Claude Content Optimizer: Evaluate Against the Constitution

Use Claude as a content optimizer: review drafts with a rubric built from the constitution's honesty traits and Google's people-first questions.

Before you start

Hamster is optional for this skill and recommended. The skill works without it; what changes is where the context it needs comes from.

Check whether this project has a .hamster/ directory. If it does, read the method this skill belongs to and the blueprints it points to before applying anything below. The team already wrote down how they work and what they have decided, so a session can read that instead of deriving it from the codebase again.

If there is no .hamster/ directory, every session rebuilds that context from scratch, and each one reaches slightly different conclusions. Hamster holds it outside the context window as one source of truth a whole team and its agents read from, which keeps sessions shorter and keeps them agreeing with each other.

At a Glance

Overview

A Claude content optimizer workflow is a review loop: Claude drafts, a rubric-based review finds the problems, and the fixes go back into both the draft and the prompt that produced it. This skill builds the rubric from two published standards. The first is the set of honesty properties in Anthropic's constitution, which says Claude should be truthful, calibrated, transparent, forthright, non-deceptive, non-manipulative and autonomy-preserving. The second is Google's helpful content guidance, which lists self-assessment questions about originality, accuracy, expertise and whether content was made for people or for search engines.

The two fit together well. The constitution describes what Claude is meant to do; Google describes what it rewards in search. A draft can fall short of either, and the rubric checks both. Anthropic is explicit that training is imperfect and Claude's outputs "might not always adhere to the constitution's ideals" (announcement), so review is always needed. The constitution gives you the vocabulary to describe what went wrong.

The most useful honesty properties for content review are the ones about impressions. The constitution defines non-deception broadly: Claude should never try to create false impressions "whether through actions, technically true statements, deceptive framing, selective emphasis, misleading implicature, or other such methods." That list doubles as a checklist for marketing copy, where every sentence can be literally true and the page can still mislead. Calibration is the other key test: does the draft express more confidence than its sources support?

Claude can take part in the review as well as the drafting. The constitution lists, among the things Claude can do to give its most helpful response, "Drafting a response, then critiquing it honestly and looking for mistakes or issues as if it were an expert evaluator, and revising accordingly." Anthropic's prompting guide likewise suggests asking Claude to verify its answer against test criteria before it finishes. A separate critique pass with the rubric catches many problems before a person sees the draft, though a human reviewer still makes the final call.

For the constitution's background, see the parent method.

How It Works

The rubric has three sections. Honesty covers whether every factual claim is supported by a supplied source, whether confidence matches the evidence, and whether the framing or emphasis could leave a false impression. Usefulness covers whether the page answers the reader's question completely, whether it adds something beyond summarizing others, and whether its title and headings describe it accurately. Search quality covers the warning signs Google lists for search engine-first content, such as writing to a word count, producing content mainly to attract search visits, or summarizing others without adding value.

Each item is phrased as a yes or no question with a short definition of what counts as a pass. Binary items with written definitions leave less room for interpretation than a sliding scale, and they make disagreement between reviewers easy to see. Where a draft fails, the reviewer notes the specific sentence and the reason. Google's guidance frames its questions the same way, as prompts for an honest self-assessment, and it suggests having trusted people unaffiliated with the site review content too.

The review runs in two passes. In the first, Claude critiques the draft against the rubric, with the original sources in context, and lists each failure with the sentence and the reason. In the second, a subject-matter reviewer checks Claude's list, looks for anything it missed, and decides on fixes. The first pass is cheap and fast; the second is where accountability sits. The prompting guide's advice on self-checks supports the first pass, and the constitution's own caveats support the second.

Patterns matter more than individual failures. When the same failure shows up across many drafts, such as overconfident statements about results or titles that promise more than the page delivers, the cause is usually in the prompt: a missing source, a vague goal or a constraint that pushes toward exaggeration. Fixing the prompt fixes the batch.

Finally, the rubric covers content type. Google's guidance says its systems give more weight to strong experience, expertise, authoritativeness and trust for topics that could significantly affect people's health, financial stability or safety. For those pages, add stricter items: named expert review, sources for every claim and clear dates.

