Quick Start

This walks you through your first session with Rocky in about 5 minutes — using the recommended Claude Code (or OpenCode) flow.

Haven’t installed yet? Start at Installation — rocky install claude-all wires up everything you need before this page makes sense.

Step 1: Work normally with your AI agent

Open Claude Code in a project, work on a real task. Don’t change anything about your normal flow.

Behind the scenes:

• Each commit silently appends its diff to ./.rocky/queue/. No LLM call on commit — there’s no interruption to your flow.
• Each prompt is appended to ./.rocky/prompts.jsonl. The skills will read this transcript later to ground topic extraction in your actual session.

After a real chunk of work, you’ll have a queue and a transcript ready for extraction.

Step 2: End of session — review

In the same Claude Code session, type:

/rocky-review

This is the recommended default. It’s an umbrella skill that runs two passes back-to-back:

1. /rocky-checkpoint — reads the queued diffs + your prompt transcript, identifies new concepts, and writes them to your PKG with a four-question bank each (generic enough that the same topic resurfacing in a different project still matches).
2. /rocky-promptiq — rescores the prompts you’ve used in this session for prompting quality, with one-line feedback per prompt.

You get one fused two-line summary at the end — what you learned and how you asked.

Underneath, both can be called individually. For CI / automation that runs after every push, prefer the underlying skill directly so failures stay isolated to their own queue:

/rocky-checkpoint    # CI: every push
/rocky-promptiq      # CI: weekly cron (slower cadence)

What you get from the checkpoint pass:

• New nodes for each distinct concept the session introduced.
• Updated nodes when something already in your PKG appeared again — Rocky bumps the encounter count and refreshes its weighting.
• Cross-project dedup — the same idea across two projects becomes one node with multiple repos[] entries, not two duplicates.

No LLM cost on the Rocky side. The agent’s own context window does the extraction.

Step 3: Quiz in Rocky View (recommended)

The terminal works, but the browser is where Rocky actually shines:

rocky view

A local server starts on 127.0.0.1:<random-port> and your browser opens straight into the Dashboard. The first thing you see is your Rocky IQ — a single 0–100 number that summarises how well you’d recall everything in your PKG right now.

Four tabs share the same data: Dashboard, Knowledge Map, Review Queue, Saga. Click Quiz top 5 in the Review Queue to start a focused session on your weakest topics. Answers go in via keyboard or the 🎤 button (voice is local via whisper.cpp; nothing leaves the machine). The IQ banner updates live as your scores come in.

Why the browser instead of the terminal? The graph view turns the PKG from a flat list into something you can actually feel — clusters, gaps, and the shape of what you’ve learned all become visible. The answer experience is also just nicer: real text wrapping, voice input, instant graph updates, and the IQ dial right there.

Dashboard

Rocky IQ banner with the live score and gradient bar, recent activity, and weakest topics at a glance.

Knowledge Map

Interactive 2-hop “wormhole” graph — bounded to ~28 nodes on screen at a time regardless of PKG size. Click any node to warp focus to it. Edges are colour-coded by kind, weakest topics glow for attention.

Review Queue

Flat list of every topic that needs attention, ranked by recall_now ascending. Quiz top 5 is a single click — the most direct path from “open the app” to “start drilling.”

The Saga tab is a cinematic timelapse of your knowledge graph growing over time — gated behind ~30 days of consistent reviewing as a reward, so it doesn’t get a screenshot here.

Step 4: Quick inline review without leaving the agent

If you’d rather not switch contexts, you can quiz directly inside Claude Code:

/rocky-quiz

The agent picks your weakest topics by recall_now, asks one question per topic from the canonical question bank, and writes the score back. Same FSRS state advancement as Rocky View — just text-only and inline.

This is best for quick reviews — a couple of questions between tasks. For a real study session, open Rocky View.

What happens over time

Rocky uses FSRS — the same family of spaced-repetition algorithm Anki moved to. Topics you know well decay slowly. Topics you barely know decay fast. Over time, Rocky surfaces the right things at the right moments without spamming you.

Your Rocky IQ is built from recall_now = retrievability × mastery averaged across your PKG. It moves down when you stop engaging and up when you can answer for what you’ve shipped. There’s no daily cap on /rocky-review or rocky view — you control when you extract and when you review.

rocky "add user login with JWT tokens to my Express API"

rocky diff           # last commit
rocky diff HEAD~3    # specific commit
rocky diff --staged  # before committing

rocky quiz           # full review
rocky quiz "redis"   # search and pick topics