What's your Rocky IQ?

Rocky is an open-source personal knowledge management platform that helps developers fight AI skill atrophy.

Why it exists

Every senior engineer knows the feeling: you stop writing something by hand, and six months later you can’t remember how it works without looking it up. That’s normal. That’s how memory works.

What’s new is the speed. AI tools don’t just accelerate your output — they remove the need to think through problems entirely. You describe what you want, the AI writes it, you ship it. Fast. But the understanding that used to come from doing the work yourself? That doesn’t happen anymore.

This is AI skill atrophy — and it compounds silently. You don’t notice it until the day the AI gives you the wrong answer and you can’t tell.

Rocky exists for one reason: to shine a light on blind spots in systems developers are responsible for — systems co-created with AI agents.

See it in action: watch the demo →

What makes it different

Most learning tools quiz you on definitions. Rocky doesn’t care if you can define JWT. It cares whether you’d catch the bug.

Instead of asking “what is a refresh token?”, Rocky asks:

You’re issuing JWTs with a 15-minute expiry. A user is halfway through a checkout flow when their token expires. What happens — and how do you make the experience seamless without storing session state on the server, question?

That’s the kind of question that separates someone who read the docs from someone who’s actually built with it. Rocky asks the second kind — every time.

How it works

Rocky’s primary mode is alongside your AI agent. When Claude Code is doing the work, Rocky watches what got built, extracts the topics that live in the changes, and turns each one into a question you’ll need to be able to answer for the system you now own.

With your AI agent (Claude Code)

This is the canonical Rocky workflow. After rocky install claude, the agent picks up five skills it can invoke at the right moments — no separate Ollama call, no per-prompt logging in the critical path:

• /rocky-review — the default — umbrella skill that runs /rocky-checkpoint and /rocky-promptiq back-to-back at the end of a session, then prints a single fused summary. Use this unless you need one of the underlying skills on its own.
• /rocky-checkpoint — extract topics from recent commits + the conversation. Called by /rocky-review, but invoke directly from CI / automation so a failure stays isolated to one queue.
• /rocky-quiz — runs a Socratic review session inside the agent session, using the canonical questions Rocky stored.
• /rocky-backfill — seeds the PKG from a project’s existing git history when you’re new to a repo or first installing Rocky.
• /rocky-promptiq — re-evaluates your recent prompts and produces a PromptIQ score with feedback. Also called by /rocky-review; suitable for a slower CI cadence (weekly, etc.).

rocky install claude     # one-time: drops the skills into ~/.claude/skills/
# ... work normally with the agent ...
# inside Claude: /rocky-review  (at end of session — runs checkpoint + promptiq)
#                /rocky-quiz    (any time — drill weakest topics)

The alternative workflows below exist for the moments you’re not in an agent session — useful supplements, not the main story.

rocky "add rate limiting to the API using Redis"

git commit -m "add Redis rate limiter"
# Rocky runs automatically:
# ~ Redis Lua scripting  (recall fading to 71%)
#   Reminder: Lua scripts in Redis run atomically — the whole script or nothing.
#   This is why you use them for rate limiting: checking and incrementing the
#   counter must be a single operation, or two requests can both pass the check
#   before either increments.

rocky quiz               # general review
rocky quiz "redis"       # targeted — search and pick topics to drill

rocky diff --staged

The knowledge model

Rocky tracks a Personal Knowledge Graph (PKG) — a local database of every topic you’ve encountered. Each topic has a recall score: a number between 0 and 1 that combines two things:

• Retrievability (R) — freshness from spaced-repetition decay. High right after a review, lower as time passes.
• Mastery (M) — the mean of your last three review scores (default 0.5 if you’ve never been quizzed). Captures how well you actually answered, not just how recently.

recall = R × M. Freshness alone doesn’t count as knowing — if you got the wrong idea last time, retrieving it quickly today doesn’t help you. Both factors have to be high before Rocky considers a topic known.

Knowledge decays. A topic you understood deeply three months ago might be at 25% today — Rocky surfaces it before you trip over it.

Want to see the whole loop end-to-end? Watch the demo →

Rocky the alien

Rocky has a personality based on Rocky the alien from Andy Weir’s Project Hail Mary — enthusiastic, direct, and genuinely rooting for you.

• Fist my bump, friend! Is correct!
• Excite excite excite! Friend get it!
• Is okay! Rocky also not know at first!
• We are crew. We solve together.

Questions end with ”, question?” — Rocky’s way of asking. Set personality = false in [ui] config for plain output.

Key concepts