メニュー
Use when reading Claude Code .jsonl session transcripts — covers locating files, filtering rewinds and compactions, focused skimming, and subagent transcripts.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when finishing a feature branch and wanting a final-state cleanup pass. Audits changed comments, structures, and defensive fallbacks with parallel subagents, fixes every validated finding caused or exposed by the change, and leaves unrelated baseline debt untouched.
Use when the user asks for a combined code-review and cleanup final pass before integration, including per-commit review-and-cleanup. Runs both passes to convergence without expanding beyond issues caused or exposed by the original change.
Capture intentional changes made directly to chezmoi targets back into the dotfiles source — modifier templates and directly managed files alike. Use when application-managed or manually edited dotfiles should be synchronized into the dotfiles repository for review without applying changes to the live home directory.
Use when a Pull Request is ready for final pre-squash verification, description and title synchronization, squash-message drafting, and an optional user-authorized squash merge.
Use when reading Codex .jsonl session transcripts — covers locating files (incl. by UUID or project), filtering rewinds and compactions, focused skimming, and subagent transcripts.
Use when an agent needs to call the `claude` CLI from another runtime, delegate work to Claude Code non-interactively, resume a Claude session, capture Claude output, or coordinate cross-runtime agent handoffs.
| name | reading-claude-session |
| description | Use when reading Claude Code .jsonl session transcripts — covers locating files, filtering rewinds and compactions, focused skimming, and subagent transcripts. |
Session files are JSONL. Every line is independent JSON — sed -n Np file | jq expands any line on demand.
Sessions live under ~/.claude/projects/<encoded-cwd>/<session-uuid>.jsonl, where <encoded-cwd> is the absolute path with / replaced by - (e.g. /Users/foo/bar → -Users-foo-bar). Easiest: ls ~/.claude/projects/ | grep <project-name>.
grep -l "<keyword>" ~/.claude/projects/<dir>/*.jsonl. Sort by mtime if the user says "recent".
/rewind branches the parentUuid tree; dead branches stay in the file. To get the live conversation:
last-prompt entry with non-null leafUuid — that's the live leaf.parentUuid back from it through user / assistant / system / attachment nodes.Typically 50–80% of message nodes in a long session are rewound; reading the file linearly without filtering will mislead you.
A system entry with subtype:"compact_boundary" has parentUuid:null (chain intentionally broken) and logicalParentUuid pointing to the pre-compact leaf. The next user entry has isCompactSummary:true — that's the summary the model continued from.
compact_boundary.logicalParentUuid and continue (still apply rewind filter on each segment).user text and assistant text content. These carry intent, decisions, corrections.tool_use.input and tool_result.content. Truncate to ~80 char preview with {line, json-path} ref, expand a specific line only if the surrounding discussion points to it.permission-mode, worktree-state, file-history-snapshot, queue-operation, hook attachments, task-reminder attachments.Plans, file names, and decisions change mid-conversation. Don't trust early statements as still-true — follow the user/assistant exchange forward, and let later turns override earlier ones. When the user asks "what did we decide about X", scan from the end backward.
When a subagent is dispatched, its full transcript is a separate JSONL at:
~/.claude/projects/<encoded-cwd>/<parent-session-uuid>/subagents/agent-<taskid>.jsonl
<taskid> shows up in the parent session as queue-operation content (<task-id>...</task-id>) and as the id returned in the dispatching tool_result. Same format, same filtering rules; subagents can dispatch their own subagents — recurse identically.