diff --git a/README.md b/README.md
index 940de84..7b640a1 100644
--- a/README.md
+++ b/README.md
@@ -63,6 +63,33 @@ pre-commit installed at .git/hooks/pre-commit
 
 3. You only need to run steps 5, 7, and 8 in the quick start during development. If the test cases need to be updated, step 6 is also needed.
 
+## How does it work?
+
+These steps are executed in runner-images. We use `sudo -u tt` to elevate the permission and run `joj3` and `joint-teapot`. All the secret files should be stored in the host machine with user `tt` and mounted into the runner (e.g. `/home/tt/.config`). Since the runner uses user `student`, we can keep the data safe.
+
+1. Run JOJ3
+   1. Parse the message.
+       - If not specified by `-msg`, it will use the git commit message from `HEAD`. The message should meet the [Conventional Commits specification](https://www.conventionalcommits.org/). We use `scope` and `description` here.
+   2. Find the configuration file.
+       - We have `conf-root` and `conf-name` specified in the CLI argument. Then the full path of configuration file is `<conf-root>/<scope>/<conf-name>`.
+   3. Generate stages.
+      - We have an empty list of stages at the beginning.
+      - We check all the stages from the configuration file. Stages with empty `group` field will always be added. And stages with `group = joj` will be added when `description` contains "joj" (case insensitive).
+      - Every stage needs to have an unique `name`, which means if two stages have the same name, only the first one will be added.
+   4. Run stages.
+      - By default, all the stages will run sequentially.
+      - Each stage contains a executor and a parser. Executor (currently only sandbox) executes the command and parser parses the output generated by the executor.
+      - The parser can return a force quit, which means all the stages after it will be skipped.
+   5. Generate results.
+      - Once the running of stages is done, it will generate a result file where the path is specified in the configuration file.
+2. Run Joint-Teapot.
+   1. Generally speaking, it reads the JOJ3 results file and output results on Gitea.
+      - We use a wrapper script `joj3-teapot` to limit the commands can be run in the original `joint-teapot` for safety.
+      - Currently, the environment file path and joj3 result file path are hardcoded, as `/home/tt/.config/teapot/teapot.env` and `/tmp/joj3_result.json`, respectively.
+   2. With `joint-teapot joj3-scoreboard`, it will update the scoreboard file in grading repo.
+   3. With `joint-teapot joj3-failed-table`, it will update the failed table file in grading repo.
+   4. With `joint-teapot joj3-create-result-issue`, it create an issue in the submitter's repo to show the results.
+
 ## Models
 
 The program parses the configuration file to run multiple stages.