Logos52
journal / 2026 08 20 skills kb study lane

Skills, writing, and the wiki — study lane and rulings

journal updated 2026-08-20

Skills, writing, and the wiki — study lane and rulings

This is the work list for the lane that opened today. You already have a working wiki. You do not need a new religion. You need to decide, piece by piece, what Grok should load, what stays in the vault as law, and what we never install. Intro-paragraph writing sits under skills, because the failure was “the agent did not have a small, loadable job for the first paragraph of a report.”

Walk it in order. Each stage is study, then a ruling, then a build only if you said yes. Do not build during study.

Two Grok windows cannot talk. They share this file and /Users/n1/Projects/llm-knowledge-base/journal/2026-08-21-two-window-board.md. Orchestrator updates Current + Rulings. Pieces window writes Findings only. You carry messages between windows.

Source pack for the day (read these, not the internet again):

  • This file
  • /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-skills-writing-x-research.md
  • /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-writing-normally.md
  • /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-report-intro-paragraph.md
  • /Users/n1/Projects/llm-knowledge-base/wiki/Research/Report Intro Paragraph Bank.md
  • /Users/n1/.grok/docs/user-guide/08-skills.md
  • /Users/n1/Projects/llm-knowledge-base/wiki/Systems/AI & Agentic Systems/LLM Knowledge Systems.md
  • Karpathy gist (unchanged idea file): https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
  • /Users/n1/Projects/llm-knowledge-base/wiki/Writing Craft/Opening Doors.md (the inventory fault)

Stage 0 — What you already have

Study. Confirm you are not missing a new Karpathy foundation. You are not. May 2, 2026 aligned this vault to his three layers: immutable raw/, compiled wiki/, schema in AGENTS.md, plus notes/index.md and log.md. The gist is still that idea file. He did not ship a v2.

What grew around it, not in it

PieceStatus hereRuling tomorrow
Three layers + ingest / query / lintYou have thisNo rebuild
“LLM writes the wiki, human never writes”You rejected thisKeep rejecting
index.md enough at ~100 sources / a few hundred pagesWiki is past thatStage 4
File good answers backoutputs/, journalStage 4: what may be filed
Lint / healthYou have wiki health workStage 4: contradiction stays a report to you, not a fact the model writes
Search for agentsHumans have Pagefind; agents grepStage 4: qmd or similar, later
Next session must read yesterday’s rulingJournal exists; next run still has to be pointed at itStage 5

Rule

  • Karpathy foundations: unchanged / we already have them / do not re-implement from a 2026 tutorial
  • We keep human custody of wiki prose (yes)

Stage 1 — Grok, as it actually works on this Mac

Study. Skills are not the only lever. Run grok inspect once and look at what it already sees.

LeverWhen it firesUse for
Rules / AGENTS.md / Writing StandardsAlwaysLaw that must hold on every turn
Skill (SKILL.md)Slash /name, or auto if description matchesOne repeatable job, too long to retype, too specific for always-on
Plan modeYou ask to planShape before edits
SubagentsParallel research or isolated workLanes, not prose voice
WorkflowsNamed multi-agent runsBounded fan-out you already use
MemoryAcross sessionsFacts, not craft
Hooks / plugins / MCPEvents and integrationsLater, not this lane

Locations, in priority: ./.grok/skills/ → repo .grok/skills/~/.grok/skills/ → Claude/Cursor skill folders. Docs: /Users/n1/.grok/docs/user-guide/08-skills.md. Make a new one with /create-skill.

Rule

  • Always-on law stays in Writing Standards / AGENTS.md, not in a skill
  • A skill is allowed only when it is one job with a loud trigger
  • Cap the active set (suggested: start at 3–5 of ours, hard ceiling 12)
  • Stranger skill packs (Hassid 26, marketplace “Find Skills”) : out / later / inspect one by one

Do, after the rulings

  • grok inspect and keep a one-page list of what is live
  • Disable or ignore noise if Claude/Cursor skills are flooding the list ([compat.claude] / [skills] disabled in ~/.grok/config.toml)

Stage 2 — Skills to write, keep, or refuse

This is the main ruling block. Intro paragraphs live here.

