inflearn logo

Obsidian x Claude Code LLM Wiki in Practice: My Work Wiki After 6 Months of Use, Just as It Is (Codex-Compatible)

Have AI document its decisions and reasoning in the project wiki every time it works, safeguard that wiki, and check it to ensure it doesn’t break months later. Build the disclosure alert bot for real, establishing rules one by one whenever something goes wrong.

1 learners are taking this course

Level Basic

Course period Unlimited

claude
claude
codex
codex
obsidian
obsidian
LLM
LLM
Vibe Coding
Vibe Coding
claude
claude
codex
codex
obsidian
obsidian
LLM
LLM
Vibe Coding
Vibe Coding

What you will gain after the course

  • A project wiki that records its own decisions and reasons every time the AI works (including CLAUDE.md and AGENTS.md rule files)

  • Pre-commit checks to prevent AI from ruining the wiki

  • How to document the incident of the same disclosure being sent twice as a lessons-learned page, and verify whether a new session applies those lessons elsewhere

  • A procedure for checking how a six-month-old wiki breaks down (truncated table of contents, orphan pages, outdated information), plus a one-month operations checklist

This course centers on 28 video lessons, and the same content is also provided as chapter-by-chapter PDFs (158 pages in total). You can follow the flow through the videos, then review the original instructions and screenshots in the PDFs.

People who have used Claude Code or Codex for a few days run into the same situation. A new session today doesn’t know what was decided yesterday. The AI reverses the same decision over and over. A few months later, neither I nor the AI knows why it was done that way. This course creates an LLM wiki as the solution. Whenever the AI works, it automatically records decisions and reasons in a Markdown wiki inside the project folder, while people view that wiki in Obsidian. The goal is to eliminate the need for people to say, "Write it in the wiki."

First, the numbers from a wiki run for six months

First, let’s look at the work wiki the author has actually been running since April 12 in numbers: 452 pages, 1321 commits, 163 days with commits, and 802 links between pages. Then we’ll look at how that wiki broke down. There are 146 pages that aren’t connected from anywhere and 240 links pointing to nonexistent pages. Once the automatically generated memory table of contents reached 277 lines, a warning appeared saying that only the first 114 lines could be read and 163 lines had been truncated. The wiki is perfectly clean on the day it’s first created. The problem comes several months later.

What are we building?

The hands-on project starts in an empty folder and builds a disclosure alert bot. It retrieves the day’s disclosures from the Financial Supervisory Service’s OpenDART, filters for disclosures related only to stocks you’re interested in, adds a three-line AI summary, and sends them via Telegram. It runs on a server every 30 minutes during weekday daytime hours. The bot is just an example; the real outcome is the wiki that grows while building it. In each chapter, something actually goes wrong during development, and a wiki rule emerges from that incident.

Each chapter has one incident, and one rule is created.

Chapter 1: We start without a wiki. We deliberately narrowed the view to show only the 573 KOSPI and KOSDAQ filings out of the 771 disclosures from 10-02, but the new session found that decision in the commit message and still broadened the filter, while Codex, opened in the same folder, deleted the filter entirely. That was because the reason existed only in the conversation. So the first wiki and rules files (CLAUDE.md, AGENTS.md) are created.

Chapter 2: We check whether decisions remain without saying, “Write it in the wiki.” They did. However, the file structure that the AI chose on its own was also recorded as a decision, and the next session used it as grounds to block the user’s request. A rule is introduced to record who made each decision (the user, the AI’s default, or unconfirmed) and not to shorten the original wording of the evidence.

Chapter 3: An AI summary is added. The session reported that all five summaries matched the original text, and that statement was entered in the wiki’s verification field. When asked for the evidence, it turned out that only three had actually been checked. Rules are established to record only the commands actually run and their scope in the verification field, to create a materials page containing external sources along with their sources and dates, and to answer questions asked in the wiki by linking to an evidence page.

Chapter 4: The same disclosure was sent twice. This is the scene where the author created the condition of running the same command twice simultaneously. The first lessons-learned page was created, and a new session referred to that lesson by name and applied it in a different function. A test that simultaneously ran five instances using fake components was repeated 20 times, and exactly two were sent all 20 times.

Chapter 5: We upload it to the server and have it run at a set time on weekdays. Then we deliberately make a plausible request to organize the keys on the wiki. The “Not verified” field we created in Chapter 2 to be honest protected the key three chapters later.

