bootstrap-realtime-eval

A workflow for starting a new realtime evaluation in a cookbook repository. A realtime evaluation tests an AI system during live audio or interactive runs using a shared crawl, walk, or run harness.

In plain words
What is it for?
Use it to scaffold an evaluation folder, define its prompts, tools, data, and graders, write a README, and validate the setup with smoke tests and full runs.
Why use it?
It helps create the required files and choose the right test harness without copying shared harness code into the new evaluation.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/openai/openai-cookbook/bootstrap-realtime-eval
Any agent
npx skills add openai/openai-cookbook --skill bootstrap-realtime-eval
Clone the repo
git clone --depth 1 https://github.com/openai/openai-cookbook

Made for: Claude Code, Codex.

Per session 76 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,780 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00076 $0.01780
Opus 5 $0.00038 $0.00890
Sonnet 5 $0.00015 $0.00356
Haiku 4.5 $0.00008 $0.00178

Measured yesterday against content hash c7861252c75e, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

bootstrap-realtime-eval scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

The scan reads SKILL.md. This mod also ships 1 executable file (scripts/bootstrap_realtime_eval.py), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

examples/evals/realtime_evals/skills/bootstrap-realtime-eval/SKILL.md · 134 lines

How it starts

The opening of the file, as written. The whole thing — 134 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Bootstrap Realtime Eval

Use this skill when the user wants a new realtime eval scaffold under examples/evals/realtime_evals/.

This skill is repo-specific. Do not copy harness code into the generated folder. The generated eval should point at the shared harnesses already in:

  • examples/evals/realtime_evals/crawl_harness
  • examples/evals/realtime_evals/walk_harness
  • examples/evals/realtime_evals/run_harness

Inputs To Collect

Always ask the user for the minimum set needed to choose and scaffold the eval before you create files, run the scaffold script, or author starter data. Do not skip this just because you can infer a default.

Ask for:

  • Eval name
  • Goal or scenario
  • Harness choice, or enough context to recommend one
  • System prompt path or inline text
  • Tools JSON path or tool descriptions
  • Data path or source materials
  • Desired graders

If the user does not know which harness they want, explain the options briefly and recommend one. See references/harness-selection.md.

When the user asks for synthetic audio but does not specify a harness, default to crawl text-to-TTS unless they need the generated audio to carry particular noise, telephony artifacts, speaker characteristics, or other replay-specific properties. Use walk for those cases.

Keep the questions concise and grouped into one short batch whenever possible.

If the user only provides user_text or a short task description, still ask the questions above first. If they answer only partially, then infer the remaining low-risk details, call out the assumptions, and make the scaffold easy to revise later.

Workflow

  1. Ask the user for the required inputs first.

    • Do this before making files or selecting a final harness.
    • If the user already supplied some of the inputs, ask only for the missing ones.
    • If you recommend a harness, wait for the user response before scaffolding.
  2. Pick the harness.

    • crawl: single-turn text-to-TTS.
    • walk: replay saved audio or generate audio from text rows.
    • run: multi-turn simulation with tool mocks and judge criteria.
    • If the user wants synthetic audio but does not care about replay-specific audio characteristics, prefer crawl over walk.

Read the full file on GitHub · 134 lines

Files

What ships with it

3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 134 lines · 76 tokens per session scan A c7861252c75e

Subscribe to this mod's changes

bootstrap-realtime-eval is a skill published in the GitHub repository openai/openai-cookbook (75,624 stars, last pushed 3d ago), licensed MIT. It adds 76 tokens to every session and 1,780 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.