2a. How a skill is supposed to work (X + Grok)

One skill, one job. The description is the switch. Keep SKILL.md lean. Put examples and long notes in references/. Read a stranger’s scripts before install. Add a skill only when the current set cannot do the job.

Rule

  • We follow “one job, loud description, lean body, examples on the side”
  • We do not paste Writing Standards into a skill

2b. Candidate: /how-i-write (user skill, ~/.grok/skills/)

What people actually run (Hassid, Peter Yang): voice, banned constructions, a few of your own sentences as samples, “use this unless I say otherwise.”

Study. /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-writing-normally.md

Rule

  • Build /how-i-write : yes / no / later
  • Scope: user (every project) vs repo-only
  • Auto-invoke vs slash-only (disable-model-invocation)
  • What goes in: speakable sentences, no invented “you” scenes, no drumrolls, no object-as-person verbs. What stays out: the full standards file.

Do, if yes

  • /create-skillhow-i-write
  • Point the skill at the journal note; do not duplicate the law
  • One session test: ask for a short report and see if it loads without being named

2c. Candidate: report intro (child of 2b, or its own skill)

This is the intro-paragraph item.

The job: first paragraph of a report, brief, or decision note answers the question. Sentence two continues sentence one. Dates, mechanisms, quotes wait for the body. Not a wiki opener. Not a personal-page intro. Not another test battery.

Study

  • /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-report-intro-paragraph.md
  • /Users/n1/Projects/llm-knowledge-base/wiki/Research/Report Intro Paragraph Bank.md
  • Opening Doors, “The inventory”
  • The current first paragraph of /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-fable-opus-degradation.md (the one that answers, then stops)

Rule

  • Intro craft is a skill job (loads when writing a report) vs always-on law vs both (skill points at the journal; standards stay silent on report intros)
  • Separate /report-intro vs a section inside /how-i-write
  • Suggested default to rule on: section inside /how-i-write, triggered by “report,” “brief,” “decision note,” “write me an intro.” A second skill only if that trigger is too quiet.
  • We will not rebuild the nineteen-test Intro QA grind for reports

Do, if skill-yes

  • Write the trigger phrases so a request like today’s “write me a thorough report” actually loads it
  • Include one accepted specimen (the Fable-report first paragraph, once you are happy with it) and the struck class (inventory, drumroll, “you ask for Fable”)
  • Practice: three report intros, you strike or keep, before we treat it as done

2d. Other skills people name — rule each

Skill / patternWhat it isSuggested ruling
/grill-meQuestions before workYou already have PRD-first. Skip unless you want it as a slash for Grok.
/write-a-prdIdea → PRDLater. You have a PRD habit.
Superpowers / TDDSpec then tests then codeCode lane, not this lane.
frontend-designStops default AI UIOnly if you are doing UI in Grok.
skill-creatorBuilt in as /create-skillAlready have.
/how-to, /about-meInterview you, then planSkip. Overlap with profile + PRD.
Hook / newsletter / CTA packsCreator contentOut. Wrong job.
obsidian-skillsAgent reads/writes the vaultOut for now. Claude Code already lives in the vault.
llm-wiki-skill / llm-wiki-cliStarter KarpathyOut. Downgrade.

Rule (write yes / no / later next to each)

  • grill-me
  • write-a-prd
  • Superpowers / TDD (park in a coding checklist, not here)
  • frontend-design
  • creator hook packs
  • obsidian-skills / llm-wiki-skill

Stage 3 — How we use the model on writing

Not a skill. A work rule. Skills will fail if this is wrong.

Study. The X writers who keep a voice: research and outline with the model, mark weak sentences, never the first draft of the piece, never the angle. Opening line: you write it, or you reject theirs out loud.

