Claude session replays
Real Claude Code sessions, replayed with commentary. Scroll through each one, or press play, and open any tool call to see exactly what went in and what came back.
- Z → ee in one prompt on the ATLAS AF One prompt to Claude Code (Sonnet). Under four minutes later there is a tutorial notebook on the UChicago AF that finds an ATLAS Run 3 sample, streams it from the facility cache and fits the Z peak. 64 events
- TTree → RNTuple on the ATLAS AF A real Claude Code session that converts 32 GB of analysis ntuples on the UChicago Analysis Facility, entirely through the atlas-af MCP server. 203 events
Want to make your own replays?
The session-replay skill turns any Claude Code session into a page like these: it captures the transcript, helps you scope and redact it, drafts the commentary with you, and builds the page. In Claude Code:
/plugin marketplace add kratsg/claude-session-replays
/plugin install session-replay@claude-session-replays
Or with the GitHub CLI (any agent it supports; without the plugin, type session-replay:start and session-replay:stop instead of the /replay-start and /replay-stop commands):
gh skills install kratsg/claude-session-replays session-replay --agent claude-code --scope user
Then fork the repository, open Claude Code in your fork, and ask something like "make a replay of my last session in ~/work/my-project". The skill stops for your review before anything is published. See the README for the details.
What a run looks like
Three runs from the skill's own tests, one per way in, condensed: the prompts are real, the output is trimmed.
> Make a replay of the Z→ee notebook session I ran in ~/mcp-demo/zee yesterday, like the other replays in this repo. ⏺ Skill(session-replay) ⏺ Bash(find_session.py --cwd ~/mcp-demo/zee) ⎿ 7365cc3f 2026-11-15 21:21 claude -p · sonnet · 1 prompt ⏺ Before I extract anything: what is private, the whole session or a range, and which timezone should the page use? > Mask my username. Knowledge-base results are internal, omit them. Whole thing, Chicago time. ⏺ Bash(extract_session.py session.jsonl --out events.json --scope scope.json --redact local-redact.json --listing listing.txt) ⎿ 64 events · 1 result omitted · hub URL → example.org/user/jovyan/ ⏺ A docs page that Claude fetched names another user's account (event 27), so I masked that too. ⏺ Bash(leak_scan.py events.json --local local-redact.json) ⎿ 0 hit(s) in 1 file(s) ⏺ Bash(new_story.py --slug atlas-af-zee-notebook --timezone America/Chicago …) ⏺ Write(stories/2026-11-15-atlas-af-zee-notebook/story.json) ⎿ 6 chapters · 23 cards ⏺ Bash(check_story.py … && pixi run lint && pixi run build) ⏺ Ready at localhost:8000. I left out the mass plot because it carries an "ATLAS Simulation Internal" label. Nothing is committed until you say so.
> /replay-start ⏺ Recording (session-replay:start). > the inventory-api healthcheck is failing in staging, can you find out why? ⏺ Bash(kubectl get pods -n staging) … six tool calls, one fix, the rollout comes back healthy … > /replay-stop > Turn that part into a replay. ⏺ Skill(session-replay) ⏺ Bash(extract_session.py session.jsonl --out events.json --listing listing.txt) ⎿ markers: rows 3..18 · 11 events (1 prompt, 4 texts, 6 tool calls) ⏺ Everything before /replay-start is dropped, including the prompt where you loaded an env file with a deploy key. ⏺ What else here is private? > My username, my colleague's name, anything on our corp domains, and anything that looks like a credential. ⏺ Bash(leak_scan.py events.json --local local-redact.json) ⎿ 0 hit(s) in 1 file(s) ⏺ Masked: your username, your colleague's name and email ("a colleague"), the corp domains, the pod IPs, a database password and a GitHub token. Here is the listing. Shall I write the commentary?
> Turn my wordfreq coding session into a replay page so I can show students how Claude Code works through a small task with tests. ⏺ Skill(session-replay) ⏺ Bash(find_session.py --cwd /tmp/wordfreq-demo) ⏺ What is private, and should I use the whole session? > Just my username. Whole thing. ⏺ Bash(extract_session.py session.jsonl --out events.json --listing listing.txt) ⎿ 15 events (1 prompt, 5 texts, 9 tool calls) ⏺ I kept the file names in the temp paths, so the pytest traceback still says which test file failed. ⏺ Write(stories/2026-11-15-wordfreq-tdd/story.json) ⎿ 5 chapters · 11 cards: red then green, why silent output isn't a pass, exit status 127, deterministic tie-breaks ⏺ Bash(check_story.py … && pixi run lint && pixi run build) ⏺ Ready for your review. No MCP servers here, so the page shows no polling or service tiles, only the run's own numbers ("10 passed in 0.34 s").