A scorecard for you vs. your AI.
Prompt it well, correct it when it's wrong, and find out if it could really replace you.
Works with Claude Code, Claude Cowork, Codex, ChatGPT and Claude.ai.
Tiers to climb, goals to hit, lessons to level up, and a card worth bragging about.
🔒 Private: everything stays on your computer unless you share it
💸 Free: the score never calls an AI, so nothing to pay for
You correct your AI all day: the function that doesn't exist, the instruction it ignored, the bug that's "fixed" but still broken. And sometimes it goes sideways because the ask was vague.
No, Actually... reads the chat history already on your computer and answers one big question, built from two smaller ones:
| The question | What it measures | Tiers | |
|---|---|---|---|
| 🧠 | Could AI replace you? | Your overall score, 0 to 100, from both questions below | Rubber stamp → Coworker → Sharp-eyed → Irreplaceable |
| 🎯 | ↳ Do you make it guess? | How often your prompts are clear the first time: no adding, clarifying or starting over | Constantly → Often → Sometimes → Rarely |
| 🧐 | ↳ Do you correct it? | Who corrects whom: you correcting it, or it correcting you | Rarely → Sometimes → Often → Religiously |
Better prompts mean fewer mistakes to correct, and that never costs you: once you've corrected it a few times (proof you check its work), more corrections add nothing.
💸 Costs nothing to run. No AI does the judging, so there's no API key, no tokens spent and no subscription. (Optional report cards are the one exception: your own AI writes them, so they use some of your tokens.) Fixed, readable rules score your chats right on your computer: the same history always gives the same score, and nothing is uploaded.
- Download it for Mac, Windows or Linux (which file?).
- Open it, click Choose sources, and Allow the AI tools you want counted. Nothing is read until you do.
- That's it. It reads your past chats first, then keeps score as you work, from your menu bar (Mac) or system tray (Windows and Linux). Prefer an ordinary app that quits when you close it? Turn off Keep running in the menu bar in Settings → General.
- A verdict and a score. "Could Claude replace you? No. Score 86, Sharp-eyed." Updated as you work, over the last 28 days.
- A prompt score. "Do you make it guess? Sometimes. 85% of your prompts were clear the first time." Plus what your prompts include, why you followed up, and one tip backed by your own history.
- Improve. One weekly goal picked from your own numbers ("Include guardrails in 25% of your prompts"), streaks, trophies, personal bests, last week's recap, and charts of how you're trending.
- Learn. 13 one-minute lessons in three tracks (the 6 habits, follow up less, correct more), quick quizzes, and a prompt checker that grades a draft as you type, with the same rules as your score.
- English only, for now. Prompts and corrections are read by rules written for English. Messages in other languages aren't graded fairly, and may not count at all.
- Any time window. 7 days, 28 days, 90 days or all time: the whole scorecard follows the picker, score included. Windows under 28 days ask for fewer corrections, so a short window is fair. Windows over 28 days ask for the same 3, so a longer window can score higher.
- Every model, side by side. All AI, one provider, one family, or one exact version (say, Claude Opus 5.5). The picker shows the numbers that page is about.
- A Hall of fame of awards, one per kind of correction: The Fabulist (made it up), Selective Hearing, Reading Is Hard, The Blind Spot, “Fixed” It, Big Brain Energy, and Touché for the times it corrected you. Each shows its best snippet, picked the same way every time. Pin the ones you want to keep forever.
- Popups for your 1st, 10th, 25th… correction, a new tier, a streak, a goal week met, or a goal done.
- Made to brag. Share cards for X, LinkedIn, Bluesky, Threads, Reddit, Facebook, WhatsApp and Instagram. You always see exactly what will be posted before anything goes anywhere.
Finish a task in Claude Code or Codex, type /grade:me (Codex: $grade-me), and get a report card:
- A letter grade, A to F, and why
- What you could have said up front, with a better first message to copy
- What the extra back-and-forth cost you, in minutes (and dollars, at pay-as-you-go prices, when the model has a price). Codex report cards show minutes only, no dollars: there are no GPT prices in the table
It talks at your level: plain words if you're new, short and technical if you're not. No, Actually... saves each report card's numbers (never your words) and shows how you're doing over time.