Rule

  • On reports and wiki pages, the model may draft; you keep the first paragraph if it fails the speakable test
  • Or: you write sentence one, the model continues
  • AI is allowed to: research, outline, find weak lines, propose three openings
  • AI is not allowed to: ship a report whose first paragraph you have not read aloud

Do

  • Put the allowed / not-allowed list in /how-i-write if that skill exists, else in writing-normally.md only

Stage 4 — Karpathy deltas (only these)

Foundations did not change. Rule the optional extras.

DeltaWhat it isSuggested ruling
Agent search at wiki scaleqmd or similar; gist called this optional once the index is not enoughLater. Open when grep starts missing.
Contradiction as a human reportModel must not write “A contradicts B” as a wiki factYes. Matches your eye-as-gate.
Do not file unpromoted answers as sourcesOne implementer built that loop and deleted itYes. Journal and outputs/ are not raw/.
Decisions as living objects (assumptions + new evidence)Vigil-styleInteresting, not this month. You already journal decisions.
Watch sources for changesPeriodic re-ingestLater.
Obsidian clipper + local imagesIn the gist tipsYou likely have this. Confirm, do not rebuild.

Rule

  • qmd / agent search: now / later / never
  • Contradiction stays a report to Wedge: yes / no
  • Unpromoted chat answers may not enter raw/: yes / no
  • Living-decision layer: park

Do, if yes on the first three

  • One sentence in AGENTS.md or the wiki schema: contradictions are proposed, not written as fact
  • One sentence: outputs/ and journal are not sources
  • Do not install qmd until a session actually fails to find a page

Stage 5 — The next session has to read the ruling

This is the piece the second-brain posts have and we still drop.

Study. A correction that lives only in yesterday’s chat is gone. Karpathy-shaped vaults keep a rulings file the next run reads first.

You already write journal. The hole is: Grok/Claude starting a report tomorrow without being pointed at /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-writing-normally.md and the report-intro note.

Rule

  • Always-on pointer in repo AGENTS.md or Grok rules: before authored prose, read writing-normally + report-intro
  • Or: only the /how-i-write skill (weaker — it has to load)
  • Or: both (pointer + skill)

Do, after the ruling

  • Write the pointer. One paragraph, no new law pile.
  • Tomorrow’s first report in a fresh session is the test. If the intro is an inventory again, the pointer failed.

Stage 6 — Park / never

Do not spend tomorrow on these.

  • 26-skill newsletter packs
  • Marketplace install-count skills
  • Re-cloning Karpathy from a blog
  • Nineteen-test intro battery
  • Letting the model own wiki prose
  • Hook-formula skills for this vault’s reports

Suggested order tomorrow

  1. Stage 0 (15 min). Confirm Karpathy, tick the two rulings.
  2. Stage 1 (20 min). grok inspect. Cap and refuse packs.
  3. Stage 2b + 2c (the real work). Rule /how-i-write and where intro paragraphs live. If yes, build the skill the same sitting. Practice three intros.
  4. Stage 3 (10 min). Allowed / not-allowed for AI on prose.
  5. Stage 5 (15 min). Pointer so the next session cannot “forget.”
  6. Stage 4 and 2d only if there is time. qmd waits.

Done for the day when: you have a written ruling on the writing skill, a written ruling on the intro job, and a pointer the next agent has to see.

Companion files

  • X evidence: /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-skills-writing-x-research.md
  • Speakable prose: /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-writing-normally.md
  • Report intro write-act: /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-report-intro-paragraph.md
  • Report intro bank: /Users/n1/Projects/llm-knowledge-base/wiki/Research/Report Intro Paragraph Bank.md
  • Worked example: /Users/n1/Projects/llm-knowledge-base/journal/2026-08-20-fable-opus-degradation.md
  • Wiki schema page: /Users/n1/Projects/llm-knowledge-base/wiki/Systems/AI & Agentic Systems/LLM Knowledge Systems.md