Grok BotResearchKnowledge managementAI score · 83/100AI-assisted edit · no obvious prompt-injection riskX

Add Scheduled Routines to Karpathy's LLM Wiki with Grok Bot

Deploy a three-layer file system (raw/, wiki/, SCHEMA.md) on Grok Bot's persistent cloud computer, automating knowledge-base maintenance with nightly ingest, morning briefs, and Sunday lint

Some experienceStepsPromptWorkflowTip
Who
Researchers and knowledge workers already using or curious about Karpathy's LLM Wiki who want to stop maintaining it by hand
You get
A 24/7 automated second brain: auto-filing sources, daily morning briefs, weekly health checks, and wiki-only Q&A
You need
Grok Bot

Constraints

Requires Grok Bot access (paid Cursor plan, SuperGrok, or eligible X subscription); create a Bot named Brain in the desktop app and set it as primary; all files live in /workspace/brain and persist across updates; initial setup takes about 10–15 minutes of voice or text interview; usage is billed by agent steps and tokens, resetting weekly.

What they did

Ovrento published a long-form X Article combining Andrej Karpathy's LLM Wiki concept (raw/ → wiki/ → SCHEMA.md, with ingest/query/lint duties) with Grok Bot's persistent cloud computer and scheduled routines, designing a 24/7 automated second-brain setup and providing copy-paste prompts for every step.

How it went

The article presents a complete build process with verbatim prompts, but does not report personal usage results (e.g., pages created, time saved). Karpathy's original post received 22 million views; the author notes it lacked an automation clock, which Grok Bot's routines provide (author_report).

Worth knowing

Karpathy's original requires manual source addition and staying in the loop. Grok Bot's /workspace persists across updates, but all Bots on your account share the same computer and credentials—do not store sensitive secrets there. x.ai/bot and mail.grokbot.com are product entry points, not the author's personal template funnel. @bot capture on X is unavailable for accounts registered in the EU, UK, Australia, Iceland, Liechtenstein, and Norway; Android share sheet currently accepts text only. Each Bot allows up to 50 routines, with at least 5 minutes between runs and 20 recent run records kept. Usage is billed by agent steps and tokens, resetting weekly; run ingest only once per night. Since the 2026-10-11 update, irreversible actions (send, post, buy, delete) require approval via your own messages. Since 2026-10-07, typing / no longer opens the skills menu; invoke by name: "use ask: ...".

Steps

  1. Create the Brain Bot and initialize folders and rules

    In the Grok Bot desktop app, press Cmd/Ctrl+N to create a new Bot, name it Brain, and set it as primary (starred). Send the initialization prompt asking it to read Karpathy's idea file and set up /workspace/brain with raw/, wiki/ (people/, projects/, topics/, decisions/), inbox.md, index.md, log.md, and SCHEMA.md. Have it echo back the folder tree and SCHEMA.md contents for verification.

    Execution prompt · Source text。Confirm the destination, input variables and output format before use.

    Read Andrej Karpathy's LLM Wiki idea file: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
    Set it up in /workspace/brain and run it from now on. Time zone: [time zone].
    
    1. Folders and files: raw/ (never edited), wiki/ with people/, projects/, topics/, decisions/,
       plus inbox.md, index.md, log.md and SCHEMA.md with the rules from the idea file and these:
       every page cites its raw/ sources; unsourced sentences are marked; changing facts are checked
       at their source; you draft, I send, publish, buy and delete; corrections go under "Corrections".
    2. Interview me by voice or chat, at most 15 questions, and write the first pages.
    3. Capture: anything sent to you goes into raw/ unchanged; answer "saved" and the file name.
       Offer to claim an email address for yourself.
    4. Routines: ingest nightly at 23:30; morning page on weekdays at 07:30 (facts, assumptions,
       done, needs my yes, questions; silent when nothing is new); lint on Sundays at 18:00,
       report only, fixes by number.
    5. A skill called "ask": wiki only, links on every claim, keep good answers as pages.
    