Turn it on: Settings → Report cards → Turn on. Then start a new chat, finish a task, and type /grade:me.
On Windows: report cards work with Claude Code, which needs Git for Windows. They aren't available for Codex yet.
Without the app: in Claude Code run /plugin marketplace add HTMLGuyLLC/NoActually then
/plugin install grade@noactually. You get report cards, graded by eye, with no history or cost figures. This follows the newest code on the default branch. To stay on a release, add HTMLGuyLLC/NoActually#v<version> instead.
Cost: report cards are written by your own AI, so each one uses some of your tokens: about 0.5k to 2k output tokens per task graded, more with all.
They only run when you ask. Tips as you work are optional and use a little in each chat.
Your score never uses AI and stays free.
Grades can vary: your AI writes the report card, so a grade is a judgment call and can differ between runs, unlike your score.
Your AI chats can hold code, keys, client names and half-formed ideas. This app is built so none of that leaves your computer.
| 🔒 Local only | No accounts, no servers, no telemetry, no analytics. Fonts are bundled. The app's only network use is an update check against GitHub, a share link you click yourself and, with report cards on, a check of the plugin's version on GitHub. The update check carries no account, ID, stats or chat data (GitHub sees your IP address, as with any download), and you can turn it off. With optional report cards on, Claude Code downloads a small plugin from GitHub (and the matching one after an app update), a small check runs after each AI reply and reads the end of that chat on your computer, and when you ask for a report card your AI reads a short summary of the chat that the app makes on your computer. See PRIVACY.md. |
| ✋ Asks first | Every source is off until you Allow it. Pause it, stop it, or Forget it (deletes what it read, leaving only finished goal weeks and trophies, which are totals) any time. |
| 🔢 Numbers, not chats | Ordinary messages are only counted, never stored. For corrections you pick full text, snippet (the default) or category only, and kept text steps down on a timer you set. Prompts, goals and lessons store numbers only, never words, and the prompt checker saves nothing you type. |
| ✂️ Redacted | Keys, emails, links, home paths and long numbers are stripped from any text kept. It's pattern-based, so it can miss things: pick category only if that matters. |
| ✅ Verifiable | MIT open source. The Mac app is signed and notarized by Apple, updates are signed, and every release ships SHA256SUMS.txt. |
The full details are in PRIVACY.md.
Does it send my chats anywhere?
No. Everything is read and scored on your computer, and the results live in a local SQLite file. There is no server to send them to. The only time the app sends text anywhere is if you share a Hall of fame card, and you see the card before it goes anywhere. If you turn on report cards, your AI reads a short summary of that chat, including the first few words of each of your messages, and that goes to your AI provider like the rest of the chat.
Does it use AI to judge my chats?
No. Using an AI to grade AI would mean uploading your chats and getting a different answer each run. Scoring is plain pattern rules you can read in rules.rs and prompt.rs. The one exception is the optional report card, which your own AI writes when you ask. It never changes your score.
Which AI tools does it work with?
Claude Code, Claude Cowork and Codex live, plus ChatGPT and Claude.ai through their data exports. More are coming; see Supported tools.
Will the lessons go out of date?
They're written not to. They teach habits that work with any collaborator, human or AI (name a target, give evidence, set a finish line, add guardrails, share the purpose, say what you already know), never tool tricks or magic phrases. A test fails the build if a lesson names a product, a model or a version, and the real scoring rules grade every example, so a lesson can never disagree with your score.
Will it slow my computer down?
The engine is native Rust, and the window uses your system's built-in web view instead of shipping its own browser. The first read of your history runs files in parallel. After that it only reads what's new since its last checkpoint.
If I prompt it better, it makes fewer mistakes. Doesn't that lower my score?
No. Corrections only have to show you check its work: 3 in 28 days is full credit (1 in 7 days), and more add nothing. After that, what counts is how rarely the AI has to correct you, per task. Most of the score (70%) is how often your prompts are clear the first time, so prompting better is the fastest way up.
Is a low prompt score bad?
It's a nudge, not a grade on you. A prompt is "clear the first time" when nothing on you followed it: you didn't add to it, clarify it or start it over. The AI getting it wrong never counts against you; that's a correction for you instead. One-shot prompts don't count at all, since they can't need a follow-up.
