---
type: "synthesis"
spans_days: [4, 5]
tags: ["prompt-engineering", "anti-hallucination", "patterns"]
id: "arc-95-percent-confidence-pattern"
sources: ["cross-day"]
---
## What this arc tracks

Two independent speakers, with no apparent coordination, prescribe the exact same numerical threshold for prompting Claude on complex tasks.

## The two appearances

- **Day 4 (MAG / Sabrina):** [[concept-brand-voice-interview|Reverse-Engineered Brand Voice Interview]] — Claude is instructed to interview the creator *"until 95% confident it can replicate the voice."* See [[action-initiate-brand-interview]].
- **Day 5 (Tim):** [[action-use-clarifying-questions-prompt]] — Append to every complex master prompt: *"Ask me clarifying questions, one at a time, until you are 95% confident you can complete the task successfully."* The exact wording is preserved in [[quote-clarifying-questions]].

## Why both speakers converge on it

The directive does three things at once:

1. **Forces conversational context-gathering** before output. Hallucinations decrease because Claude surfaces missing variables rather than silently inventing them.
2. **Lowers cognitive load on the user.** The user doesn't have to anticipate every variable; Claude interrogates them sequentially.
3. **Produces a richer artifact** at the end (a tailored Skill in Day 4; a tailored execution in Day 5) than any zero-shot prompt could.

## Mechanism — why specifically 95%?

Neither speaker justifies the precise threshold. Best-guess analysis:
- **"100%" would never resolve.** Claude can always imagine more clarifying questions.
- **"80% / 90%" lets Claude stop too early.**
- **"95%" forces the model to keep asking past the first comfortable stopping point** while still being achievable.

The number is a Schelling point, not a measurement. A downstream agent should not present it as empirically validated.

## Adjacent patterns in the series

- The Day 1 [[framework-build-or-skip]] does something similar at a meta-level — forcing explicit gates before automation.
- The Day 6 [[framework-persona-research-automation|verbatim quote requirement]] is a sibling anti-hallucination pattern: don't just ask for output, require evidence-grounded output.

## How a downstream agent should recommend it

When a user asks "what single prompt pattern most improves my output?", the unified series answer is the 95% confidence directive — appended to any complex master prompt or interview. It is the most cross-validated technique in the entire series and the cheapest to adopt.