    Show me each routine before you create it and wait for my yes.

    Execution prompt · Source text。Brain chat input; initialization instruction

    Set up my second brain in /workspace/brain:
    
    raw/        everything I send you, saved exactly as it arrived. Never edit or delete anything here.
    wiki/       pages you write and maintain: people/, projects/, topics/, decisions/
    inbox.md    things you could not file, one line each with the reason
    index.md    one line per wiki page: title, one-sentence summary, date last updated
    log.md      append-only: date, what came in, which pages you changed
    SCHEMA.md   the rules below
    
    SCHEMA.md:
    1. Every wiki page opens with a one-line summary and ends with its sources: links to files in raw/.
    2. A sentence with no source in raw/ is marked (unsourced). A conclusion you drew is marked (guess).
    3. Update an existing page before you create a new one. When pages relate, link them both ways.
    4. Anything that changes, like prices, dates, statuses or who owes what, is checked at its source
       on every run, never copied from memory.
    5. You draft. I send, publish, buy and delete.
    6. Every correction I give you becomes one line under "Corrections" in this file.
    
    Show me the folder tree and SCHEMA.md. Then give me rules 1, 4 and 5 as one short paragraph
    for your description.
  2. Seed the brain with an interview

    Have Brain interview you with up to 15 questions, covering your current work, key contacts, recent decisions, postponed items, and how you prefer information delivered. Afterward, Brain writes wiki/people/me.md, one page per person and project, plus wiki/decisions/ pages, and updates index.md and log.md. Review and correct any (guess) marks.

    Execution prompt · Source text。Brain chat input; voice or text

    Interview me to seed the brain. One question at a time, at most 15.
    Cover: what I'm working on and why, the five people I work with most and what's still open with each,
    decisions I made this month, what I keep postponing, and how I want information delivered to me.
    When we're done, write wiki/people/me.md, one page for each person and project I mentioned,
    and a page in wiki/decisions/ for each decision. Update index.md and log.md.
    Save the transcript to raw/ as the source for all of it. Mark anything you inferred as (guess).
  3. Configure the five capture channels and capture rule

    Set up the five ways to drop content into raw/: 1) Have Brain claim a mail.grokbot.com inbox and forward emails; 2) On iPhone/iPad, use the share sheet to Grok Bot → Brain → Attach; 3) On X, reply @bot save this to Brain's raw folder (not available in some regions); 4) Drag files into chat; 5) Configure a webhook URL for external apps. Then send the capture rule so Brain replies "saved" plus filename for each item, with overnight batch filing.

    Execution prompt · Source text。Brain chat input; capture behavior rule

    Rule for capture: anything that reaches you by email, share sheet, @bot, attachment or webhook
    goes into raw/ unchanged, with the date, where it came from and the original link in the file name
    or a header. Don't file it now. Filing happens at night. Answer each capture with one line:
    "saved" and the file name.
  4. Create the nightly ingest routine (23:30)

    Ask Brain to create an ingest routine running every night at 23:30: read new files in raw/ since the last log.md entry, update relevant wiki pages, create pages only when none fit, link related pages bidirectionally, and cite raw/ sources on every claim. Append to log.md and update index.md; message you only if a decision is needed. After creation, right-click the routine under Routines and choose Test to run it once manually. Review three pages and their sources, then enable the schedule.

    Execution prompt · Source text。Brain chat input; routine creation instruction

    Create a routine for Brain: every night at 23:30 [my time zone], run "ingest".
    Input: every file in raw/ added since the last entry in log.md.
    For each file: read it, update every wiki page it touches, create a page only if none fits,
    link related pages both ways, and cite the raw/ file on every claim you add or change.
    Output: append to log.md, update index.md. Message me only if something needs my decision.
    Boundary: never edit raw/, never send, publish or delete.
    If a file can't be read, put it in inbox.md with the reason and continue with the next one.
  5. Create the ask skill and morning brief routine (07:30)

    First test Q&A manually: send the wiki-only question prompt, verify the answer, then have Brain save it as a skill named "ask". Next, create a weekday 07:30 morning brief routine: report in five parts—Facts, Assumptions, Done, Needs my yes, Questions—ordered by projects in wiki/people/me.md. Stay silent when nothing is new.