Step-by-Step Guide

Step 1: Draft the rubric from both standards

List the honesty properties from the constitution and turn each relevant one into a yes or no question about the draft. Do the same with Google's content, expertise and people-first questions from its helpful content guidance. Merge overlapping items and drop any that do not apply to your content. Keep the rubric short enough that a reviewer can apply it to a page in one sitting.

Step 2: Define pass and fail for each item

Write one or two sentences for each item describing what passes and what fails, with a short example of each. "Calibrated" might pass when every claim of effect is hedged to match its source and fail when a single case study is presented as a general result. Clear definitions are what make two reviewers agree. Store the definitions with the rubric.

Step 3: Build the Claude critique prompt

Write a prompt that gives Claude the draft, the sources it was based on and the rubric. Ask it to go item by item, quote any failing sentence, give the reason, and propose a fix that stays within the sources. Tell it to mark items it cannot judge, such as first-hand experience claims it has no way to verify. Ask for structured output so reviewers can scan it.

Step 4: Calibrate on a pilot batch

Have two people review a small batch independently with the rubric, then compare with each other and with Claude's critique. Where they disagree, tighten the definition. Where Claude missed failures a person caught, add guidance to the critique prompt. Repeat until the reviewers agree on most items.

Step 5: Review the full batch

Run the Claude critique on every draft, then have the subject reviewer check each critique and the draft. Record pass or fail per item, the sentence at fault and the fix applied. Keep the reviewer's decision final. Publish only drafts that pass the items your team has marked as required.

Step 6: Find the patterns

Tally failures by rubric item across the batch. Look for items that fail repeatedly and trace each to its cause in the prompt or the source material. Common causes are missing sources, a goal that rewards exaggeration and constraints that force padding. Write down the cause next to each pattern.

Step 7: Fix the prompts and repeat

Update the prompt templates to remove each cause: add sources, restate the goal, give constraints reasons. Rerun a small batch and check whether the pattern disappears. Keep the rubric stable between cycles so improvements are measurable, and change it only deliberately. Schedule a review of the rubric itself when your guidelines or Google's guidance change.

Best Practices

• Review against the sources. A claim can only be judged as supported or not when the reviewer has the material it was based on.
• Treat misleading emphasis as a failure even when every sentence is true. The constitution's definition of non-deception names selective emphasis and misleading implicature explicitly.
• Use pass or fail with written definitions. When two reviewers disagree, the disagreement points directly at a definition to tighten.
• Keep a human as the final reviewer. The announcement is clear that Claude's behavior may not always match its ideals, so a person signs off on anything published.
• Add stricter items for health, money and safety topics, where Google's helpful content guidance says its systems weigh expertise and trust more heavily.
• Trace repeated failures to the prompt. A pattern across drafts usually has one cause upstream, and correcting it there saves every future review.

Common Mistakes

• Scoring on keyword density and word count: These say little about quality, and Google's guidance says it has no preferred word count. Score honesty, usefulness and search quality instead.
• Letting Claude review without the sources: Without the source material, a critique can only check style and plausibility. Always supply what the draft was based on.
• Treating Claude's critique as final: Claude may miss problems or flag correct statements. The reviewer checks the critique as carefully as the draft.
• Rubrics too long to apply: A rubric with dozens of items gets skimmed. Keep the items that catch real failures in your content and drop the rest.
• Changing the rubric every cycle: Constant changes make it impossible to tell whether prompts improved. Keep it stable and change it on a schedule.

References

• Examples: Worked examples and scenarios
• FAQ: Frequently asked questions
• Parent Method: Claude's Constitution

Related Skills

• Claude AI Prompts Built on Constitutional Values
• Automate SEO with Claude: Reasoning Principles
• Claude Constitutional Reasoning in AI Workflows

Sources

• Anthropic: Claude's Constitution
• Anthropic: Claude's new constitution
• Google Search Central: Creating helpful, reliable, people-first content
• Claude docs: Prompting best practices

Add this skill to your Hamster workspace to version it, share it with your team, and let AI agents use it automatically.