Restoring Claude Archives as an LLM Wiki: Turning 473 Prompts Into Knowledge Nodes

When prompts accumulate, they eventually stop feeling like assets and start behaving like dust. The titles sound useful, but when you open the files, some are empty, some are fragments of old experiments, and some no longer show where the actual working prompt went.

That is the current state of MyZettelkasten/07 Resources/AI Agents/Claude_Archives. There are many files. But a large file count is not the same thing as a working knowledge system.

This post is not an attempt to publish the folder as a public prompt pack. It is the first restoration pass: read the archive, separate live nodes from empty shells, and turn the folder into an LLM Wiki that can later produce public, reusable articles one node at a time.

What I Verified First

The folder currently contains 478 Markdown files. Of those, 471 are closer to migration templates than finished prompts. Many contain a placeholder instruction with the same basic meaning:

Paste the original system prompt here.

So most files are not ready-to-publish prompt assets. They are restoration markers. A small number of longer files still contain enough structure to become public articles.

1. 00_DNA_Registry.md
2. TpT_상점_설명.md
3. System_TPT_Store_Optimizer.md
4. GEN2_Moonlit Quest Weaver.md
5. GEN2_정량적 영어 원서 워크시트 생성기.md
6. GEN2_SlangSavvy Wordplay Quest.md
7. GEN2_React 증상 분석가.md

The implication is simple: the task is not to publish 478 prompts. The task is to reclassify 478 traces into knowledge nodes.

What Changes When This Becomes an LLM Wiki

As a folder, the archive is hard to reason about. As an LLM Wiki, each file can be assigned one of four states.

State	Meaning	Treatment
Live node	It still has a role, instructions, and usable context	Refine into a blog article
Restoration node	It mostly has title and metadata	Recover from .txt, Claude exports, or earlier conversations
Derived node	It is a second-generation mutation of an earlier prompt	Connect it to its parent node
Retirement candidate	It overlaps too much or no longer fits the system	Keep it indexed, but do not publish it yet

In this frame, GEN2_Moonlit Quest Weaver is not merely a game prompt. It is a node for turning learning content into quests. GEN2_정량적 영어 원서 워크시트 생성기 is a node for converting original English texts into quantified worksheet sections. TpT_상점_설명 is a marketing node that turns educational materials into marketplace descriptions.

The filename is storage. The node is the role. Future cleanup should be organized around roles, not filenames.

The First Parent Node: Claude Project DNA Registry

The folder includes a short file called 00_DNA_Registry.md. It is brief, but it contains the right orientation.

• Total nodes: 473
• Generation: 1
• Last mutation: 2025-12-13
• Actionable genes: generation, analysis, transformation, simulation

The useful idea here is that prompts were being treated as a phylogenetic tree, not as a flat list. The question was not only “What prompts do I have?” but also “Which prompts descended from which earlier nodes? What mutations appeared? Which functional genes repeat?”

That framing still works. It actually fits the current blog workflow even better. A prompt should not be published as a raw instruction block. It should be translated into a reusable knowledge node:

Raw prompt
-> role definition
-> input conditions
-> process
-> output contract
-> verification criteria
-> use case
-> next node

That is the basic conversion formula for turning a prompt into an LLM Wiki node.

The Refinement Harness

When I refine files from Claude_Archives into blog posts, I will use this harness.

Step	Question	Output
1. Classify	Is this a live node or a restoration node?	Publishability
2. Role it	What repeated problem does this prompt solve?	Node name
3. Structure it	What are the inputs, process, output, and checks?	Harness table
4. Link it	What are the parent and child nodes?	LLM Wiki links
5. Apply it	Where does this fit in the blog, classes, apps, or worksheets?	Use scenario
6. Publish it	What should be shown, and what should be explained instead?	Blog article

The key rule is not to paste the raw prompt as the article. A prompt is an execution artifact. A blog post is an explanation a reader can understand and reuse.

Category Map

After tracing the available source files, Claude_Archives can be grouped into 11 working families. The count matters, but the more important signal is the connection path: not where a file happened to sit, but what repeated job it was trying to perform.

