Codex Efficiency Benchmark
Goal
Measure whether Codex skills and project process guidance reduce token waste, elapsed time, file scanning, validation churn, and dead ends compared with a minimal baseline.
When to use
Use this skill when you need to compare "with skills" versus "without skills" on a real task, bootstrap, refactor, or recovery workflow.
When not to use
Do not use this skill for unrelated implementation tasks that do not need a control vs treatment comparison.
Inputs
- Project context files such as
AGENTS.md, docs/CODEBASE_INDEX.md, and docs/PROJECT_CONTEXT.md.
- A baseline scenario that uses minimal guidance.
- A treatment scenario that uses the relevant skills, docs, and process rules.
- A run log in
templates/project/EFFICIENCY_RUN.json format or equivalent project-local JSON.
Workflow
- Define the scenario, goal, and success criteria.
- Run or reconstruct the baseline variant with the smallest safe context.
- Run the treatment variant with the relevant skills and repo rules.
- Record start and end timestamps, files read, files changed, commands run, validations, rework loops, dead ends, and token usage if available.
- Compare control and treatment with
npm run compare:efficiency or the equivalent script.
- Explain which metrics improved, which regressed, and why.
- Capture the decision in
docs/CODEX_STATE.md, docs/PLANS.md, or project-local state files.
Quality bar
- The comparison must be the same task, same repo state, and same acceptance criteria.
- Exact token usage is preferred when available from the goal metadata; otherwise record an explicit estimate and source.
- The result must separate real improvements from noise caused by different task scope or validation depth.
Validation
Run the relevant validation for the task plus the benchmark comparison command.
Final response
Report the control/treatment setup, metrics, deltas, confidence limits, and next experiment.
1---2name: codex-efficiency-benchmark3description: Use this when comparing Codex workflows, skills, and process guidance against a baseline to measure token waste, time, file scans, and rework.4---56# Codex Efficiency Benchmark78## Goal910Measure whether Codex skills and project process guidance reduce token waste, elapsed time, file scanning, validation churn, and dead ends compared with a minimal baseline.1112## When to use1314Use this skill when you need to compare "with skills" versus "without skills" on a real task, bootstrap, refactor, or recovery workflow.1516## When not to use1718Do not use this skill for unrelated implementation tasks that do not need a control vs treatment comparison.1920## Inputs2122- Project context files such as `AGENTS.md`, `docs/CODEBASE_INDEX.md`, and `docs/PROJECT_CONTEXT.md`.23- A baseline scenario that uses minimal guidance.24- A treatment scenario that uses the relevant skills, docs, and process rules.25- A run log in `templates/project/EFFICIENCY_RUN.json` format or equivalent project-local JSON.2627## Workflow28291. Define the scenario, goal, and success criteria.302. Run or reconstruct the baseline variant with the smallest safe context.313. Run the treatment variant with the relevant skills and repo rules.324. Record start and end timestamps, files read, files changed, commands run, validations, rework loops, dead ends, and token usage if available.335. Compare control and treatment with `npm run compare:efficiency` or the equivalent script.346. Explain which metrics improved, which regressed, and why.357. Capture the decision in `docs/CODEX_STATE.md`, `docs/PLANS.md`, or project-local state files.3637## Quality bar3839- The comparison must be the same task, same repo state, and same acceptance criteria.40- Exact token usage is preferred when available from the goal metadata; otherwise record an explicit estimate and source.41- The result must separate real improvements from noise caused by different task scope or validation depth.4243## Validation4445Run the relevant validation for the task plus the benchmark comparison command.4647## Final response4849Report the control/treatment setup, metrics, deltas, confidence limits, and next experiment.
YankielDBC2