fucking-cheap

The Problem

Cheap models are unreliable.
They forget constraints, bịa láo,
and loop endlessly on bugs.

10×

More fix iterations than GPT-4

∞

Infinite loops without stop conditions

$0

Budget left after debugging

The Solution

Impose discipline.
Not intelligence.

fucking-cheap is a modular skill system that wraps cheap models with hard rules: plan before code, test before commit, stop when stuck. No more infinite loops. No more forgotten constraints.

How It Works

Four phases. Zero ambiguity.

01

INTAKE

Read the task. Classify scope: Spike, Bounded, or Architectural. Ask clarifying questions before assuming.

02

DECOMPOSE

Break into atomic subtasks. Each = 1 file, 1 function, 1 test. Present plan. Wait for approval. Never skip this.

03

TDD LOOP

Write test first. Write code to pass. Fix max 2 times. Stop if still failing. No infinite loops allowed.

04

VERIFY

Run full test suite. Playwright for UI. Integration tests. Report what passed, what failed, what changed.

Architecture

22 modules. Lazy-loaded.
Thông minh hơn.

Core

7 modules

workflow.md — state machine
constraints.md — anti-drift
evidence.md — verify facts
verification.md — test/build/lint
completion-gate.md — DONE rules
failure-recovery.md — debug flow
vibecoding.md — plan→TDD→delegate

Domains

6 modules

frontend.md
backend.md
database.md
security.md
concurrency.md
performance.md

Protocols

5 modules

anti-loop.md
context-compression.md
git-safety.md
regression-guard.md
model-escalation.md

State Machine

INTAKE → RECON → PLAN → IMPLEMENT → VERIFY → REVIEW → DONE

No IMPLEMENT → DONE shortcut

In Action

One task. One flow.
No guesswork.

$ User: "Add margin import to Lagtuz"

→ INTAKE     Scoping: Architectural (multi-file)
→ RECON      Reading codebase... found 3 related files
→ PLAN       Decomposed into 5 subtasks
              ┌─────────────────────────────────────┐
              │ Subtask 1: API route     → Test: 200 │
              │ Subtask 2: Parse Excel   → Test: JSON│
              │ Subtask 3: UI component  → Test: render│
              │ Subtask 4: Connect       → Test: flow│
              │ Subtask 5: Integration   → Test: all │
              └─────────────────────────────────────┘
✓ APPROVED   User said "OK"

→ IMPLEMENT  Subtask 1: test ✅ code ✅
              Subtask 2: test ✅ code ✅
              Subtask 3: test ✅ code ✅ (fix 1x)
              Subtask 4: test ✅ code ✅
              Subtask 5: test ✅ code ✅

→ VERIFY     Unit: 12/12 pass
              Integration: 3/3 pass
              UI: 2/2 pass

→ DONE       5 files changed, 847 insertions(+)

Không lo model đắt đỏ, fucking cheap,
bắt đầu vibing.

Install Now Read Docs

Cheap models are unreliable. They forget constraints, bịa láo, and loop endlessly on bugs.

Impose discipline. Not intelligence.