diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index b009f4b0..d0157131 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -4,6 +4,8 @@ on:
   push:
     branches:
       - main
+    tags:        
+      - v**
   pull_request:
     branches:
       - main
diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 3550cb4a..833bbb64 100644
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -5,6 +5,9 @@ Changelog
 Not released
 ------------
 
+v0.4.0 (2022/05/31)
+-------------------
+
 * SASCA: support modular ADD and MUL operations (#18)
 * TTest: Performance improvement by using a mix of 2 passes and 1 pass algorithms 
 * MTTest: First implementation of multivariate T-test.
@@ -15,8 +18,8 @@ Not released
 * Revamp multi-threading handling thanks to new `scalib.threading` module.
 * AVX2: Wheels on PyPi are built with AVX2 feature. 
 
-v0.3.4
-------
+v0.3.4 (2021/12/27)
+-------------------
 
 * Release GC in SASCA's `run_bp` .
 * Release GC in `rank_accurary` and `rank_nbin`.
@@ -24,46 +27,46 @@ v0.3.4
 * Hide by default the progress bar of `SASCAGraph.run_bp` (can be re-enable
   with the `progress` parameter).
 
-v0.3.3
-------
+v0.3.3 (2021/07/13)
+-------------------
 
 * Solving minor issues in `MultiLDA` and `LDAClassifier`. Allowing multiple
   threads in `predict_proba()` and add a `done` flag to `solve()`.
 
-v0.3.2
-------
+v0.3.2 (2021/07/12)
+-------------------
 
 * Chunk `SNR.fit_u` to maintain similar performances with long traces and
   adding a progress bar 
 
-v0.3.1
-------
+v0.3.1 (2021/06/03)
+-------------------
 
 * Add `max_nb_bin` parameter to `postprocessing.rank_accuracy` (that was
   previously hard-coded).
 
-v0.3.0
-------
+v0.3.0 (2021/06/01)
+-------------------
 
 * Rename `num_threads` parameter of `modeling.MultiLDA` to `num_cpus`.
 * Fix rank estimation when there is only one key chunk.
 * Improve performance of `SNR.get_snr`.
 
-v0.2.0
-------
+v0.2.0 (2021/05/20)
+-------------------
 
 * Remove OpenBLAS and LAPACK, use Spectra and nalgebra instead.
 * Use BLIS for matrix multiplications (Linux-only for now).
 * Make `modeling.LDAClassifier` incremental (breaking change).
 * Add `modeling.MultiLDA`.
 
-v0.1.1
-------
+v0.1.1 (2021/04/26)
+-------------------
 
 * Fix "invalid instruction" bug for CI wheel on windows.
 
-v0.1.0
-------
+v0.1.0 (2021/04/16)
+-------------------
 
 * Initial release, with the following features:
   * LDA and Gaussian templates modelling
diff --git a/DEVELOP.rst b/DEVELOP.rst
index 8a30194c..b02affc5 100644
--- a/DEVELOP.rst
+++ b/DEVELOP.rst
@@ -140,3 +140,13 @@ Also, when you develop, looking at generated assembly may help
     RUSTFLAGS="-C target-feature=+avx2" cargo asm scalib::module::function --rust
     # or
     RUSTFLAGS="-C target-cpu=native" cargo asm scalib::module::function --rust
+
+Make a release
+--------------
+
+0. Start from main branch
+1. Add any missing element in CHANGELOG.rst
+2. Add the new release with the release date in CHANGELOG.rst
+3. Commit, create pull request and merge it (after CI succeeds).
+4. Create and push release tag: `git tag vX.Y.Z && git push origin vX.Y.Z`
+