It counted something wrong. Now what?
Rules miss sometimes. Open a Wrong correction issue with the message redacted. Fixes usually come with a test for your example, so they stay fixed. When rules change, your score is rebuilt from history automatically.
Can my employer see my score?
Not through the app: there's no account, no dashboard and no sync. Your score lives in a file on your machine (not encrypted, so anyone with access to your computer could read it, like any other file).
It's also a for-fun self-check, not a measure of job performance, and it shouldn't be used to evaluate anyone.
Does a report card change my score?
No. Your score comes from fixed rules and never looks at report cards.
Why does a report card cost tokens when the rest of the app doesn't?
Your own AI writes it, inside Claude Code or Codex, so it uses some of your tokens. The app itself never calls an AI. Report cards are off until you turn them on, and tips are a separate setting that is off by default.
What are tips?
An optional setting for Claude Code: Settings → Report cards → Give me tips as I work. When a follow-up could have been in your first message, your AI ends that reply with one line, like "Tip from No, Actually...: next time, say it only breaks on Safari when you first ask." Most replies get nothing. The part in brackets at the end, like [evidence, redid 50%], is for the app: the habit, and how much work was redone. Letter grades still come only from /grade:me.
It adds about 480 tokens when a chat starts and again after a long chat is compacted, plus a few for each tip. That came to about 2.5% more tokens overall in testing. It applies to new or reopened chats. To stop sooner, tell your AI "no more tips." Settings shows the exact instructions your AI gets. The app counts tips and keeps which habit each was about, never the words. Your AI decides when to give a tip, so it can miss some, and Claude Haiku misses more and is less accurate.
Can I see my old report cards?
The app keeps each report card's numbers, so you can see how you're trending. The words stay in your chat history.
How do I delete everything?
Click Forget on a source to delete what was read from it, including its pins, report cards and tips. Finished goal weeks and trophies stay, as totals. Delete report card and tip history in Settings clears them all. Or click Delete everything in Settings: it turns off report cards and Launch at login, deletes the data and logs, and quits. If the app is already gone, the manual steps are in PRIVACY.md. Uninstalling alone leaves the data behind.
Is it really free? What does it cost to run?
Free, and it costs nothing to run. The score never calls an AI, so there's no API key to add and no tokens on your bill, however much history you have. Optional report cards are written by your own AI, so they use some of your tokens, only when you ask, plus a little in each chat if you turn on tips. MIT licensed, no paid tier, no ads.
Over the window you pick (the last 28 days unless you choose another), once it has at least 10 tasks with more than one message:
score = 70% × direction + 30% × judgment
direction = % of your prompts that were clear the first time (see "How a prompt is graded")
judgment = 100, minus 5 for each 1% of your tasks where the AI had to correct you
(counted against at least 30 tasks, so a quiet month isn't wiped out),
scaled down until you've corrected it 3 times
(a share of that in a window under 28 days: once in 7 days)
| Tier | Score | Answer |
|---|---|---|
| Irreplaceable | 90+ | No. |
| Sharp-eyed | 80–89 | No. |
| Coworker | 65–79 | No. |
| Rubber stamp | under 65 | Hmm. |
Each question below the score gets a one-word answer too, and that word is its tier. Each tier also has a short name, shown in the quick glance and on share cards.
Do you make it guess? goes by the share of prompts that were clear the first time (see "How a prompt is graded"):
| Answer | Clear the first time | Name |
|---|---|---|
| Rarely. | 90%+ | Architect |
| Sometimes. | 80–89% | Tech lead |
| Often. | 70–79% | Ticket writer |
| Constantly. | under 70% | Vibes only |
Do you correct it? goes by judgment:
| Answer | Judgment | Name |
|---|---|---|
| Religiously. | 90+ | Hawk-eyed |
| Often. | 75–89 | Reviewer |
| Sometimes. | 50–74 | Spot-checker |
| Rarely. | under 50 | Autopilot |
A few things this gets right on purpose:
- Prompting better never costs you. Past 3 corrections, more corrections add nothing, so an AI that makes fewer mistakes because you prompted it well doesn't lower your score.
- No corrections at all caps you. Getting everything clear the first time but never correcting anything is a Coworker (70) at best: there's no sign you check its work.
