Three Deep Research reports on suppressing AI writing tells and matching a specific author voice. Used as inputs for the blog reviewer agent.

Prompt

Research the perfect prompt to give Claude Opus 4.6 and Sonet so their writing style feels like a casual human. It should identify all of the notable tells. I'm not sure if a rule list (do not do x,y,x), best practice guide (try to do a,b,c), or something else is best. Research both the best way to prompt and the best content for the prompt, then derive your research into a prompt I can use as part of the definition of my Reviewer agent. It's going to be used for my blog posts on kyle.pericak.com (if you check if for style, posts before 2025 are entirely human, 2025 and later are AI-supplemented).

Reports

• ChatGPT Deep Research: AI Slop / Voice Prevention — AI tell taxonomy (20 classes), prompting strategies for Claude 4.6, author voice profiling, and ready-to-use reviewer agent prompts.
• Gemini Deep Research: AI Slop / Voice Prevention — Synthetic indicator taxonomy (lexical, syntactic, structural), prompt engineering strategy comparison, voice analysis, and reviewer agent prompt.
• Claude Deep Research: AI Slop / Voice Prevention — Tiered AI tell taxonomy, hybrid prompt format research (Saxifrage experiment), voice profile analysis, and production-ready XML reviewer prompt with two-pass workflow.

---

Cross-Source Synthesis

Compare and contrast of findings across all three reports. No outside opinions added. Every point below is derived directly from the source material.

Shared Findings (present in 2+ sources)

• Ban a common AI vocabulary list — all three name nearly the same words: "delve," "tapestry," "landscape," "multifaceted," "nuanced," "robust," "seamless," "vibrant," "leverage," "embark," "unpack," "crucial," "comprehensive," "innovative" (all 3)
• Reduce or eliminate em-dashes — flagged as the single most recognizable punctuation tell of AI writing (all 3)
• Vary sentence length — AI produces metronomically uniform sentences; human writing alternates short punches with longer constructions (all 3)
• Eliminate formal transitions — "Moreover," "Furthermore," "Additionally" used with mechanical regularity are AI markers; use headers and horizontal rules instead (all 3)
• Remove hedging phrases — "It's important to note," "It's worth noting," "Generally speaking" signal abnormal qualifier density (all 3)
• Remove meta-commentary / self-referential language — "Let me explain," "I should note that," "Great question," "Absolutely" (all 3)
• Never open with throat-clearing — "In today's fast-paced world," "In the ever-evolving landscape of" are near-diagnostic of AI (all 3)
• Never write summary/conclusion sections — end when the content ends; don't restate what was already said (all 3)
• Use positive framing in prompts — "do Y" outperforms "don't do X"; negative framing causes worse output in larger models (all 3)
• Use XML tags to structure prompts — separate role, voice profile, rules, examples, and task into tagged sections (all 3)
• Two-pass flag-then-rewrite workflow — flag issues first, then targeted rewrite; better than flag-only or full automatic rewrite (all 3)
• Provide few-shot examples from the author's actual writing to calibrate voice (all 3)
• Code blocks should dominate technical posts — 60%+ code, prose as connective tissue (all 3)
• Assume reader competence — link to docs instead of re-explaining; don't over-scaffold (all 3)
• First person for personal experience only — never editorial "we" or "let's" (all 3)
• Use contractions everywhere (all 3)
• Express frustration, uncertainty, and pragmatic compromise directly — don't diplomatize (all 3)
• Mix informal vocabulary with technical precision — "pain," "sane," "weird," "super tedious" alongside exact tool names (all 3)
• Open with one functional declarative sentence stating the problem or what the post covers (all 3)
• Avoid symmetric/formulaic structure — equal-length paragraphs and predictable organization are AI tells (all 3)
• Don't inflate significance — ban "revolutionary," "game-changing," "plays a crucial role in shaping," "stands as a testament to" (all 3)
• Allow grammatical imperfection — fragments, sentences starting with "And" or "But," bent rules for naturalness (all 3)
• Ban "Not only X, but also Y" and tricolon (rule-of-three) constructions (Claude, Gemini)
• Opus 4.6 is better suited for reviewer/tone tasks than Sonnet (ChatGPT, Gemini)
• Keep prompt length to 1,500-2,500 tokens for best results (ChatGPT, Claude)
• Score drafts on voice match + AI detectability and skip rewriting if the draft already scores well (ChatGPT, Claude)
• Track which flags recur most often and update the suppression list over time (ChatGPT, Claude)
• End posts abruptly when the task is done — "That should do it" rather than a wrap-up (ChatGPT, Claude)
• Dry, understated humor as ironic asides — never forced or performed (Claude, Gemini)

