Why do in 2 lines what you could do in 2,000?
An AI coding agent can solve FizzBuzz in a few lines. Mullet helps it recover from that mistake.
Mullet is a skill that makes working software as complex as possible. More layers. More indirection. More languages involved in returning a string. If a function can call another function directly, there is still work to do.
You have an AI that can write thousands of lines a minute. Why are you asking it to stop?
Companies want a return on their investment in AI. Fortunately, lines of code are very easy to count.
This repository contains 257,080 lines of example code. So far, it can print FizzBuzz and remember to buy milk. Both features worked much earlier in development, but we kept improving them.
Mullet defaults to 200 layers of abstraction, with 16 separately priced contracts behind each layer in full mode. It gives those layers something to do. Money moves, messages change formats, claims collide, and the answer eventually comes out the other end carrying a receipt. The tests still have to pass. Otherwise you could just paste the word Factory into a text file and go home.
If you aren't personally responsible for maintaining 1,000,000,000 lines of code, your job is already at risk.
The examples contain 220,209 generated lines in three statute books and 36,002 JVM assembly instructions, plus the application and shared runtime source. Those statutes execute on the request path. The total counts physical source lines, including blanks. Reproduce it and check that the applications still work:
git ls-files 'examples/*.mjs' 'examples/*.rb' 'examples/*.awk' 'examples/*.py' 'examples/*.sql' 'examples/*.java' 'examples/*.jasm' | xargs wc -l
python3 tests/ratification.py
That leaves 999,742,920 lines. At the current rate, we may need a third feature. Benchmark methods
For Windows builds, Mullet prefers to put business logic in SQL Server stored procedures that call .NET assemblies through SQL CLR. The JVM still gets its ten percent. Both runtimes can now bill for the same checkbox.
In Claude Code, install Mullet from GitHub.
/plugin install mullet --marketplace russell-dot-js/mullet
Then ask your agent:
Use $mullet to build a URL shortener. Every redirect should be an architectural decision. I want the URLs to get shorter and the codebase to get longer.
Plugins, hooks, mode switches, editor rules and seven skills ship together. Claude Code gets a task board so you can watch your AI spending money in real time. Installation for Codex and other hosts · Integration coverage
Every business operation deserves three independent implementations. At runtime Mullet draws one at random. In FizzBuzz the registry sorts the claims, the queue clerk inserts each claimant ahead of its juniors, and the auctioneer keeps selling first place until everybody has one. All three offices agree that 15 says FizzBuzz. None has agreed to close.
Run node examples/fizzbuzz/app.mjs --case 15 --minutes to see which bench heard your case. Add --bench auction to subpoena the same office again.
node examples/fizzbuzz/app.mjs --case 15
JavaScript punches an 80-column card. Ruby surveys it. Fizz and Buzz both claim the number 15, so a court settles their dispute. AWK counts the hisses. A local Kafka-style log moves the paperwork through consumer groups. Two hundred hearings settle 3,200 contracts and extend a hash chain.
The mint now owns ten percent of the repository by line count. Its die has 9,000 numbered teeth. A Java assembler turns the die sheet into a JVM method, which stamps each printer permit and each Todo deed. Removing a tooth changes the serial number. The Treasury checks its stake with python3 tests/mint.py and counts even the README against itself.
At this point we have almost printed a word.
Python receives the printing request as JSON inside a SOAP envelope. It checks the quarterly forecast. The forecast missed, so printing moves to a sale-and-leaseback dot-matrix subsidiary. XML specifies the newline in octal.
Authorized roadside sign: FizzBuzz
The newline is leased. Please return it when you finish reading.
For the first hundred numbers, run node examples/fizzbuzz/app.mjs 100. Add --minutes to the case command for every roll call. Add --edition restated for BuzzFizz, because the requirements changed after we finished the other 73,000 lines.
From a fresh checkout, try this against a new database:
python3 examples/todo/app.py --db charts.sqlite3 add "Buy milk"
python3 examples/todo/app.py --db charts.sqlite3 add "BUY MILK"
python3 examples/todo/app.py --db charts.sqlite3 done 2
python3 examples/todo/app.py --db charts.sqlite3 history 2
Todo began as a list of things to do. Then someone asked who owns the X.
