Once for all.

Ephapax (ἐφάπαξ) is a Greek word used in the New Testament to describe an act done so completely and finally that it never needs to be repeated. One act. Settled forever.

That is the discipline this skill enforces. Reason through a problem once. Codify the result. It stands forever. The same cognition must never be paid for twice.

You ask an agent to run the tests. It spends 800 tokens reasoning about how to run tests: checks the language, infers the framework, constructs the command, and arrives at npm test.

You ask again tomorrow. It does the same thing. 800 more tokens. Same answer.

You are paying for cognition that produces no new information. The answer was already known. It just was not captured anywhere a machine could find without thinking.

This is the problem ephapax solves.

Agents are general-purpose reasoners. They handle what they have never seen before very well: uncertainty, ambiguity, novel problems, edge cases.

The mistake is using that same machinery for problems that have already been solved. Once something is understood and repeatable, an agent reasoning through it again produces nothing new. It just burns tokens to reach the same answer.

Agents should handle uncertainty, exploration, and exceptions. Once a behavior becomes understood and repeatable, it should be captured in deterministic code, API/CLI tools, or workflows — not reasoned through again from scratch.

The goal is not to eliminate agents. It is to stop paying for the same cognition over and over again. Tokens should fund discovery, not rent-seeking on already-solved problems.

CHECK -> EXPLORE -> CODIFY

CHECK. Before reasoning about anything, look for a deterministic path that already exists. A Makefile target. A CLI command. An existing script. A runbook. If it exists, use it. Stop there.

EXPLORE. If no path exists, reason through it. Declare the scope before you start, and document what you learn before you stop. A finding that is not recorded means the next session pays the same cost again.

CODIFY. Once something is understood, capture it in a form that requires no reasoning to use: a script, a pipeline step, a runbook entry, an OpenCode skill. Verify it works. Register it. It now costs nothing to use again.

This is not about replacing agents with scripts. Agents are essential for exploration, debugging unknown systems, handling exceptions, and designing things that have never been built before. That work justifies every token.

The discipline is about where agents are deployed. On the frontier of what is not yet known, not on ground that already has a Makefile target.

Copy SKILL.md to the skills directory of your agent.

Agents without native skill directories (Aider, for example) can read the prompt with /read <path/to/SKILL.md>. The Markdown prompt works with any LLM.

MIT — see LICENSE.

"He entered once for all into the holy places... thus securing an eternal redemption." — Hebrews 9:12