Smallest Complete — Stop one task from becoming ten.

A lightweight Skill for Codex and ChatGPT agents: finish exactly what was asked, prove it works, and stop before the side quests. For non-trivial coding and architecture work, it also applies an elegant-architecture lens—clear control, explicit ownership, and only the state the real job requires.

Four-panel comic: a developer asks an agent to fix one button; the agent builds a state machine, sync layer, recovery system, and version history; Smallest Complete brings it back to a fixed and tested button

Install with one prompt

Paste → send → start a new task.

Install Smallest Complete from https://github.com/JetXu-LLM/smallest-complete.
Follow INSTALL.md exactly.

Codex reads the installation contract, protects your existing setup, installs the Skill, and verifies the result. When it finishes, start a new task.

Review every install, update, and uninstall step →

Installs Preserves Never adds
One Skill directory + one global activation block Existing Skills and AGENTS.md instructions Runtime, hooks, dependencies, accounts, or telemetry

One Skill, two disciplines

Smallest Complete watches for the same failure at two moments: doing more than was authorized, and making necessary code or architecture more tangled than the evidence requires.

Scope discipline Architecture discipline
Applies to Complex Codex and ChatGPT Work tasks Non-trivial coding, debugging, refactoring, migration, system design, or architecture work
Question Is this inside what was actually authorized? Is this the clearest structure the evidence requires?
Stops Scope creep and adjacent “helpful” work Extra decision owners, shared workflow state, and tangled coordination
Source Core SKILL.md Conditional elegant-architecture.md reference

First it keeps the result inside the ask. Then, when the task genuinely requires code or architecture, it keeps the structure clear.

When the task really needs architecture

Scope discipline decides whether something belongs in the job. Architecture discipline decides whether the necessary software structure has earned its ongoing cost.

For architecture design, non-trivial coding, refactoring, migration, or debugging that may change ownership, control flow, state, interfaces, or operations, the Skill reads elegant-architecture.md before planning or editing.

Architecture comparison: accidental complexity with overlapping decision owners and shared state versus elegant architecture with one control path, independent capability modules, honest state, and bounded agent judgment

Read the complete architecture guidance →

The core contract

Principle Meaning
Smallest No adjacent deliverables, speculative systems, or permanent machinery.
Complete The requested result actually works and required behavior is preserved.
Proven Completion claims match observable evidence.
Stop Useful discoveries do not silently become new work.

In practice

You ask Smallest Complete response
“Fix CSV export when descriptions contain commas.” Fix escaping at the owning boundary, test it, stop. No export platform.
“Refactor this ingestion workflow.” Keep one clear control path, a thin coordinator, capability-owned rules, and only required state. No parallel architecture “just in case.”
“Turn these notes into five slides.” Research what the deck needs, deliver five strong slides, stop. No brand system.

For research, writing, analysis, and other ChatGPT Work tasks, inquiry stays as broad as the requested result needs. The boundary applies to deliverables and actions—not to useful thinking.

Deliberately small
Runtime None
Background process None
Skill network calls None
Telemetry None
Guarantee None—it is guidance for capable agents, not an enforcement layer

The complete mechanism is one Skill, one conditional architecture reference, and one activation paragraph. The project practices what it asks agents to do.

Go deeper

Bring us the case we missed

The most useful contribution is not agreement. It is a concrete task where the Skill helped, failed, made no difference, activated at the wrong time, or gave architecture advice that was wrong for the real system.

Comparative runs and counterexamples are especially welcome. You do not need to propose a fix—a sanitized prompt, expected result, observed behavior, and the available evidence are enough to start.

Open a behavior report → · See what makes a useful contribution →

Star it if it earned it

If Smallest Complete stopped one bounded task from becoming an architecture project—or helped you build the architecture the task actually needed—click Star at the top of this page. It helps the next developer find it before their next five-line fix becomes a framework.

License

MIT. Independent project; not affiliated with or endorsed by OpenAI or Anthropic.