The Brotherhood Pension Trust leases one diagonal. The Right Bracket HOA leases the other. SQLite refuses to mark the task done until both landlords have been paid. The X is then pressed as a record, which raises a separate question: who wrote the song?
BUY MILK is a cover of Buy milk. Capitalization and a dramatic are insufficient to claim authorship. A second committee convenes, bringing completion to 400 hearings and 6,400 settled contracts. The original task collects a 35-cent songwriting royalty while remaining unchecked. It has finally found a way to earn money without doing anything.
The app explains its decision:
Capital letters and a longer between words do not entitle the band to a songwriting credit.
Run history 1 against the same database to inspect the original task's publishing income. Run list and all this becomes [x] 2 BUY MILK. Repeating done 2 pays the royalty once. The royalty accountants have filed an appeal.
Both examples run locally with a JDK 11+, Node.js 18+, Ruby, AWK and Python 3.8+ with SQLite. Todo uses mullet.sqlite3 unless you supply --db. Task IDs come from add. Example source and eleven more ways to complicate your afternoon
We wanted the skill itself to set a good example.
Choose municipality, receivership, submarine or shareholder, then specify the layers, quorum, insurers, validators, schema generations, obsolete wire format and disaster you would like the architecture to survive. You can also choose who writes the minority report.
node skills/mullet/scripts/commission.mjs submarine layers=320 \
'sacred_asset=the last dry pixel' escrow_percent=97 \
'legacy_medium=a fax from the seabed'
This prints a design brief for your agent to implement. Explicit settings override the preset. If you can explain what every setting does, raise layers. All 46 configuration inputs
Mulletize keeps adding complexity to a working application:
Use $mulletize on examples/todo. Keep going until I . Every round should add more code and another abstraction the application actually uses.
Each round reads the previous result, expands the implementation and verifies the original behavior. The headless supervisor runs your acceptance command and checks for a net increase in source lines before recording the round. Then it launches another round. There is no default iteration limit.
The ledger survives interruptions in the Git worktree. `` suspends the expansion. resume allows it to continue. The foreground supervisor needs Python 3.9+ and a running coding agent. Set --cycles 2 if you only want to make things a little worse. Run the supervisor
We slipped an extra cent into the price of an X. The books still balanced according to a witness who had been calling the function and admiring its return value. That witness now has a test.
Run npm test for the full inquiry, including the original applications, interrupted settlements, rival claimants and the witness lineup. The lineup tries a reversed naming-rights decision and a fraudulent diagonal surcharge against the real assertions. Internal Affairs then substitutes a witness who agrees with everything and another who calls in sick. Both must fail their audit. We then replace Internal Affairs with a rubber stamp and test whether anyone notices.
Every one of the 9,600 underwriters also undergoes five compulsory insolvency rehearsals. Every hearing must practice losing its own vote. The newspaper runs a 10,000-copy recall because one subscriber ate page 417. npm run coverage prints line, branch and function coverage for the Node test run.
The example tests went through two rounds of mulletize. In round one, Green Ink Mutual wrote 3,200 policies on the right to call a test green. In round two, the typesetter won two bytes of the checkmark in a divorce. To print ✓, the test must redeem a coupon for the third. Trying to expense the FizzBuzz coupon to Todo is a test case.
Run the inquiry with node --test tests/cross-examination.test.js. To reprint its 73,403 lines of insurance policies from the charter, run:
node skills/mullet/scripts/charter.mjs tests/green-ink/charter.json tests/green-ink/policies.mjs
| Skill | Purpose |
|---|---|
mullet |
Add the feature and enough architecture to make it worth a rewrite. |
mulletize |
Keep adding. |
mullet-review |
Inspect a diff for defects and dangerously direct function calls. |
mullet-audit |
Find places in the whole repository where something is still understandable. |
mullet-debt |
List shortcuts that could be replaced with longcuts. |
mullet-gain |
Count the code so you can say “productivity” with a straight face. |
mullet-help |
Seven skills was apparently already too many. |
By russell-dot-js. Inspired by Ponytail and FizzBuzzEnterpriseEdition. Host adapters build on Ponytail's MIT-licensed code. Ponytail · FizzBuzzEnterpriseEdition · Upstream license