Lookback — Development skill for Claude Code

Lookback

**Your Claude Code sessions already know what your configuration is missing.**

Every session leaves a transcript, and those transcripts record which skill, which subagent and which plugin drove each tool call, what each turn cost, and which calls failed. Nobody reads them.

Lookback reads them, compares what you actually do against what you have actually configured, and hands you the change to make: the permission rule to add, the instruction you have retyped in nine sessions that belongs in `CLAUDE.md`, the skill nobody invokes, the devloop a repository never wrote down.

**Everything stays on your machine.** There are no network calls in collection or in reporting — including the plugin suggestions, which match against the catalogue Claude Code has already cached. The single exception is opt-in and named: `lookback bootstrap --org ` reads your organization's repository listing through your own `gh` CLI, and sends nothing from the local store.

The one exception is the update check, and it is worth being exact about: at most one request a day, for a release tag, from a cache beside the store. It carries no query, no body and no identifier, and nothing about your corpus is in it. `LOOKBACK_NO_UPDATE_CHECK=1` turns it off. Collection and reporting still make no network call at all.

**Lookback proposes; you apply.** Nothing it reports edits your `settings.json`, `CLAUDE.md` or `AGENTS.md` — not the CLI, and not the agent reading its findings. You are handed the exact text to paste. The two exceptions are commands you run on purpose: `lookback install --write`, which adds the plugin to a repository, and `lookback bootstrap`, which writes a new repository's scaffolding after asking.

---

Install

There are two levels, and **the first one is all most people need.**

1. Your account — the binary and the plugin

curl -fsSL https://raw.githubusercontent.com/LLM-Lineage/lookback/main/install.sh | sh

This is what "install Lookback" means