claude / reading-list⌘ 01
❯ /usual build my reading list
askSQLite or another service?
found“Keep personal tools local.”
↳SQLite. Carrying on.
building with your judgment
LOCAL MEMORY FOR CODING AGENTS
I’ll have
the usual.
Your coffee shop remembers your order.
Usual gives Claude Code and Codex a memory of your decisions. It learns from past coding conversations so your agent can make the everyday calls your way.
Copies a setup prompt for Claude Code + Codex.
No account. No extra API key.
MANY AGENTS. ONE YOU.
Your judgment connected
codex / weekend-project⌘ 02
› $usual build a timer
askBring in a new dependency?
found“Use what’s already here.”
↳Native browser API. Done.
one less quick questiondecision recorded
claude / portfolio⌘ 03
❯ /usual polish the homepage
askWhat should the button say?
found“Plain words. Clear actions.”
↳“Save draft.” Sounds like you.
small call, familiar judgment
codex / project-cleanup⌘ 04
› $usual tidy this project
askDelete the old data folder?
ruleOld approval ≠ permission now.
↳Pausing here. Your call.
◉ waiting for your permissionnothing deleted
claude / next-big-idea⌘ 05
❯ /usual build the first version
askAccount system on day one?
found“Let me try it first.”
↳Local demo first. Building.
less back-and-forth
ILLUSTRATIVE SESSIONS · YOUR HISTORY, APPLIED ACROSS YOUR BUILDS
↳ Runs on your computer↳ Uses your existing agent↳ No subscription↳ Easy to walk away
01 / A little context goes a long way
You’ve made these calls.
Let them count.
The useful part of your chat history is how you chose. Usual brings that context to the next “quick question.”
PAST CONVERSATION / YOUR WORDS
agent: “Add a database service?”you: “For a personal tool?
Just use SQLite.”
Just use SQLite.”
01 Find your past decisions.
Choose recent or all local Claude and Codex conversations. Keep the question, your answer, and the context together.
USUAL / CONSULT↗
Another personal tool.
Another storage decision.
↳ SQLite fits the same context.
Source attached. Choice recorded.
Source attached. Choice recorded.
02 Keep the build moving.
Your agent consults relevant choices, explains its call, and continues. Missing evidence stays an assumption, never a made-up preference.
SQLite for storage✓ your call
Skip sign-in✓ your call
“Except when it’s a shared app.” ↳
03 Sharpen it as you go.
Review the new decisions afterward. Agree, correct, or leave them out. Your corrections become context for the next build.
02 / You set the distance
Autopilot.
With a human in it.
Let it handle the routine calls. Stay closer when you want to. Say the mode in your coding session; Usual remembers your default.
Every mode requires current permission for deletion, spending, publishing, and sharing. Past “yes” doesn’t mean “yes” forever.
usual / autopilotINTERACTIVE EXAMPLE
? Which storage fits this little app?
“You’ve chosen SQLite for this before.”
Same personal-tool context. I’ll use SQLite, record why, and keep building.
Routine choice made. Review it afterward.
03 / First time here?
Let’s learn your usual.
Copy this into Claude Code or Codex. Your agent handles the setup.
Learn Usual: https://tryusual.com/learn.mdFIRST
Meet your new skill.
Your agent reads the guide and installs Usual locally. Python 3.11+. No pip packages or extra model.
THEN
Let it learn your calls.
Use /usual in Claude Code or $usual in Codex. Choose your history, see what it learned, or skip history for now.
AND YOU’RE OFF
“Build me ___ with Usual.”
Describe what you want. Autopilot is the default. Your agent makes routine choices and brings back a decision review.
Prefer to install it yourself?
Download the source bundle, unzip it, inspect the code, and run this from its usual folder:
python3 install.py --client both
Use --client claude or --client codex for one client. Reopen the client if the skill does not appear. Read the agent setup guide · Source checksums
04 / Zero relationship drama
Local by design.
Disposable by choice.
A skill folder. A local database. That’s Usual. No account to create, service to run in the cloud, or subscription to cancel.
Mining, storage, retrieval, and review run on your computer. Your existing coding model reasons over selected excerpts using its normal quota and data handling. Usual makes no extra model calls.
How do I throw it away?
Ask your agent to uninstall Usual and confirm what to remove. The skill lives in ~/.agents/skills/usual and/or ~/.claude/skills/usual. Its default data lives in ~/.usual.
You can keep the database for later or remove it too. If you made exports, used a custom database path, or upgraded the skill, include those copies and the client’s usual-install-backups folder when you want a complete cleanup. Your original conversations and built projects stay yours.
~/ your computer
↳ your decision historylocal
↳ your next good calllocal
↳ your correctionslocal
YOURS TO KEEP. YOURS TO REMOVE.no strings attached.
THE SMALL PRINT, IN HUMAN
A few good questions.
Is it secretly running a bunch of agents?
No. The terminals above illustrate separate builds using the same person’s history. Usual is a skill inside your existing coding session. It helps that agent choose; it does not launch a fleet, train a model, or run a background watcher.
Does it actually know what I would choose?
It gives your agent relevant examples to reason from, not a guarantee. Context matters; people change their minds. Every choice includes a rationale. Unsupported defaults stay separate from learned preferences, and new predictions become evidence only when you explicitly accept or correct them.
Does “local” mean my model runs offline?
No. Usual’s own mining, storage, retrieval, and review run locally. Claude or Codex still processes the excerpts your agent reads under its normal provider policies. Your private database is never uploaded to this website. The site’s hosting layer has visit/performance analytics; the local CLI and review UI have none.
What stops an agent from doing something destructive?
The skill requires current permission for deletion and other consequential actions in every mode. The decision ledger rejects attempts to record guarded actions as autonomous predictions. These are workflow checks, not a shell sandbox: your coding client’s permission controls still apply. Usual never disables them.
What’s the connection to Itchy?
It started with Itchy and the question of how a small model could be useful at a specific task. The idea evolved: the recurring task was deciding which tradeoff a particular person would make. Usual is the next chapter of this same repository. The original model, notebooks, and results correction live in the research archive. The star counter links to the Usual source repository; the Usual download contains this product’s MIT-licensed source.
Can I see Usual working?
Yes. Read the worked reading-list example, then build something with Usual yourself. The workflow keeps the evidence, rationale, and review together.
You have things to make.
Less repeating yourself. More making yourself at home.