Unique Findings (from one source only)

ChatGPT only

• AI tells should be treated as a weighted bundle of signals, not single smoking guns — one tell alone isn't proof
• Prompt sensitivity is under-addressed in AI detection research; the same model looks much more human under different prompting strategies
• Prefilling and format-forcing tricks are deprecated on Opus 4.6 and Sonnet 4.6
• Adjust temperature OR top_p, not both simultaneously
• Use stop_sequences to halt on specific strings
• Opus can do multi-pass rewriting in a single call due to 128k max output
• Use a rubric + self-check rewrite loop to force the model to notice its own tells
• Run automated detectors (token rank distribution, curvature analysis) when available as a supplementary check
• Ground style analysis in peer-reviewed detection research (GLTR, DetectGPT, watermarking)
• Place longform data near the top and queries at the end of prompts — Anthropic reports ~30% quality improvement on long inputs

Claude only

• An AI-analyzed voice profile outperforms raw few-shot samples by ~20% on embedding similarity (validated by 180-run Saxifrage experiment)
• Write the prompt itself in the target style — the prompt's own style influences output style
• Mark each change with inline HTML comments like <!-- fixed: em_dash --> so the author can review what changed
• Explicit limitation: reviewer catches style only, not content fabrication — factual accuracy is a separate concern
• Explicit limitation: the approach cannot inject experiences the author hasn't had or opinions they don't hold
• Keep the suppression list to 6-8 hard constraints max — long laundry lists are counterproductive

Gemini only

• Use Extended Thinking for a "Tonal Audit" pass — have Claude diagnose AI tells in a thinking block before editing
• Ask the model to explain its editorial reasoning in the thinking block to ensure active voice processing rather than superficial filtering
• Opus is particularly good at compression (2,000 words to 1,200 while preserving human voice)
• Sonnet tends toward "softening" conclusions to be more pro-social, potentially stripping opinionated edge
• Smaller models are more likely to box themselves into examples without generalizing
• Adopt a "builder" persona — sound like someone in a datacenter explaining to a peer, not an assistant
• Suggest internal/external links smoothly for SEO rather than forced keyword stuffing

Contradictions (points where sources disagree)

• Em-dash policy — ChatGPT says never use em-dashes (zero tolerance). Claude calls them the "single most discussed punctuation tell" but discusses reduction, not elimination. Gemini prescribes a specific 80% reduction, keeping some for emphasis. The spectrum runs from total ban to controlled use.
• Number of example passages — ChatGPT recommends 3-5 before/after pairs. Claude says 1-2 brief excerpts are sufficient for final calibration. Gemini says 3-5 high-quality examples to create "prosodic memory." Claude is the outlier, favoring a lean approach.
• Suppression list size — Claude explicitly caps it at 6-8 hard constraints and warns that long rule lists are counterproductive. ChatGPT and Gemini both present substantially longer suppression lists (15-20+ items) without noting diminishing returns.
• Full rewrite risk — Claude explicitly warns that full automatic rewrite is risky because it can introduce new tells and removes human oversight. ChatGPT recommends an aggressive two-pass rewrite with less emphasis on the risk. Gemini sits in the middle, focusing on flag-then-fix but encouraging Extended Thinking as a safeguard.
• Sonnet's limitations — Gemini warns Sonnet "softens" conclusions and strips opinionated edge, framing it as a substantive voice concern. ChatGPT notes Sonnet needs tighter style guides and recommends one-pass rewrites to avoid verbosity creep, framing it as a practical constraint. Claude doesn't call out Sonnet-specific limitations.
• Voice profile approach — Claude asserts that an AI-analyzed voice profile outperforms raw samples by ~20% and recommends profile-first. ChatGPT and Gemini both emphasize direct few-shot examples as the primary calibration mechanism, with voice characteristics as supplementary.