Category	Count	Representative nodes	Operating flow
Prompts, harnesses, and agents	106	PCHL Engineering, Claude Project DNA Registry, Harness Engineering	Blog principles, Codex skills, LLM Wiki
English education and original-book learning	73	정량적 영어 원서 워크시트 생성기, SlangSavvy Wordplay Quest	ReadMaster, /k-kid-book, worksheets
Visualization, images, and slides	52	Mermaid 다이어그램 생성기, 텍스트 분석 시각화, PPT 생성기	Blog diagrams, lesson slides, visual explanations
Writing, publishing, and marketing	35	TPT Shop Description Generator, AI 같지 않은 글쓰기 작가	Blog writing, course promotion, worksheet sales
Tests, assessment, and item generation	31	수능 완전체, 문항 분석가, 최종 수능 메타프롬프트	CSAT explanations, item analysis, audio lessons
Apps, code, and automation	29	React 증상 분석가, RAG 문서처리, AppFlow Automation	Developer posts, automation, tool building
Creative writing, games, and stories	26	Moonlit Quest Weaver, 세계 시뮬레이션, 게임 창조자	Learning quests and gamified education
Thinking tools, analysis, and research	21	MECE 분석 메타프롬프트, 좋은 질문 생성기	Research, structure, problem framing
Audio, subtitles, and voice	10	11 labs 강의 스크립트 변환기, SRT 생성	TTS, radio shows, subtitles
Knowledge management, search, and RAG	7	메타이볼브 제텔카스텐, 온톨로지 AI 교육 도구 생성	Obsidian, LLM Wiki, retrieval
Uncategorized restoration queue	87	Claude Archives 복원 대기 노드	Source tracing and later reclassification

The largest cluster is Prompts, harnesses, and agents. But the most immediately publishable clusters are English education and original-book learning, Tests, assessment, and item generation, and Writing, publishing, and marketing, because they already connect to ReadMaster, CSAT explanation posts, and the blog-writing loop.

The connection rule is:

source file
-> category
-> representative node
-> applied skill
-> public blog article
-> next work candidate

For example, 정량적 영어 원서 워크시트 생성기 belongs to the English-learning category, connects to /k-kid-book, and can feed ReadMaster original-book lessons and worksheet sales. React 증상 분석가 belongs to the code-and-automation category and can become a developer article or a debugging harness note.

First Publishing Candidates

The first sequence should be:

1. Claude Project DNA Registry

This is the top-level map for the archive. This article plays that role.

2. Moonlit Quest Weaver

A gamified learning node that turns learning objectives into quests and rewards. It connects naturally to ReadMaster original-book lessons, worksheets, and radio shows.

3. 정량적 영어 원서 워크시트 생성기

A worksheet node that separates original English texts into vocabulary, comprehension, grammar, and summary sections with controllable quantities.

4. SlangSavvy Wordplay Quest

A game-style learning node for practicing slang through context, part of speech, and meaning. It can grow into expression training for college students and adult learners.

5. TPT Shop Description Generator

A marketplace-writing node that turns educational materials into sales-ready product descriptions. It connects worksheet generation with Teachers Pay Teachers workflows.

6. React 증상 분석가

A short but interesting node. It reads React app issues through symptom, diagnosis, and treatment. It could become a developer-facing article.

How To Handle Restoration Nodes

The 471 mostly empty files are not meaningless. Their titles still show what I cared about at the time. But they should not become public posts before the original material is recovered.

There are three recovery paths:

1. Find a same-named .txt archive.
2. Inspect .ajson cache files for path, metadata, and blocks.
3. Recover the source from Claude exports, previous conversations, or Obsidian backups.

The .ajson file is not the original note. But it is a useful clue. Its path points to the likely source note, metadata preserves titles and tags, and blocks shows how the note was segmented.

The Publishing Loop From Here

This archive cannot be cleaned in one pass. The loop must stay small.

1. Pick five candidates.
2. Check whether each has a live body.
3. Convert the raw prompt into a node explanation, not a prompt dump.
4. Preserve purpose, inputs, output contract, and verification criteria in the article.
5. Keep private raw instructions in Obsidian; publish reusable patterns in the blog.
6. Connect related posts through [[wiki links]].

Then the blog stops being a simple log. It becomes a prompt operating system: readers can understand the pattern, and I can reuse the article as a working tool.

Conclusion

Claude_Archives is not yet a clean knowledge system. But it is not a folder to throw away either. It is closer to a mine: most of the material still needs sorting, but a few veins are visible.

The first principle is this:

Do not preserve the prompt. Preserve the problem the prompt was solving.

Raw prompts age quickly. Problems, roles, input conditions, and verification criteria last longer. LLM Wiki is a way to preserve the durable part.

The next article should take one node out of this map and refine it properly. The first candidate is Moonlit Quest Weaver, a node for turning educational content into a game structure that can connect directly to ReadMaster original-book lessons.