# Session Review > Wraps up a session with two outputs — a sweep into the auto-memory system to preserve reusable insights for future Claude sessions, and a human-readable recap saved to the Obsidian notes vault. Use for retrospectives, debriefs, post-mortems, or end-of-session reflection. - **Type:** Skill - **Install:** `agentstack add skill-philoserf-claude-code-config-session-review` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [philoserf](https://agentstack.voostack.com/s/philoserf) - **Installs:** 0 - **Category:** [Productivity](https://agentstack.voostack.com/c/productivity) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [philoserf](https://github.com/philoserf) - **Source:** https://github.com/philoserf/claude-code-config/tree/main/skills/session-review ## Install ```sh agentstack add skill-philoserf-claude-code-config-session-review ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Session Review End-of-session wrap-up that produces two distinct outputs: 1. **Memory sweep** — write reusable session insights into the auto-memory system (`~/.claude/projects//memory/`) so future Claude sessions in this project carry the context forward. 2. **Human recap** — a readable summary saved to the Obsidian `notes` vault for the user to skim later. These outputs serve different audiences: auto-memory is for future Claude; the Obsidian recap is for the user. ## Autonomy Exercise judgment during analysis and drafting. Once the recap is drafted and shown to the user, proceed with all writes (memory files and Obsidian save) without requesting further confirmation. ## When to Use - After significant debugging, problem-solving, or design work - When user preferences or project facts emerged that weren't already captured - At the natural end of a working session worth reflecting on ## Process ### 1. Walk the session Reread all human and assistant turns in the current session, including tool call results, from the first message to the most recent. Use the 5-dimension framework in [analysis-dimensions.md](references/analysis-dimensions.md) to structure the walkthrough, watching for: - The user correcting an approach (feedback) - A user preference, role, or workflow detail surfacing (user) - A project goal, deadline, decision, or motivation coming up (project) - An external system, dashboard, or doc being referenced (reference) - A technical insight that was non-obvious at session start and would change how a future session approaches this codebase or problem domain ### 2. Memory sweep Apply the criteria from `~/.claude/CLAUDE.md` ("auto memory" section): - **Keep:** durable facts about the user, persistent feedback rules, ongoing project context, pointers to external systems. - **Skip:** code patterns derivable from current state, git-tracked history, ephemeral task details, anything already covered in CLAUDE.md, **and any item containing credentials, API keys, passwords, or PII** — if such content surfaced, record only the category (e.g., "API key rotation discussed") without the value. **Cap:** at most 5 new memory files per session. If more qualify, prioritize by durability (user preferences > project facts > technical insights) and note the rest in the recap under "Suggested CLAUDE.md updates." Then execute these sub-steps in order: **2a.** Run `mkdir -p ~/.claude/projects//memory/` and read `MEMORY.md` (treat as empty if it does not yet exist). **2b.** For each kept item: - If a matching entry exists in `MEMORY.md`: read the linked file, update changed fields, rewrite the file. - If no match: write a new `.md` with frontmatter per `~/.claude/CLAUDE.md`, then append a one-line index entry to `MEMORY.md`. **2c.** Verify `MEMORY.md` has no duplicate slugs after all writes. If no items qualify, say so in the recap and skip 2a–2c. ### 3. Human recap Draft a narrative recap covering: - **What we worked on** — goal and final state - **Decisions & rationale** — choices made and why - **Surprises & dead ends** — what didn't work or wasn't expected - **Open threads** — anything left unfinished or worth revisiting - **Suggested CLAUDE.md updates** — guidance that belongs in a tracked instructions file rather than auto-memory (if any) **Format:** use the full template by default. Use the compact template only when the session is short (≲15 turns) or fewer than 3 notable events occurred. See [output-templates.md](assets/output-templates.md). Present the recap to the user before the save step. Do not pause for confirmation — see Autonomy above. ### 4. Save to Obsidian Save the recap to the `notes` vault via `obsidian-cli`. Always pass `vault=notes` explicitly — the implicit default vault is unreliable on machines with multiple registered vaults. Before running, verify the composed content: - YAML frontmatter block present at top with `created: YYYY-MM-DD` - No body line begins with `# ` (no H1 headings) - No standalone `---` lines outside the frontmatter fences ```bash obsidian create vault=notes path="Session Reviews/YYYY-MM-DD .md" content="" silent ``` **On failure:** if `obsidian create` exits non-zero or the subsequent `obsidian read` returns no content, surface the error verbatim and emit the full recap inline as a fenced markdown block so the user can save it manually. **Empty session:** if no memory items qualified and fewer than 3 notable events occurred, still save a minimal compact recap noting "No memory updates warranted" so the user has a dated session record. ## Verification - All 5 analysis dimensions were considered (mark "N/A" for any with no findings) - Each memory written cites a specific session moment, not a hypothetical - The Obsidian save succeeded (confirm via `obsidian read vault=notes path="..."`) ## Reference Files - [analysis-dimensions.md](references/analysis-dimensions.md) — 5-dimension framework with questions, formats, and examples - [output-templates.md](assets/output-templates.md) — Full and compact recap formats ## Do not use when - Reviewing code changes for bugs — use `diff-review` - Auditing an existing CLAUDE.md against a template — use `claudemd-audit` - Short, straightforward sessions with no corrections or surprises - The user just wants a quick task done, not a retrospective ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [philoserf](https://github.com/philoserf) - **Source:** [philoserf/claude-code-config](https://github.com/philoserf/claude-code-config) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-philoserf-claude-code-config-session-review - Seller: https://agentstack.voostack.com/s/philoserf - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.