts excerpt "2011-08-01_KERA Think Radio with Krys Boyd_qafixed.md"
   fof-qrag transcripts excerpt 2011-08-01-kera-think-radio-with-krys-boyd --kind qa
   

3. Read the full transcript only when the user asks for it and excerpts are not enough:

   fof-qrag transcripts full 2011-08-01-kera-think-radio-with-krys-boyd --format vrb --out kera.md
   

   Transcripts are long (often 50,000+ characters). Write to a file with --out and read the part you need, rather than printing the whole text into the conversation.
4. Transcript format: a ## metadata header (links, length), then ### transcript with speaker-labelled blocks whose timestamps link into the recording, or ### qa with QUESTION / TIMESTAMP / ANSWER / TOPICS / STARS blocks.

Copyright rules for full transcripts

Full transcripts are copyrighted by their rights holders and are shared with selected collaborators for private research and study. Every full transcript comes back with a notice; honor it:

• Quote short passages with attribution and the timestamp link. Do not reproduce a full transcript, or a substantial part of one, in any output, file, post or document meant for other people.
• Do not upload full transcripts to other services, datasets, or shared folders.
• Summaries, analysis and short quotation are fine.
• A transcript with release: open has been released by its rights holder; its notice states the terms. Excerpts are already public.

When full access is refused

code	Meaning	What to do
NOT_SIGNED_IN	Full transcripts need an account	The user runs fof-qrag login
TRANSCRIPT_ACCESS_REQUIRED	The account has no transcript access	Full transcripts are for invited collaborators. The user can ask with fof-qrag access request (skills/qrag/qrag-account/README.md). Offer the excerpt meanwhile
TERMS_REQUIRED	The transcript terms changed since the user accepted	The user runs fof-qrag access terms --accept
TRANSCRIPT_NOT_FOUND	No full transcript is published under that id	Check the id with transcripts list --access
ACCOUNTS_UNAVAILABLE / NOT_CONFIGURED	This environment has no account or access service yet	Excerpts still work; report it

Help a user set up the fof-qrag toolkit and choose how their conversations are stored: local only, or synced with their Focus on Foundations account. Also covers special access (extra AI models, full transcripts) and the shared lifetime allowance.

Several steps here involve the user's credentials or their agreement to terms. Those are the user's to do, in their own terminal. Your part is to explain, to run the read-only checks, and to run sync when asked.

When to use

• First-time setup, or any QRAG skill reported PRIVACY_CONSENT_REQUIRED, NOT_SIGNED_IN, TRANSCRIPT_ACCESS_REQUIRED, TERMS_REQUIRED or LLM_ACCESS_DENIED.
• "Sync my conversations with the website", "Do I have special access?", "Are my chats private?", "How much of my allowance is left?"

Setup

skills/qrag/_shared/references/setup.md — start with fof-qrag status --json.

What is stored where

Mode	Conversations	The account service sees
Local only (default)	Files in ~/.fof-qrag/conversations/, owner-readable only	Nothing. No account needed
Synced	Local files plus a copy in the user's Focus on Foundations account; the website's conversation history shows them and vice versa	The conversation text, unless encryption is on
Independently of this, the hosted services log each question and answer (with a keyed hash of the IP address) to operate and improve the service — that is what the privacy notice covers. fof-qrag ask --no-save keeps an exchange out of the user's own history; it does not change service logging.

Procedures

Accept the privacy notice (user decision)

Tell the user what it covers (above) and where to read it (<site>/privacy/ from fof-qrag status). They run fof-qrag init, or instruct you to run fof-qrag init --accept-privacy. Never accept without being told to.

Sign in (user does this)

fof-qrag login                 # email + password, or leave the password blank for an emailed code
fof-qrag login --email-code    # emailed one-time code only

Accounts are created on the website. Tokens are kept in ~/.fof-qrag/credentials.json (owner-readable only); the password is never stored. fof-qrag logout revokes and deletes them. If status says accounts are not available in this environment, only local mode applies there for now.

Sync conversations

fof-qrag sync --dry-run --json   # show what would move
fof-qrag sync                    # two-way: newer copy wins
fof-qrag config set sync auto    # optional: push after every answer

Sync never deletes on its own. If a conversation was deleted on one side it is reported (removed_in_account, removed_locally); fof-qrag sync --prune carries the deletion across. Confirm with the user before --prune or fof-qrag conversations delete.

Special access

There is no code to enter. Focus on Foundations grants special access to an email address; an account with that verified address simply has it. It can unlock additional AI models, full transcripts, or both. Everyday use — search, standard AI answers, public excerpts — does not need it.

• Check: fof-qrag access status says "You have special access." or not, and lists the models and lifetime website allowance.
• Ask for it (the user's decision; it sends their address and a short message to Focus on Foundations): fof-qrag access request --message "who I am and what I'd like to use it for". One request per address per day is accepted; do not repeat it.
• Full transcripts additionally need the user to accept the transcript terms of use once — copyright acknowledgement and no publishing without permission: fof-qrag access terms --accept. Do not run that for the user with --accept-terms; agreeing to terms is theirs to do.

Check access and allowance

fof-qrag access status --json

access.specialAccess, access.allowedModels (standard model first), access.spend {period, scope, usedUsd, capUsd}, access.transcriptsFull, access.llm, access.termsAccepted, and policy (what Focus on Foundations currently requires). The allowance defaults to $1.00 total across paid website features, with period: "lifetime" and scope: "account"; it never resets monthly. When it is used up, signed-in paid calls stop on every model. Continue with search or ask Focus on Foundations for a higher limit; never drop the account ticket to bypass the cap.

Encryption (opt-in, early)

fof-qrag config set encryption aesgcm-pbkdf2-v1 encrypts conversations with a passphrase (FOF_QRAG_PASSPHRASE, or prompted) before they are written or synced, so nobody with database access can read them. Tell the user the current limits before they turn it on: the website cannot display encrypted conversations yet, and a lost passphrase cannot be recovered. It needs the optional cryptography package.

Failure modes

code	Meaning	What to do
SIGN_IN_FAILED	Wrong email, password or code	The user retries; password reset is on the website
TOO_MANY_ATTEMPTS	Sign-in attempts are rate limited	Wait; do not loop
TOO_MANY_REQUESTS	A special-access request from this address is already waiting	Nothing to do; do not resend
LLM_SPEND_CAP_REACHED	The shared lifetime website allowance is used up	Continue with search; ask Focus on Foundations for a higher limit
MODEL_NOT_ALLOWED	The account cannot use that model	fof-qrag access status lists what it can use
CHILD_ACCOUNT	Child accounts cannot hold transcript or AI-answer access	A guardian account is needed
CONVERSATION_TOO_LARGE	Over the 300 KB per-conversation sync limit	It stays local; start a new conversation
DECRYPT_FAILED / PASSPHRASE_REQUIRED	Wrong or missing passphrase	The user supplies it; it cannot be reset