- Being corrected counts, per task. If the AI keeps correcting you, judgment drops. It's measured per task, never against your corrections, so it doesn't depend on how often the AI is wrong.
| Label | Rule |
|---|---|
| Confirmed | You push back and the AI admits the mistake |
| Admitted | The AI's next reply opens with "you're right", "good catch", "my mistake"… or it checks first and then says you were right |
| Claimed | Your message opens with "no, that…", "still not working", "you missed"… (not when you're answering its question) |
| Humbled | The AI corrects you ("Actually…", "Not quite") |
Adding context ("also", "fyi", "I forgot to mention") never counts against the AI. Each correction is charged to the model that wrote the reply you corrected.
-
A task starts with a chat's first message, or a later one that opens with "now…", "next:" or "new task". Hellos, "continue" and questions ("what does this do?") aren't tasks.
-
A prompt's grade counts what it includes, never its length:
- Target: a file, function, page or link
- Evidence: an error, output, an example, anything pasted
- Finish line: the expected result, tests, a format
- Guardrails: don't, only, keep, without
- Purpose: who it's for or what it's part of
Three or more is specific, two is decent, fewer is vague. Pasted blocks count as evidence, but their words aren't graded as yours.
-
A follow-up is a later message that redirects the task. Either the AI got it wrong (a correction), or it's on you: added more ("also…", "I forgot to mention"), clarified ("I meant…", "to be clear") or started over ("scrap that", "revert it").
-
A prompt is clear the first time when nothing on you followed it. The answer (Rarely, Sometimes, Often, Constantly) comes from the share of prompts that were clear the first time over the last 28 days, counting tasks with more than one message, once there are at least 10 of them.
The tip names the thing your prompts include least. It only quotes "prompts with guardrails were clear the first time X% vs Y%" when your own history shows a real gap.
- One goal at a time, across all AI: a habit ("include guardrails in 25% of your prompts"), a ceiling on one kind of follow-up ("add more after sending in 15% of tasks or fewer"), or a tier to reach.
- Suggested from your numbers. A habit goal asks for 15 points more than your last 8 weeks, capped at 80%, so it's a stretch, not a wall.
- Weeks run Monday to Sunday, local time, counting only tasks from the day you set the goal. A week needs 5 prompts to count; quieter weeks pause the goal instead of breaking your streak.
- Met 4 weeks, it's done, and it goes on your trophy shelf. Finished weeks are frozen, so a rules update or an import can't take back a week you earned.
There's no public benchmark for weekly per-person scores, so every threshold was checked three ways:
- The author's history: 8,600 messages.
- Public chats: about 3,400 English chats from WildChat-1M, run through these exact rules.
- Published figures: Claude Code users interrupt in 5–9% of turns (Anthropic); in multi-turn developer chats in the DevGPT dataset, 14–17% of prompts refine the request.
What that showed:
- The correction rate agrees across sources: 1 in 47 messages for the author, 1 in 41 in WildChat.
- Public chats are short and almost never need follow-ups, so the prompt floors come from one heavy user, whose typical month had 86% of prompts clear the first time (Tech lead). Expect them to move as more people's numbers exist.
- Longer prompts didn't mean fewer follow-ups. Bigger tasks get both longer prompts and more back-and-forth, which is why the tier scores outcomes, not length.
- The overall score puts the author's typical month at 86 (Sharp-eyed): 83% of 150 prompts clear the first time, 41 corrections, and the AI corrected them twice.
| Tool | How |
|---|---|
| Claude Code | Live |
| Claude Cowork | Live |
| Codex (CLI and desktop) | Live |
| ChatGPT | Import your data export |
| Claude.ai | Import your data export |
Planned: Gemini CLI, Ollama, Open WebUI, LM Studio, VS Code Copilot, Cursor and Aider, plus a browser extension for live ChatGPT, Claude.ai and Gemini. Want one sooner? Request it, or add it yourself: an adapter is about 100 lines plus a test (see CONTRIBUTING.md).
