Developer
News and Updates
Get Support
Sign in
Get Support
Sign in
DOCUMENTATION
Cloud
Data Center
Resources
Sign in
Sign in
DOCUMENTATION
Cloud
Data Center
Resources
Sign in
Last updated Jul 30, 2026

Run a benchmark

  1. Open the agent desktop app or CLI you configured (Codex or Claude Code).
  2. Using the skill TWG AI Context Benchmark (/twg-ai-context-benchmark), ask the agent to run a benchmark. You can use a suite, a single scenario, or a custom prompt.
  3. A browser window opens to track the run.
Prompt exampleDescription
Use the TWG AI Context Benchmark skill with suite cc.Runs the full default suite (10 workflows).
Use the TWG AI Context Benchmark skill with scenario swe_ooo_catchup and topic {{TOPIC}}.Runs one default scenario. Fill in any {{VARIABLE}} placeholders with real values from your company.
Use the TWG AI Context Benchmark skill with prompt "<your prompt>".Runs one custom prompt.

Default suite scenarios

The cc suite includes these scenarios:

ScenarioPrompt shape
em_pr_based_status_rollupGive me a 14 day leadership briefing for the org reporting to {{LEADER_NAME}}, based only on merged PRs. Focus on what shipped and what needs leadership attention.
swe_ooo_catchupI was out for {{LOOKBACK_DAYS}} days. Help me restart work today, focused on {{TOPIC}}. If this means my active work, infer up to two priority workstreams. Tell me the material changes, decisions, blockers, and conversations that should shape what I do next.
em_incident_reliability_reviewGive me a {{LOOKBACK_DAYS}}-day reliability review for {{PLATFORM_NAME}} and its major services in {{LEADER_NAME}}'s org. Identify recurring reliability risks and where leadership should focus prevention work.
swe_topic_deep_diveI need to get up to speed on {{TOPIC}} as if I am joining tomorrow. Explain the central workstream, its current state, key people, important work and resources, material dependencies, recent changes, and where I should start contributing.
swe_dependency_mapGiven project {{PROJECT_KEY}}, map its material upstream blockers, sibling dependencies, and downstream consumers. Identify the critical path, explain how the dependencies affect delivery, and recommend the next move.
swe_sme_identificationI'm trying to find the right people to talk to about {{TOPIC}}. Identify its main workstreams and the best person to consult for each. Tell me who to contact first and why, and whether there is a confirmed overall owner.
search_rovoFind the most useful current design documents and PRDs for {{TOPIC}} using the workspace context available in this run. Tell me which ones to start with, what each establishes, and note important conflicts or freshness gaps.
pm_launch_readinessPrepare a launch-readiness decision brief for the Atlas project {{PROJECT_NAME}}, using work clearly connected to that project. Assess its readiness and recommend go, conditional go, or no-go, including what would change the decision.
daily_standup_personPrepare a concise standup for {{PERSON_NAME}} on {{PROJECT_NAME}} covering the period since yesterday in {{STANDUP_TIMEZONE}}. Focus on the work, blockers, and decisions that matter today.
search_codeMap the public reverse-dependency footprint of {{TOPIC}} across the connected code sources. Identify repositories that directly use it in production and explain how.

Custom prompt examples:

  • What should I work on today?
  • Catch me up on the latest decisions, blockers, and open work for Project X.
  • Who should I consult about packaging and distribution, and why?

Manual CLI run (optional)

Use this only if you are not using the agent-backed workflow.

Open three terminals and keep them open until the run finishes.

Terminal 1

1
2
benchmark live

A browser window opens and waits for the other two terminals.

Terminal 2

1
2
benchmark run control --agent codex --suite cc --parallelism 1

Use exactly the same selector in both lane commands: one --suite, one --scenario, or one --prompt. Do not combine selectors.

Terminal 3

1
2
benchmark run test --agent codex --suite cc --binary --parallelism 1

If started correctly, the run is trackable in the browser from Terminal 1.

For a custom prompt, pass the exact same --prompt to both lane commands. Example:

1
2
benchmark run control --agent codex --prompt "What should I work on today" --parallelism 1
benchmark run test --agent codex --prompt "What should I work on today" --binary --parallelism 1

Next step

Review and share results

Rate this page: