Tutorials7-Prompt Toolkit (Copy & Go)
Build, ingest, query, retrieve, restock, write, maintain — 7 prompts, each with 'when to use → prompt → how to adapt → how to verify.' Copy, paste, run.
This is the course's "arsenal." Each prompt has: when to use → prompt → how to adapt → how to verify.
How to use: replace content inside 【】 with your own; copy everything else verbatim.
Build: The First Prompt
When to use: the first time you build a knowledge base. AI will design, recommend methodology, interview you, then auto-build.
Help me build a "【your knowledge base name】".
【My background】I am ____ (one sentence: profession / role / scenario)
【Problem I want to solve】____ (e.g., materials too scattered / want to distill work experience / want a second brain)
【What I want it to help with】____ (e.g., answer business questions / organize and retrieve materials)
Please do NOT start building yet. Follow this order:
Step 1: Based on my request, explain how this knowledge base should be designed to be reliable;
Step 2: Recommend authoritative, reliable AI knowledge base methodologies (including but not limited to Karpathy's theory),
and summarize in plain language I can understand;
Step 3: If my needs are still unclear, conduct a deep interview with me,
asking one question at a time, until you fully understand my needs.How to adapt: fill in 【base name / your background / your pain / what you want it to do】.
How to verify: AI follows "design → methodology → interview" three steps, and after you finish the interview, auto-generates Schema.md, raw/, wiki/ structure = success.
Ingest: Short Command "Digest"
When to use: after dropping new materials into the library. Turns scattered materials into reusable knowledge cards.
DigestIf the command isn't recognized or in a new conversation, use the full version:
Invoke the knowledge-compiler flow to perform an incremental compilation of the current knowledge base:
Read rules under 90_系统规则, only process materials in 00_原始资料/00_待处理,
complete four-way routing (knowledge-only / project-only / both / pending), generate source summaries,
update topic knowledge, archive originals, log the operation. Report when done.How to adapt: don't change anything, just send.
How to verify: materials landed in raw/, new cards appeared in wiki/, index.md updated, log has entries = success.
Key check: if AI only gave a summary in the chat but didn't create files, it's not a successful compile. Send the full version again.
Query: Ask with Sources
When to use: when you want to pull knowledge from the library. AI synthesizes cards, labels sources, distinguishes confidence levels.
Query the current knowledge base to answer: 【your question】.
Distinguish confirmed, inferred, and pending, and provide clickable source links.Advanced version (after setting a work profile):
First read my work profile, then query the current knowledge base to answer: 【your question】.
Organize the answer based on my profile, goals, and expression preferences, while distinguishing confirmed, inferred, and pending, with clickable source links.How to adapt: replace 【your question】 with what you actually want to know.
How to verify: the answer distinguishes three confidence levels, and each claim links back to a source = success.
Three confidence levels:
| Level | Meaning |
|---|---|
| Confirmed | Library has evidence, directly verifiable |
| Inferred | AI's synthesis of multiple sources |
| Pending | Not in the library, flagged, not fabricated |
Mantra: if you can't trace it back, don't trust it.
Retrieve: Find Old Materials Before Starting New
When to use: before starting something new — let AI pull related materials from the library first. This is where the library "pays off."
I'm about to work on 【a new piece about XX / a new task】.
First, retrieve related materials from the current knowledge base: read index.md → find cards by topic/tag →
return sorted by reusability: ① directly usable card links ② relevant quotes/cases/methods ③ suggested combinations. Label sources at the end.How to adapt: replace 【a new piece about XX / a new task】 with what you're about to do.
How to verify: the returns are cards from your library (with links), not AI's generic content.
Restock: Fill Gaps from Authoritative Sources
When to use: when the library lacks content on a topic. Only supplement from authoritative sources; fabrication is forbidden.
Perform a "knowledge restock": for 【topic】, the current knowledge base is insufficient; authoritative materials are needed.
1. [Gap assessment] List what the library already has on this topic; pinpoint the gap (missing facts? cases? latest info?);
2. [Sources] Only supplement from authoritative, reliable sources: official docs/websites, academic/industry reports, credible media, professional institutions;
Do NOT use sources of unknown origin, marketing content, or unverified self-media;
3. [Source locking] Each supplemented item must be labeled: source name, link/reference, publication date;
4. [Quality control] Only use verifiable information; unverifiable or unreliable content is labeled "unverified" and excluded from conclusions;
5. [Ingest] Organize restocked content into new cards, store in the knowledge base, and update the index;
6. [Report] Return: what was added, source list, which gaps remain unfilled and why.
Note: restocking ≠ fabrication. If the library doesn't have it, say so — label "to be supplemented" rather than using unreliable sources to fill the gap.How to adapt: replace 【topic】 with your topic.
How to verify: a source list is returned, each verifiable; gaps are honestly stated = success.
Write: Library → A Deliverable
When to use: produce a file (report / article / outline / PPT) using materials from the library.
Based on the current knowledge base materials, help me write a document.
【Goal】What this document should solve: ____ (e.g., project progress for management / intro guide for students / personal review)
【Audience】Target reader: ____ (e.g., management / novice students / yourself)
【Format】____ (e.g., a Markdown article / a report / a one-page PPT outline)
Execute as follows:
1. [Library-only] Content must come from the knowledge base; no fabrication; label "to be supplemented" for gaps;
2. [Source locking] Mark sources for key conclusions so readers can trace back;
3. [Structure] Organize for the target reader: conclusion first, then detail;
4. [Deliver] Save the file to the output location;
5. [Report] Return: file title, structure outline, which library cards were used, what was labeled "to be supplemented" and why.
Note: this is "using" not "creating" — write from library materials, don't invent; it's final only after you confirm.How to adapt: fill in 【goal】【audience】【format】.
How to verify: all content has library backing, key conclusions have sources, no fabrication = success.
Maintain: Reflow + Health Check
When to use: not every time — long-term maintenance: reflow after finishing, health check periodically.
Result Reflow
After finishing → write back reusable insights from this round:
Reflow insights from this round's results that have source support and are worth long-term reuse into the relevant topic knowledge;
preserve the results and source links; do not write personal opinions as facts.Health Check
Periodically → diagnose first, fix later; report first, modify after:
Perform a read-only health check on the current knowledge base.
Group issues as "normal, warning, needs confirmation," with path, cause, and fix suggestion for each.
Do NOT fix anything this time — just produce the report.Mantra: done is not the end — keep the reusable, flag the unverified, drop the one-off.
Health check produces a report only — no unauthorized modifications; after confirmation, fix one issue at a time.
Quick Reference
| Action | When | Which prompt |
|---|---|---|
| Build library | First time | First prompt |
| Drop materials → ingest | Every time you get new materials | "Digest" |
| Want to ask → query | Pull knowledge from library | Query with sources |
| Start new → retrieve | Before starting something new | Retrieve prompt |
| Library thin → restock | Topic lacks materials | Restock prompt |
| Library → write | Produce a deliverable | Write prompt |
| Long-term maintain | After finishing + periodically | Reflow + health check |
Three Questions + Two Tools
What is a second brain / AI knowledge base? It's not a cloud drive or a chatbot — the core logic is Karpathy's LLM Wiki 'compiler' paradigm. Meet WorkBuddy (does the work) and Obsidian (shows the structure).
Hands-On: Build Your Library from Zero in 30 min
Steps 0–5: create folder → send first prompt → drop in materials → query with sources → view structure in Obsidian. Each step has acceptance criteria.