待翻譯：How to Use Harness Engineering to Remove AI Tells

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Petar Djukic, PhD Aug 17, 2026 The initial draft was produced by an AI. What you read now is roughly the fifteenth pass. Between that first version and this one I stripped out the AI tells that wear out a reader. Below I walk you through that process. The skill here is building an agentic harness for AI tells. Setting up the flag, rewrite, recheck loop that expunges machine patterns. Everyone knows the most common tells, and as the harness clears them, your reading can catch even the rare ones. I write Mesh Intelligence so you have a leg up. It’s free, and it stays that way. I compose prose in a sort of vibe-code way. While writing, I never edit the document directly; I stay in chat, tell the agent what the article should say, it rewrites the file, I read the output, and I guide again. Guide, read, guide. I run the harness described below to make my part easier. Without it, I get tired reading lifeless AI prose. This article walks you through the mechanics of my writing harness: which patterns the scripts flag, what the rewrite pass does with them, and why the recheck has to run each time. Read it because once the harness runs, the common tells no longer reach you, and you can spend your time on the substance of your ideas. The term ‘harness’ in the title is borrowed. Geoffrey Huntley arrived at it first, presenting the Ralph loop, a coding agent enclosed in a simple bash loop that repeatedly feeds it the same prompt until the output satisfies its checks [10]. Later, Mitchell Hashimoto christened the practice ‘harness engineering.’ Whenever the agent errs, you let it reshape its work so the error goes away [9]. Neither approach depends on a more intelligent model; the intelligence gaps are closed by the harness’s iterative loop. The loop applies deterministic checks to define completion, signals the model about missing pieces, and the model delivers the next attempt. Figure 1. The harness, in its simplest form. You set the goal and the pass bar. The model writes the next attempt and deterministic scripts grade it. A failure goes back with the line and the reason. The loop ends when the counts hit zero, or when a count stalls and the call returns to you. Coding agents know this shape as the Ralph loop. Here it runs on drafts. If you came for a way to pass the AI check, this is the wrong article; there are guides for that, and some reach for the very fixes below (kill the em-dashes, cut “delve”) [7]. However, that is a losing race. AI detectors are trained on detecting something else. My point here is that if you are going to use AI for writing, focus on the reader. They will read content if it is good. The writing harness exists to yield prose a reader actually wants to read, not merely to slip past an AI detector. Though the two aims sound similar, they are separate tasks. Pangram, the Substack tool, records its approach. It supplies human prose, raw model output, and “synthetic mirrors” (AI-generated text that mimics a specific human passage) to the classifier, teaching it to recognize the precise cases where a machine authored or rewrote the text [8]. Pangram does not judge a sentence’s readability; it only flags machine origin. I ran the full cleanup on an article below. The harness excised every AI tell it detected, yet the detector’s score remained unchanged. One author, noting the same tool flip a human piece between machine and human labels, phrased it differently: “A score that shifts when you add or remove a tool is measuring tools, while the person who wrote the sentences stands right here, unmoved” [6]. Readability matters to me, because I am, by definition, the first reader of any text I author with AI. I run the model in a loop, occupying the thinking half. I state the claim I want the article to convey; the model produces a draft; I read it, spot the flaw in the argument, and return it. Geoff Woods calls this using AI as a thinking partner [5]. The loop only works if I can read the draft without being mentally drained by poor form. When the output turns unreadable, my feedback shifts from ideas to texture, and the thinking loop collapses into a copy-editing loop. Figure 2. The thinking-partner loop. You describe what the article should say. The model drafts the prose. You read, redirect, and the ideas improve each round. When the output is unreadable, feedback addresses texture instead of ideas and the loop stalls. Describing the cleanup takes only a moment. Learn the tells, then give the draft three sweeps, one for words, one for shape, one for argument, and continue until no further gain appears. The following sections walk you through it, tells first. A field guide to AI tells (a list that keeps growing) Machine-generated technical prose feels foreign. At first glance the facts align and the reasoning holds, but the phrasing gives it away. Word choice is the first layer. Models reach for a recognizable vocabulary: “leverage,” “robust,” “seamless,” a “landscape” that keeps “evolving,” the obligatory “delve.” They announce that a point is important before making it. They reach for the em-dash the way a nervous speaker reaches for “um.” None of these is wrong alone; stacked, they read as a machine clearing its throat. The pattern is measurable. A Stanford analysis of nearly a million papers found specific words surging after ChatGPT arrived (”intricate,” “pivotal,” “showcasing”), with computer science posting the steepest rise [2]. A separate study of fifteen million biomedical abstracts found the same surge concentrated in style words, and its headline example is the one everyone already jokes about: “delve” [4]. Shape forms the second tier. Sentences line up with almost identical length, one after another, flattening the rhythm. Paragraphs match in size. Each sentence opens with “The.” A single figure the model favors most: a claim followed by a brief, clipped echo that flips it. “The tooling is