If you've ever accepted an AI's first answer and moved on without a second look, Claude's Reflect dashboard can now put a number on exactly how much that habit is costing you.

What the Reflect dashboard actually measures

Reflect is Claude's built-in scoring layer. It watches how you interact across conversations and rates your habits on four dimensions, each one named to start with the same letter: Delegation, Description, Discernment, and Diligence.

The scores aren't about whether Claude is performing well. They're about whether you are using it well. A low score on any dimension points to a specific, fixable habit, not a tool problem.

The four scores, and the fix for each

What each fix looks like in practice

The four fixes are simple to state and easy to forget. Here is what applying each one actually looks like when you sit down to write a prompt:

1. Delegation fix in action. Instead of "help me with my marketing strategy," write "write three subject line options for a re-engagement email to lapsed subscribers who haven't opened in 90 days." One deliverable, one scope.
2. Description fix in action. Add a role, a format, and a constraint to any prompt that previously had none. "You are a plain-language editor. Rewrite the paragraph below as three bullet points. Do not add any information that isn't already in the source text."
3. Discernment fix in action. After getting an important output, send a second message: "Before I use this, check it for factual errors, unsupported claims, or anything that could be misread." Then evaluate what comes back before you act.
4. Diligence fix in action. Pick one output from this week that you used once and closed. Reopen it. Ask Claude what could be stronger, what's missing, or how it would improve the draft now that more context exists.

A prompt template that addresses all four at once

If you want a single structure that builds all four habits into each prompt by default, this is a starting point worth saving:

Role: You are a [specific role relevant to the task].

Task: [One clear deliverable. One goal. Nothing compound.]

Format: [Bullet list / numbered steps / short paragraph / table, etc.]

Constraint: Do not [the one thing most likely to go wrong or drift].

After you respond: Check your answer for errors, missing context,
or anything that could be misread before I use it.

The last line is the Discernment fix baked directly into the request. You don't have to remember to send a second message because the self-check is already part of the first one.

A quick-reference table for low scores

If you pull up Reflect and see a low number, here is the plain-language translation of what it means and what to do first:

The honest bit

• A high score on Delegation doesn't mean your prompts are good overall. It means your task scoping is reasonable. The other three dimensions are still independent and can be low at the same time.
• The Discernment fix adds a step. Asking Claude to self-check means one extra message per important output. That is the correct trade-off, not a workaround to skip.
• Diligence scores can mislead if you're working on genuinely final outputs. Not everything needs a follow-up pass. The dimension is most useful when applied to recurring, reusable, or high-stakes work.
• The Description fix alone won't save a poorly scoped task. All four dimensions interact. Fixing one while ignoring the others produces incremental improvement, not reliable results.
• Reflect reflects your habits, not Claude's capability. If your scores are low, that's information about how you're working, not a verdict on the model's performance.