Meeting 11

Author

Kwok-leong Tang

Published

April 15, 2026

Modified

April 15, 2026

Today’s Schedule

Part 1 — Sharing Your LLM-Wikis

  • Each student presents their LLM-wiki (Assignment 3 progress)
  • What worked, what didn’t, what surprised you
  • Common problems and solutions
  • Discussion: what makes a good schema?

Part 2 — Model and App Independence: Why It Matters

  • Why the LLM-wiki pattern frees you from vendor lock-in
  • Your knowledge lives in markdown, not in any app
  • Hands-on: Set up Opencode with the Z.ai Coding Plan
  • Run the same wiki tasks with a different agent and a different model

Part 1: Sharing Your LLM-Wikis

Last week you began building research wikis using the LLM Wiki pattern. Today we start by hearing from everyone: what did you build, and what went wrong?

Presentation Format

Each student will share their screen and walk through their wiki for 5 minutes:

  1. Show your vault structure — how did you organize raw/, your wiki pages, and your schema?
  2. Pick one ingest — show us a source you ingested and the wiki pages it produced. Were the cross-references correct? Did the agent follow your templates?
  3. Name one problem — what was the hardest part? A bad template? A hallucinated fact? A broken wikilink? An agent that ignored your AGENTS.md?

Discussion Points

As you listen to each other’s presentations, think about:

  • Schema design: Did students who wrote more detailed AGENTS.md / CLAUDE.md files get better results?
  • Template adherence: When did the agent follow the template faithfully, and when did it improvise? Was the improvisation good or bad?
  • Cross-referencing: Did the agent create wikilinks to pages that don’t exist yet? Is that a bug or a feature?
  • Source quality: Did the agent handle Chinese-language sources differently from English ones?
  • Correction workflow: When the agent got something wrong, how did you fix it? Did you edit the wiki page manually, or did you prompt the agent to correct itself?
Tip

Forward-linking is a feature, not a bug. If the agent creates a wikilink like [[Su Shi]] before a Su Shi page exists, that is correct behavior — it is marking a gap in the wiki. When you later ingest a source about Su Shi, the link will resolve automatically. This is how wikis grow.

Note

Assignment 3 is due next week (April 22). Use today’s feedback to improve your wiki before submission.

Part 2: Model and App Independence

The Key Insight

Look at what you built last week. Your LLM-wiki is:

  • A folder of markdown files with YAML frontmatter
  • An AGENTS.md (or CLAUDE.md) file with your schema
  • A raw/ directory with your sources
  • Wikilinks connecting everything together

None of this is proprietary. There is no database, no special file format, no app-specific encoding. Everything is plain text.

This means your wiki is not locked into any specific:

Layer You can switch freely
AI model Claude, GPT, GLM, Gemini, Llama, Qwen — any model can read and write markdown
AI coding agent Claude Code, Codex, Opencode — they all read AGENTS.md and write files
Note-taking app Obsidian, VS Code, any text editor — they all open .md files
Operating system macOS, Linux, Windows — plain text works everywhere

This is the real power of the LLM-wiki pattern: your knowledge is yours. If Obsidian disappears tomorrow, your wiki still works. If you switch from Claude to GPT, your wiki still works. If you move from Codex to Opencode, your AGENTS.md still works.

Why Markdown?

Markdown is the lingua franca of AI coding agents. Every major agent — Claude Code, Codex, Opencode — reads and writes markdown natively. When you write your schema in AGENTS.md, you are writing in the format that every agent already understands.

Compare this to traditional research tools:

Tool Format What happens if the tool disappears?
Evernote Proprietary database Export to HTML, lose structure
Notion Proprietary JSON Export to markdown, lose databases and views
OneNote Proprietary format Limited export options
Google Docs Cloud-only Download as .docx, lose comments and suggestions
Your LLM-wiki Plain markdown + YAML Nothing changes — files are already portable
Important

This is not an argument against Obsidian. Obsidian is excellent precisely because it uses plain markdown. Your vault is just a folder of .md files. Obsidian adds a beautiful UI on top, but the data is never locked in.

Proving It: Same Wiki, Different Agent, Different Model

To demonstrate model and app independence, we will now take your existing wiki — the one you built with Codex or Claude Code — and run the same tasks using Opencode with a completely different model (GLM from Z.ai).

If the pattern works, the results should be comparable. The agent and the model are different, but the wiki, the schema, and the instructions are the same.

Hands-On: Set Up Opencode with the Z.ai Coding Plan

In Meeting 09, we introduced Opencode and configured it manually with opencode.json. Today we will use a simpler method: Z.ai’s built-in authentication for Opencode.

Step 1: Install Opencode (if you haven’t already)

macOS (Homebrew — recommended):

brew install anomalyco/tap/opencode

Any platform (npm):

npm i -g opencode-ai@latest

Any platform (curl):

curl -fsSL https://opencode.ai/install | bash
Note

If you already installed Opencode in Meeting 09, update it to the latest version:

brew upgrade opencode    # Homebrew
npm update -g opencode-ai  # npm

Step 2: Log In with Your Z.ai Coding Plan

Z.ai has native integration with Opencode. Instead of manually editing config files, you can authenticate directly:

opencode auth login

When prompted, select Z.AI Coding Plan as the provider and enter your Z.ai API key.

Tip

Where is my API key? Go to z.ai/manage-apikey/apikey-list, sign in with your Z.ai account, and copy your key. You need an active GLM Coding Plan subscription — your instructor has already set this up for the class.

Note

What is the Z.ai Coding Plan? It is a subscription from Z.ai (the international platform of Zhipu AI) designed specifically for AI coding tools. It gives you access to GLM models through an OpenAI-compatible API, so it works with Opencode, Codex, and other agents. See the official documentation for details.

Step 3: Select a Model

After logging in, launch Opencode in your vault directory:

cd ~/your-humanities-vault
opencode

Use the /models command inside the Opencode TUI to select a model. The recommended model is:

Model Description
glm-4.7 Default model for the Z.ai Coding Plan. Reliable and cost-effective.
Note

The Z.ai Coding Plan default model has been upgraded to GLM-4.7. This is the model you should use for today’s exercises.

Step 4: Verify It Works

In the Opencode TUI, type a simple prompt to verify the connection:

What files are in this directory? List them.

If Opencode responds with a file listing that includes your AGENTS.md, templates, and wiki pages — you are connected and ready.

Alternative: Manual Configuration (from Meeting 09)

If opencode auth login does not work, you can configure Opencode manually by creating an opencode.json in your project directory. This is the same method we used in Meeting 09:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "z-ai": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Z.ai GLM Coding",
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4",
        "apiKey": "{env:ZAI_API_KEY}"
      },
      "models": {
        "glm-4.7": {
          "name": "GLM-4.7",
          "limit": { "context": 200000, "output": 65536 }
        }
      }
    }
  },
  "model": "z-ai/glm-4.7"
}

Then set the environment variable:

export ZAI_API_KEY="your-zai-key-here"

See Meeting 09 for the full manual configuration walkthrough.

Exercise: Ingest a Source with Opencode + GLM

Now that Opencode is connected to Z.ai, let’s prove that your wiki is truly model-independent. You will run the same ingest operation you did in Meeting 10 — but with a different agent (Opencode) and a different model (GLM-4.7).

Step 1: Open Your Vault in Opencode

cd ~/your-humanities-vault
opencode

Make sure Opencode detects your AGENTS.md file. You can verify by asking:

What does my AGENTS.md say about how to create Person entries?

The agent should summarize your schema rules — proving it read the same instructions that Codex or Claude Code used.

Step 2: Ingest a New Source

Choose a source you have not yet ingested — either from your raw/ directory or a new one. Give Opencode the same kind of prompt you used before:

I have a new source in raw/my-new-source.md. Please ingest it into the vault:

1. Read the source text
2. Create or update wiki entries using the appropriate templates
3. Add cross-references with wikilinks
4. Tell me what you created and what connections you found

Step 3: Compare the Results

After Opencode finishes, open the wiki pages it created in Obsidian and evaluate:

Question What to look for
Did the agent follow your templates? Check YAML frontmatter fields against your schema
Are wikilinks correct? Click them in Obsidian — do they resolve?
Is the content accurate? Compare against the raw source — any hallucinations?
How does it compare to Codex/Claude Code? Is the output better, worse, or just different?
Tip

Different models have different strengths. GLM models may handle Chinese-language sources differently from Claude or GPT. Try ingesting a Chinese primary source and compare the quality across models — this is valuable data for your research workflow.

What You Just Proved

By running the same wiki operation with a different agent and model, you demonstrated that:

  1. Your wiki is not locked into any tool. The same AGENTS.md, the same templates, the same folder structure — it all works across agents.
  2. Your schema is the real product. The schema you wrote is more valuable than the tool you used to run it. A good schema produces good results regardless of the model.
  3. You can switch freely. If one model handles Chinese text better, use that for Chinese sources. If another is better at structured data, use that for bibliography entries. Mix and match.

This is the practical meaning of model independence: your research infrastructure is not a bet on any one company.

Takeaways

  1. Share early, share often — seeing how others structured their wikis reveals patterns and pitfalls you would not discover alone
  2. The LLM-wiki is model-independent — your knowledge lives in markdown, not in any model’s memory or any app’s database
  3. The LLM-wiki is app-independent — Codex, Opencode, Claude Code all read AGENTS.md and write the same markdown files
  4. Your schema is the most portable artifact — a well-written AGENTS.md works with any agent and any model
  5. Different models, different strengths — use model independence to your advantage by matching models to tasks

Resources