real. The results are real.” Once you hear that cadence you cannot stop hearing it. The argument’s body language sits at the third layer. Everything falls into groups of three. A single idea is presented and then broken into two or three parts, like a numbered list poised for the eye. Each paragraph ends by signalling what the next one will accomplish. None of them is an error. It is a writer performing thoroughness instead of having a thought. Identify the three layers: lexical, structural, semantic. The first two are purely mechanical: a script can line up the words and gauge their shape. The third does not fit that pattern. To see it you must grasp what the paragraph is trying to convey, and no script can understand that. I processed my entire archive, fifty-one articles, through the detector below, and the recurring artifacts fell into those three layers. The quoted excerpts come from my drafts and from published work. Figure 3. A field guide to the tells. The artifacts that recurred across a fifty-one-article run, grouped by the layer that catches them. The scripts catch the lexical and structural columns; only a reader catches the semantic one. Quoted examples are real, from my own drafts and published pieces. The run uncovered more tell varieties than the earlier paragraphs enumerate. The figure assembles the recurring types, sorted by layer, and presents an example for each. The three passes The tool is a coding agent skill, filter-tells: two inexpensive scripts that catch the mechanical tells, and a model pass for what semantic problems they cannot identify. Invoked as a chat command, it runs the loop itself — it flags, rewrites, reruns the scripts, and checks whether the counts fell. That division of labor is the harness from Figure 1, with the scripts as its deterministic half. If you’ve dealt with a coding agent, this loop will already be familiar. When a test fails, you supply the model with the failure report, identifying the test and the expected output, then it retries, and you keep looping until the suite turns green. The cleanup preserves that structure and swaps only the test. Instead of functional correctness, it imposes a human-like quality of the prose. A tell is just a failing test; the highlighted line and its rationale constitute the failure you return, and green means no line on the page still snags the reader on the machine. Figure 4. A test loop, but for reading. In coding, the model writes, the tests run, and a failure goes back to the model with the reason until the suite is green. The cleanup is the same loop with a different test, one that asks whether the draft reads like a person, a flagged tell is a failing test, and the loop repeats until nothing is left to flag — or until a count stalls and a person takes over. I’ll walk a real cleanup through the loop, using my article on coding loops [1] as an example input. Initially this one came out of the model accurate, but packed with AI tells. Figure 5. Three passes, cheapest first. A generated draft moves through the lexical grep, the structural metrics, and the semantic read. Flagged passages are rewritten and revalidated against the scripts until the draft stops improving; then a person reads it, because voice is a judgment. Pass one is a shell script, a grep over the lexical tells. It carries the banned words, the cliché phrases, the false-emphasis adverbs, the canned transitions, and prints every match with a line number. The list varies by venue, because a tell is relative to a room. The academic lexicon forgives “critical,” which is ordinary in a paper, and hunts padded phrasing instead; the newsletter list is the strictest. No model, no cost, instant. Its output is bare. Each line gives the file, the line number, the flagged token, and why: loop.md:23 banned leverage loop.md:23 false-emph it is important to note that loop.md:57 cliche at the end of the day On the loop article it flagged three banned words, and each is a one-for-one fix that needs no judgment: “leverage” becomes “use,” “robust” becomes the specific thing you meant, and the “it is important to note that” in front of a sentence simply comes out, because the point was already in the next clause. That is why this pass runs first and free. Run it on the vocabulary paragraph in the field guide above and it flags every word I quoted, which is the point. Cheap also catches the most embarrassing class of defect. One of my published articles ended, for months, with a line no reader was meant to see. It read, “Ready for Substack. Want me to create a suggested author bio or any other supporting materials?” The assistant’s sign-off, pasted along with the article, published under my name. A one-line grep would have caught it. Nobody ran one. Pass two runs as a Python script that measures textual form. It reports sentence-length variance, the range of opening constructions, the count of dashes and colons per five hundred words, and it records flip-cadence pairs directly, before assigning a label of clean, minor, suspicious, or likely AI. What matters are the figures, not the tag. When variance drops beneath a threshold, the rhythm becomes overly uniform; when opening diversity falls under seventy percent, too many sentences share the same starter. Commercial detectors call this burstiness, the irregularity human writing has and machine-generated text lacks [3]. Applied to the Loop article, the pass logged six flip-cadence pairs and em-dashes at nearly three times the comfortable frequency, and returned a likely-AI verdict. Two remedies follow. First the flip pair, a claim and its trimmed echo: The tooling is real. The results are real. which becomes a sentence that says one thing and moves on: The tooling works and the output ships. Then a trailing echo, two back-to-back sentences hitting the identical pulse: …so you can see why each one is there. Read the next section and you will want to break each one first. where the fix rebuilds one ending so the rhym [truncated for AI cost control]