A lightweight Skill for Codex and ChatGPT agents: once a complex job is clear
enough to act, finish exactly what was asked, prove it works, and stop before the
side quests. For non-trivial coding and architecture work, it conditionally applies an
elegant-architecture
lens—clear control, explicit ownership, stable handoffs, and only the state the
real job requires. When the choice of tests or other validation evidence
materially affects completion, it applies
evidence-calibrated-testing
to derive evidence from the real claim rather than from test volume. It works
backward from the receiver's next action, then delivers one value path within
the minimum hard boundaries that must hold now. Handoffs are complete relative
to the receiver's current state, and reader-facing results are accepted in their
actual medium and surrounding context rather than from producer checks alone.
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 |
Smallest Complete keeps scope, architecture, and validation tied to the same real job.
| Scope discipline | Architecture discipline | Testing discipline | |
|---|---|---|---|
| Applies to | Complex Codex and ChatGPT Work tasks | Non-trivial coding, debugging, refactoring, migration, system design, or architecture work | Material test/evidence design, escaped failures, receiver/operational/real-run claims, or focused-versus-full selection |
| Question | Is this inside what was actually authorized? | Is this the clearest structure the evidence requires? | What failure must the evidence distinguish, at which real boundary? |
| Stops | Scope creep and adjacent “helpful” work | Extra decision owners, brittle handoffs, speculative defenses, and tangled coordination | Self-certified fixtures, proxy-green completion, and low-information reruns |
| Source | Core SKILL.md |
Conditional elegant-architecture.md reference |
Conditional evidence-calibrated-testing.md reference |
The runtime references load only when their decisions are material. A simple task
stays simple. The package also includes a
casebook and a
self-contained evaluation rubric
for learning, revising, and testing the method; ordinary execution does not load them.
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.
Read the complete architecture guidance →
For non-trivial test strategy, escaped defects, receiver, operational, or
real-run claims, or a consequential choice among focused, broad, full, and
other completion evidence, the Skill reads
evidence-calibrated-testing.md.
It derives failure scenarios from real losses, receivers, operation, and semantic
impact; chooses the matching oracle and boundary; and limits every green claim
to the path actually exercised. Routine local checks do not load the reference.
Read the complete testing guidance →
| Principle | Meaning |
|---|---|
| Smallest | No adjacent deliverables, speculative systems, or permanent machinery. |
| Complete | The requested result works for its next receiver in the real medium, with required behavior preserved. |
| Coherent | Long work keeps one route; local fixes fit it or replace part of it, and declared handoffs stay stable. |
| Proven | Completion claims match observable evidence. |
| Stop | Useful discoveries do not silently become new work. |
| 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.” | Exercise representative real input through the receiver's next action, keep one control path, and add hardening only when a current boundary or evidence requires it. |
| “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.
| 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 runtime mechanism is one Skill, two conditional references, and one activation paragraph. Two non-runtime evaluation references ship beside it so the method can be tested without making evaluation ceremony part of every task.
- The complete elegant-architecture reference
- The complete evidence-calibrated-testing reference
- The phase-specific casebook
- The self-contained evaluation rubric
- Why capable agents expand the mission
- Design and architectural choices
- Evaluation without invented success rates
- The complete Skill source
- The exact global activation block
- Contributing
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 →
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.
MIT. Independent project; not affiliated with or endorsed by OpenAI or Anthropic.