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.

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.

Point it at a finished session; it asks what is private before it extracts anything.
> 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.
Mark the part worth showing while you work; everything outside the markers stays out.
> /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?
Nothing about it is specific to one domain: a plain session with no MCP servers works too.
> 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").