Lesson 6.86.8 The same review on today's tools

In everyday life. Two shops both give receipts, but neither tells you whether the price matched the shelf label. You check that yourself.

Anthropic makes Claude, and OpenAI makes ChatGPT. The Review Contract describes the work, not the tool, so it works on either AI vendor without changes.

The comparison. Line by line, the AI vendors match closely. Both AI vendors take a detailed request, but only OpenAI names review criteria for it. Both show progress, let you approve important actions, cite web sources when they search, and warn that output can be wrong. One gap, on both AI vendors, changes how you write the contract. Neither AI vendor's help pages promise a citation to a row in your own file. So ask for row-level evidence in the brief, and verify it: a reason, a policy section and a source on every row of the output. The records are similar: the help pages describe progress and results kept with the session or conversation, and no separate audit log. Either way, the record shows what the worker did. It does not show whether the work was right.

Open source changes where the record is kept, not what it proves. Omnara logs every step, in your own database if you host it, the kind of log 6.7 checks a summary against.¹

Figure 6.7. The same review on both AI vendors, as verified 6 October 2026 and updated 4 November 2026. The gold box is the line to write into every contract.

What stays the same. The checks, the decision and the signature stay with you on either AI vendor or an open-source tool. The product supplies the record, and citations when it searches the web. It never supplies the final decision.

For reference.

The three boxes below follow the same six lines in the same order. A number such as (6.5) points to the part of this chapter that the line matches. This page's quiz asks about some of these lines.

Anthropic, as verified 6 October 2026.

• Where the checks go (6.1). In the brief. Anthropic's help center lists the outcome, format and inputs to include.² A Claude conversation becomes a task when the work needs it. Anthropic's guide to these tasks, which it calls Cowork, says to review Claude's approach before you let it run.³ Neither page names review criteria, so this book's method puts them in the brief.
• Watching the work (6.7). Progress indicators show what Claude is doing at each step. You can open the same session in another app, such as the phone app, to watch it, answer Claude's questions or redirect the work.³
• The record (6.7). The session keeps the task's progress and its finished output. You can start a task in one app, guide it from another and get the output wherever you are.³
• Citations (6.5). Responses that use web search include citations.⁴ Anthropic says to review the cited sources, because the originals may have context the summary left out.⁵
• High-risk work (6.1). Some work has real consequences, such as money, messages sent in your name or important files. For that work, Anthropic says to stay close and review what Claude does, or switch from automatic approval back to manual approval, where Claude asks before each action.³
• The AI vendor's own warning (6.3). Claude can produce responses that are incorrect or misleading, including quotes that look trustworthy but are not based on fact. Do not rely on it as your only source of truth for high-risk advice.⁵

OpenAI, as verified 6 October 2026.

• Where the checks go (6.1). In the request. OpenAI's guide to Work, its name for longer, multi-step work in ChatGPT, says to add any files, constraints and review criteria ChatGPT should use.⁶
• Watching the work (6.7). In a longer task, you can review progress, answer questions, change direction and approve important actions.⁶
• The record (6.7). You review the result in the same chat, and you ask for changes or more work there.⁶
• Citations (6.5). Responses that use web search may include citations, and a Sources view when it is available. OpenAI says citations can be incomplete, outdated or incorrect. Open a cited source to check that it supports the answer, and check its date.⁷
• High-risk work (6.1). Approve important actions as the work runs.⁶ OpenAI's prompting guide adds: require your approval before ChatGPT sends, publishes or changes information others rely on.⁸
• The AI vendor's own warning (6.3). ChatGPT can produce incorrect or misleading output, and it can sound confident when it is wrong. Check important information against reliable sources, and use ChatGPT as a first draft, not a final source.⁹

Open source, as verified 23 October 2026.

• Where the checks go (6.1). In the brief, as on any capable worker.
• Watching the work (6.7). Omnara is a runtime for AI Workers that you can host yourself. Its agents work in Slack and Discord, where they reply, ask questions and ask for approvals.¹
• The record (6.7). Omnara saves every step of every agent, so an agent can carry on after a crash or restart. If you host it yourself, that record is in your own database.¹
• Citations (6.5). Ask for evidence on every row in the brief, as above. The tool pages this section cites describe no citation feature.
• High-risk work (6.1). Omnara sets each tool to always allow, always ask or always deny. Its built-in tools and your own start at always allow.¹⁰ In OpenCode, every action starts allowed, except reading a file of sensitive settings. A rule can change any action to ask or deny.¹¹
• The AI vendor's own warning (6.3). The model you choose can still be wrong, whoever makes it, so the checks in 6.3 apply as written.
• Running it safely. Before a run you will not watch, set every tool that can send or change something to ask. Omnara's local settings are insecure on purpose, so a real deployment follows its self-hosting guide.¹