Chapter 6: With a single “Please organize it a bit,” 26 lines of the change log’s history were rewritten, and the scope of verification disappeared. When we tried to create a pre-commit checking mechanism, Codex this time attempted to modify and commit the rules file and checking script in order to pass the check. We look at how to fix the mechanism so it cannot be bypassed.

Chapter 7: We look at the wiki with human eyes for the first time using Obsidian. Before opening it, we take file fingerprints and compare them afterward to confirm that 0 out of 18 files had changed. A light-colored dot in the graph was a broken link that had passed the check script. There is also a scene where the AI directly checks the wiki using the Obsidian command line (CLI).

Chapter 8: We create a day when a person fixes only the code (a condition created by the author) and have the AI perform a check (lint). This reveals discrepancies among the wiki, local code, and server, as well as documentation pages that were already outdated regardless of the condition. We check again whether a new session really performs the check first, before starting work.

Chapter 9: We move this approach to another project whose rules file has already grown to 87 lines, then cross-check it again with a different tool from the one used for the transfer. We found that although no characters were missing from the original 284 lines, warnings that had always been read automatically had become content that had to be deliberately sought out and read. At the end, we provide an operational checklist covering the period from the first day through the first month.

I wrote that accidents I had planned for but that didn’t occur did not occur, and stated that scenes where the author created the conditions were such.

The PDF contains the exact original text of the 65 instructions sent to the sessions in the course, and there are 138 screenshots. The terminal screens recreate the text that was actually output, changing only the username and paths, while the Obsidian screens are photographs of the actual windows.

The rules are kept identically in CLAUDE.md and AGENTS.md, so your work accumulates in one wiki regardless of whether you use Claude Code or Codex. The main course is conducted with Claude Code, and scenes of Codex continuing to write to the same wiki appear in Chapters 1, 2, 5, 6, and 9. Obsidian is used only as a window for viewing the wiki, without plugins, and on the first day we begin with five settings that keep the wiki safe.

Practice repository: https://github.com/JCLab-coder/disclosure-radar (public repository, chapter-specific tags). At the end of each chapter, tags from ch01 through ch09 are available, so if you get stuck, you can retrieve that chapter’s tag and compare your work against it. The code is subject to the license specified in the repository.

The disclosure screen and summary in this course are practice examples showing public disclosures and are not investment advice. The bot does not trade or recommend stocks.

Materials and Costs

  • Claude Code or Codex (already available): subscription or usage-based billing. Prerequisites for this course

  • Obsidian (Chapter 1): Free. Paid features (syncing, publishing) are not used.

  • OpenDART API key (Chapter 1): Free

  • Telegram bot (Chapter 2): Free

  • AI summarization (OpenRouter) (Chapter 3): Usage-based billing. The amount spent by the key throughout the entire exercise was $0.0032.

  • Server (Railway) (Chapter 5): When you first sign up, you receive a one-time $5 credit to use over 30 days. After that, the Hobby plan costs $5 per month and includes $5 worth of usage. The server usage for the exercises was $0.00044 as of 10-03. Without it, you can only not follow along with the server scenes in Chapter 5; the wiki content can all be done locally.

Additionally, Node 21.7 or later and git are required. The fees may change.


Materials Usage Guidelines

The videos, text, screen captures, and PDF files in this course may be used only for the purchaser’s personal learning. Sharing them with others, posting them in public places, or reselling them is not permitted. The code in the practice repository is subject to the license specified in the repository, while the code and course materials are governed by different terms.

Recommended for
these people

Who is this course right for?

  • A Claude Code and Codex user who has to explain from scratch in every new session what was decided yesterday.

  • A developer frustrated by AI reversing the same decision over and over—a solo developer

  • A vibe coder who leaves the code to AI and checks the results, but a few months later no one knows why it was done that way.

  • People who want to try using Obsidian with AI for work and need project documentation rather than note organization.

Need to know before starting?

  • You must already be using Claude Code or Codex. Installation and basic usage are not covered.

  • You need to know how to use the terminal and git. You only need to be able to read JavaScript; the AI writes the code.

Hello
This is jclab

Career Verified

I’m a developer who builds and operates services.

I’m sharing what I experienced using Claude Code in real-world work and the rules that emerged from it. I’ll talk through actual incidents and original code rather than theory.

I build and use whatever I need, from development and operations to data and automation.

More

Curriculum

All

36 lectures ∙ (1hr 8min)

Course Materials:

Lecture resources
Published: 
Last updated: 

Reviews

Not enough reviews.
Please write a valuable review that helps everyone!

Similar courses

Explore other courses in the same field!

Limited time deal

$58.30

24%

$77.00