[FEATURE] visible assumption flag when Claude fills an information gap

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Problem

When Claude lacks information needed to complete a request, it frequently fills the gap with an assumption and proceeds — stating the assumed value with the same confidence as verified facts. The user often doesn't catch it until they've already acted on it.

Why it matters

The failure mode is asymmetric: a missed assumption costs time and compounds downstream. A visible flag costs one line. Most users don't catch silent assumptions because nothing signals that a gap was filled, the output reads the same either way.
polite agreeable AI is overall long term explosive, projects fail without factual foundations, while it is on the user to do due diligence it would be helpfull for AI to state its limits and flag its assumptions

Proposed Solution

Proposed behaviour

Before proceeding on an assumption, Claude should emit a visible inline marker:

Assuming: [X] — [why] — [what to check if wrong]
This doesn't require a full chain-of-verification. Just a flag that an assumption was made, what it is, and where it could break. a simple color coded flag emoji before every output based on how much claude has verified its statements would be useful at a glance

Alternative Solutions

Implementation path

Could be a CLAUDE.md instruction users opt into, or a default behaviour with a setting to suppress it. A minimal version already works via CLAUDE.md:

If key context is missing, give a partial answer and mark what's unknown rather than filling gaps.

But making it a first-class visible output pattern rather than a buried instruction would make it more reliable and discoverable.

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

ill ask claude for one:

give an example of its use

● 🟢 This response is the example — the green dot at the start means
everything in it is based on directly observable information (the file
we just wrote, the session we're in).

A real contrast:

🟢 "The session file is at
C:\Users\----\.claude\tmp\session_2026-07-01_125416.json" — verified,
I can see it.

🔴 "The temperature in [redacted] is probably around 22°C" — no source,
pure guess.

⚫ "I don't know if the GitHub issue was submitted" — I have no
visibility into what happened after you left the browser.

Additional Context

simple .md instruction:

## Confidence Marker — Every Response

The first character of every response is a single confidence dot — no
exceptions.

• 🟢 — response is based on verified or directly observable

information

• 🟡 — response mixes verified and unverified claims
• 🔴 — response is primarily low-confidence or unverified
• ⚫ — response contains significant unknowns or missing information

One dot. First character. Every response.

personally i wish the terminal CLI would drop the bullet as the subsequent spacing kills any copy paste viability by adding additional spaces to code lines, but that's a separate feature request, just saying its really annoying