The problem I kept running into
Using Claude Code in the past couple of months I have noticed a clear bottleneck in my workflow - reviewing code written by AI.
My workflow looked like this:
- Finish a Claude Code session
- Open a draft PR to review the diff
- Hit a question on a specific line
- Copy file name + line numbers back into Claude
- Try to help the session to remember the context
- Repeat
Then I realized something: in the pre-AI world, I never did this.
I never took diffs back to the code author. I just left a comment
in the PR and the author answered directly, because they had full
context of why and what.
So I built askdiff.
What askdiff does
Askdiff is a Claude Code skill + an NPM package that opens a GitHub PR-style diff viewer in your browser, attached to the same Claude Code session that wrote the code.
Launch /askdiff from any Claude Code session and it will open the working directory diff in a UI attached to the Claude Code session.
How to try it
cd /path/to/your/project
npx -y askdiff install-skill
# Or, install user-level (available from any project):
npx -y askdiff install-skill --global
Then run /askdiff from any Claude Code session.
/askdiff with natural language descriptions for diff and session
Askdiff can also be launched by describing the diff and the session you want it to be attached. Let's say you have made a change days ago and it had a dedicated session that has all the context. You can do
/askdiff the commit where we implemented auth attached to the session that wrote it
The technical bit
Askdiff uses claude --resume without --fork-session, so, if launched or attached to a session that has the context your question contains all the relevant information that the AI needs to answer it.
This means if you ask "why did you use a Set here instead of an
array?", the answer will reference the actual reasoning from when
the code was written, not a generic explanation.
What's next
Currently Claude Code only, no Anthropic API key required, no configuration.
The repo is open source: https://github.com/narghev/askdiff
I'm still using askdiff daily and refining the UI and the skill itself.
If you're using Claude Code in your team workflow — how are you reviewing AI-generated code today? Is the back-and-forth between PR and Claude a pain for you too?