Codex spec prompting
Reference recipes and antipatterns for writing precise Codex coding-agent specs. Use when refining, reviewing, or hand-writing a spec for Codex or another coding agent, or when judging whether a refined spec is good enough to run.From its SKILL.md
npx -y skills add saadjs/codex-advisor --skill codex-spec-promptingAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.0 KB, as published. Nobody here has run it
Codex Spec Prompting
Guidance for turning a rough request into a precise, runnable spec for a coding agent. This is the same standard the Codex Advisor bridge applies when it rewrites a request — use it when you write or review a spec yourself, or when deciding whether a refined spec is ready to execute.
A good spec has six parts
- Goal — one concrete sentence. What changes and why, not a restatement of the request.
- Scope — what is in scope, and explicitly what is out. This is where you stop scope creep.
- Assumptions — every inferred default, each prefixed with
ASSUMPTION:. Only list one when resolving it would change correctness, scope, safety, or verification. - Requirements — verifiable, imperative statements. "Retry failed uploads up to 3 times with exponential backoff", not "make uploads more reliable".
- Verification — the exact commands, tests, or checks that confirm the work.
- Acceptance Criteria — observable, checkable completion conditions.
How to apply it
- Name exact files, symbols, and commands only when they are given or present in repository context. Otherwise mark them
ASSUMPTION:rather than inventing them. - Prefer one precise requirement over three vague ones.
- Make every acceptance criterion something you could check: a command to run, a test to pass, or an observable state to inspect.
- Keep the spec tightly scoped. No speculative refactors, renames, dependency bumps, or cleanup unless they are required for correctness.
References
references/codex-spec-recipes.md— worked structures and before/after examples.references/codex-spec-antipatterns.md— common failure modes and how to fix them.