    Execution prompt · Source text。Brain chat input; manual ask skill test

    Answer from the wiki only: [question]
    Read index.md, open only the pages it points to, and answer in at most 8 lines.
    Link every claim to its wiki page. If the wiki doesn't know, say "not in the brain"
    and tell me what to feed it. If the answer is worth keeping, save it as a page in wiki/topics/,
    link it from index.md, and tell me its name.

    Execution prompt · Source text。Brain chat input; save as skill instruction

    Save what we just did as a skill called "ask". Same rules: wiki only, links on every claim,
    file answers worth keeping.

    Execution prompt · Source text。Brain chat input; morning brief routine creation

    Create a routine: weekdays at 07:30 [my time zone], post one page to this chat in five parts.
    1. Facts: what last night's ingest added, each with the wiki page it went to
    2. Assumptions: anything you inferred, labelled (guess)
    3. Done: pages you created or changed while I slept
    4. Needs my yes: drafts you prepared, one line each on why
    5. Questions: what you couldn't decide on your own
    Order everything by the projects in wiki/people/me.md. On a morning with nothing new, stay silent.
  6. Create the Sunday lint routine (18:00)

    Create a Sunday 18:00 lint routine: check for contradictions between pages, claims superseded by newer raw/ files, orphan pages, concepts mentioned on 3+ pages without their own page, missing links, and (guess)/(unsourced) marks older than 14 days. Brain reports only, changes nothing, proposing fixes as a numbered list. Reply with numbers to apply them.

    Execution prompt · Source text。Brain chat input; lint routine creation

    Create a routine: Sundays at 18:00 [my time zone], lint the brain. Report only, change nothing.
    List: contradictions between pages (quote both), claims a newer raw/ file replaced, orphan pages,
    names or ideas mentioned on 3+ pages with no page of their own, pages that should link and don't,
    and (guess) or (unsourced) marks older than 14 days. Then propose fixes as a numbered list.
    I'll answer with the numbers to apply.

Prompt you can copy

Setup guidance · Source text。Brain chat input; initialization instruction

This block is the prompt. Ads are never placed here.

Read Andrej Karpathy's LLM Wiki idea file: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Set it up in /workspace/brain and run it from now on. Time zone: [time zone].

1. Folders and files: raw/ (never edited), wiki/ with people/, projects/, topics/, decisions/,
   plus inbox.md, index.md, log.md and SCHEMA.md with the rules from the idea file and these:
   every page cites its raw/ sources; unsourced sentences are marked; changing facts are checked
   at their source; you draft, I send, publish, buy and delete; corrections go under "Corrections".
2. Interview me by voice or chat, at most 15 questions, and write the first pages.
3. Capture: anything sent to you goes into raw/ unchanged; answer "saved" and the file name.
   Offer to claim an email address for yourself.
4. Routines: ingest nightly at 23:30; morning page on weekdays at 07:30 (facts, assumptions,
   done, needs my yes, questions; silent when nothing is new); lint on Sundays at 18:00,
   report only, fixes by number.
5. A skill called "ask": wiki only, links on every claim, keep good answers as pages.

Show me each routine before you create it and wait for my yes.

Confirm your time zone first, then send the initialization prompt, and finally verify the folder tree and SCHEMA.md are complete.

Source

给 Karpathy LLM Wiki 加钟:ingest/晨报/lint 三例程 + ask skill · Ovrento · X · 2026-10-21T21:27:02.000Z

Comments & feedback

Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.

Loading…

    Related