Open the latest release and scroll to Assets (click it to expand the list if needed). Download the one file for your computer:
| Your computer | Download | Install |
|---|---|---|
| Mac (macOS 12+, Apple silicon or Intel) | No-Actually-<version>-Mac-universal.dmg |
Open it, drag No, Actually... into Applications, and open it from there. It's signed and notarized, so there are no warnings. |
| Windows (64-bit, with Microsoft Edge WebView2) | No-Actually-<version>-Windows-Setup-x64.exe |
Run it. If you see "Windows protected your PC", click More info → Run anyway (the Windows build isn't code-signed yet). |
| Linux (x86_64, built on Ubuntu 22.04 with WebKitGTK 4.1) | No-Actually-<version>-x86_64.AppImage |
Right-click → Properties → allow it to run as a program (or chmod +x), then open it. |
The other files are for special cases:
-Windows-IT-Installer-x64.msiis for IT teams deploying to many PCs.-Linux-Ubuntu-Debian-amd64.deband-Linux-Fedora-x86_64.rpminstall through your package manager.SHA256SUMS.txtlets you verify a download.
The app checks for new versions and shows Restart to update when one is ready. Nothing is downloaded or installed until you click it, and an update only installs if its signature matches the project's key. Turn the check off in Settings → General. To get new versions first, set Update channel to Beta there. Betas may have bugs. Going back to Stable keeps your version until a newer full release comes out.
- Silent install (Windows):
msiexec /i <file>.msi /qn, using the-Windows-IT-Installer-x64.msifile. - Updates: the settings are per user, in Settings → General. There is no machine-wide switch yet.
- Data: it lives per user, at the paths listed in PRIVACY.md.
- Code signing: the Windows build isn't code-signed.
Needs rustup (rust-toolchain.toml picks the Rust version), Node 20.19+ and pnpm 10. On Linux, also install the Tauri packages listed in CONTRIBUTING.md:
git clone https://github.com/HTMLGuyLLC/NoActually.git
cd NoActually/app && pnpm install && pnpm tauri build -c '{"bundle":{"createUpdaterArtifacts":false}}'The app lands in target/release/bundle/. (The -c part skips the signed update files, which need the project's private key.) For development, use pnpm tauri dev. To try the UI in a browser on sample data, run pnpm dev and open http://localhost:1420.
The same engine and database work from a terminal. Releases don't include the command line, so install it from a checkout of this repo:
cargo install --path crates/cli
noactually sources # what it can read, and what you've allowed
noactually allow claude-code
noactually sync
noactually stats --family "Claude Opus"
noactually check "Fix the crash in checkout.ts. Done when the tests pass."stats prints the overall score and both halves. check "<prompt>" grades a draft prompt, like the app's checker. Set NOACTUALLY_DB=/path/to/test.db to use a different database than your real one. Every command is in docs/ARCHITECTURE.md.
crates/engine correction rules, prompt grading, the overall score, model names, redaction: pure Rust, no I/O, WASM-ready
crates/sources one adapter per AI tool (Claude Code, Codex, exports…)
crates/store SQLite: consent, checkpoints, daily counts, tasks, goals, retention, rules versioning
crates/cli the `noactually` command
plugins/grade the report card plugin: the /grade:me skill and its hooks
app/ Tauri 2 desktop app: menu bar, scorecard, Improve and Learn, notifications
- Crash-safe checkpoints. A turn is only counted once its reply exists, and each file's writes commit in a single transaction.
- Hashed message keys. Resumed sessions, Claude Code forks and re-imports are never counted twice.
- Rules versioning. When the rules change, history is counted again. Chats whose files are gone, or whose source is stopped, keep their old counts.
The full design is in docs/ARCHITECTURE.md.
Issues and PRs are welcome. Good places to start:
- Support for another AI tool
- A rule that miscounted one of your chats (use the Wrong correction form, and redact the text first)
- Polish on the app
Read CONTRIBUTING.md, then run the same checks as CI with just (brew install just or cargo install just):
just ciWithout just, the same commands are listed by hand in CONTRIBUTING.md.
MIT © HTMLGuy LLC. Provided as is, without warranty. Bundled fonts (SIL OFL 1.1) and dependency licenses are listed in THIRD_PARTY_NOTICES.md.
Not affiliated with or endorsed by Anthropic, OpenAI or any other AI provider. Claude, ChatGPT, Codex and other product names are trademarks of their owners, used here only to say which tools it can read.