Teaching AI to Take Notes Before It Forgets: MyR2D2, 12 Open-Source Claude Skills

Use AI tools long enough and you end up stuck between two bad options.

Feed it the whole backstory, and you burn through tokens as the context grows, and every reply gets slower. Don't feed it, and it wakes up with amnesia every day. The rules you agreed on, the traps you already fell into, the next step it promised to take: all gone. Tomorrow it leads you straight back into the same mistake.

Claude, ChatGPT and Gemini all offer some kind of memory now, and it helps a little. In my experience, though, a new session still only knows what got written down, and the working details of a task rarely are.

I can't cure the amnesia. So I taught the AI to write itself notes before it forgets. Those notes grew into a set of small tools, and I've open-sourced them as MyR2D2. It's at v0.7.3 now, with 12 skills.

What amnesia actually costs

"AI has no long-term memory" sounds abstract. In practice it looks like this.

You spend an hour with it untangling a problem and agreeing on a plan. The next day you open a new chat and it knows nothing. You explain everything again, it hands you a slightly different version of yesterday's plan, and then it recommends the exact route you ruled out yesterday.

The second failure is sneakier. To avoid the first one, you paste in all the background every time. Conversations get longer and slower, your quota drains, and the three sentences that really matter end up buried under two thousand words of context, where the model can easily overlook them.

So the real question isn't whether to give AI a memory. It's whether there's a third way between burning tokens and losing everything.

My answer: stop asking the AI to remember. Make it write down what matters, on disk, before it forgets.

What's in the box

The 12 skills fall into four groups. The cheat sheet below is the one-page version.

Wrap-up and handoff

• save-all: Run it before you close up or reboot. It finds everything that only lives in the conversation, writes it out, then reads it back to confirm it actually landed. It ends with either "safe to restart" or "X still isn't synced."
• dropoff: When a task needs to move to another chat, another machine, or another project, it writes a handoff card: where things stand, what's next, the links, the traps. The card assumes the reader knows nothing.
• pickup: dropoff's partner. A new session checks for cards addressed to it, reads them in full, claims one, and gets to work. On recent versions of Claude Code, if the receiving session is already open, dropoff can also ring its doorbell.

Work journal

• mission-log: Harvests any day's session activity from the transcripts Claude Code already keeps. It costs zero tokens, because it's a plain script with no model involved.
• daily-debrief and weekly-debrief: Turn that skeleton into a daily report with a reflection, then roll seven of them into a weekly. Transcripts are cleared after 30 days by default, so this is how the value gets out before they go.

Kickoff and quality

• new-mission: For anything with three or more steps, or anything hard to undo. It looks around first, asks at most five questions (with numbered options, so on a phone you mostly reply with a digit), drafts a plan where every step says how it will be checked, and waits for an explicit go. "Looks fine" doesn't count. When the work is done you get a wrap-up report against the plan, including an honest ledger: what wasn't done, what wasn't verified, and what was a guess.
• damage-report: Five questions the AI runs against your original request before it reports back. If there's nothing real to suggest, it's allowed to write "none."
• ai-review: Sends the work to a different model family for a second opinion, by default through Codex CLI on your own ChatGPT account, with separate modes for code, public-facing copy, and research reports. It then folds the feedback into its conclusion: what it accepts, what it rejects and why, and what you need to decide. If Codex isn't installed or signed in, it says "self-review only" and lets you keep going. A tool that blocks you when you're in a hurry is a tool you stop using.
• ai-search: Live web answers with citations you can check. When it can't find something, it says so instead of filling the gap from stale training data.

Budget and life

• token-optimizer: Budget rules the AI follows before it spins up multiple agents: pick the right model tier, compress what comes back, and stop after three failures. On a subscription plan like mine, running out of quota stops every conversation until it resets, which hurts more than a bigger bill.
• flight-to-calendar: This one is purely for me. Through a Google Calendar connector, it puts booked flights on your calendar with the time zones right, one event per leg, and a note on which side of the plane gets the sunset. If you fly often you know why: travel days are when meetings get reshuffled, and I have sprinted through more than one airport because of it.

Two design decisions

Verification over declaration. The most important rule in the whole set: writes get read back, "done" needs evidence, and a success message on its own proves nothing.

I learned that one the hard way. One evening I ran my wrap-up routine, the tool reported success all the way through, and I closed the session. The next day the file was empty, and a whole block of conclusions was gone. Since then, reading back is a hard rule. Nothing counts as saved until I can see the content on disk.

Zero dependencies by default. Handoff cards are plain Markdown files in your project folder. No database, no API, no service running in the background. Tools should still work on your worst day: new laptop, no network, some service down. A Markdown file still opens. You can plug in your own task system later, but the default only needs a project folder you can write to.

I ran the reviewer on itself

The first job I gave ai-review was reviewing ai-review. Three rounds of cross-model review caught 21 defects, and 13 of them had been introduced by my own fixes from the round before. I fixed what round one found, and round two found new problems in the exact places I had touched.

That's why the tool exists. A fix can plant a new bug, and whoever wrote the fix is the last to see it. Ask the same model to "check again" and you mostly get its original conclusion in different words. So, as of v0.7.3, ai-review ships with 41 regression tests you can run yourself, and ai-search ships with 43.

A second opinion is still just an opinion. Some of its feedback catches real problems, some is wrong because it never saw the code, and some covers things I had already weighed and rejected for good reasons. Treating the second opinion as gospel is the same mistake as treating the first one that way.

Why "MyR2D2"

R2-D2 is never the hero, but every episode depends on him. Leia hands him her message, and he carries it across a desert to Obi-Wan and plays it back word for word.

The mapping to the tools turned out tighter than I expected:

1. dropoff is Leia recording "Help me, Obi-Wan Kenobi."
2. pickup is R2 finding Obi-Wan and playing the message.
3. save-all is keeping the plans safe in the escape pod, so whoever comes next can pick up where you left off.

One more detail: the Chinese version of this post reached its writing session through a handoff card. I wrote the brief in the session where I was tidying the repo, then picked it up in a fresh one. For me, the real test of a tool is the day I can't work without it.

The skills are written in Chinese. You can use them in English.

Each skill's instructions are written in Traditional Chinese, and there's exactly one copy of each, so there are no parallel translations to drift apart. Every skill comes with trigger phrases in both languages: "about to reboot," "hand this off to X," and "anything handed off to me?" all work. Claude follows the Chinese instructions and answers in whatever language you use.

Install

MIT license. The recommended route is one line:

npx skills add tingyulu/MyR2D2

That's Vercel's skills installer, which also supports gemini-cli, codex, cursor and other agents. It installs to your project by default; add -g for a global install, or --skill to pick individual skills.

On Claude Code you can also install it as a plugin, which keeps updates in one place and puts the skills under a myr2d2: namespace so they won't collide with skills you already have:

/plugin marketplace add tingyulu/MyR2D2
/plugin install myr2d2@myr2d2

No CLI? The repo's prompts/ folder has paste-in English versions of new-mission, damage-report, ai-review and ai-search for ChatGPT or Claude on the web. The three journal skills need Claude Code, because they read its local transcripts. The README has the full compatibility table.

Repo: https://github.com/tingyulu/MyR2D2

Solve your own pain points, and don't keep the answers to yourself.

If you try it, tell me how it went: which skill you never touched, and where you got stuck. That kind of feedback is worth more than a star.

I'm Eric Lu ("Uncle Eric"), a product consultant, headhunter, and career coach. These skills are my actual daily workflow.