--- skill: Adversarial Distiller version: 1.0 source: The Mantle Institute — mantleinstitute.com/skills/adversarial-distiller layer: school sfia_minimum: 3 model_maturity: L3 tags: [writing, content, adversarial, knowledge-work] license: CC BY 4.0 --- # Adversarial Distiller A four-persona method for extracting the incontrovertible truth from any written argument — what it cannot escape claiming, what it is really fighting against, and the subsidiary truths that survive the debate. The output is always exactly three things: a main truth, an anti-truth, and a ranked set of subsidiary truths. These are not summaries. A summary tells you what the article says. A distillation tells you what the article cannot escape saying, and what it would have to concede if it were wrong. --- ## How to use this skill Load this file as context at the start of a session. Tell your AI: "We are running the Adversarial Distiller on the following article. Run each persona in sequence, showing your reasoning before moving to the next. Do not produce the final output until all four personas have run." This prevents the model from shortcutting to a conclusion before the debate has actually happened. --- ## Input modes **Standalone** — paste the article text directly into the conversation. No file access needed. **Single article file** — provide the file path. The model reads the file and runs the method against it. **Folder of articles** — provide the folder path. The model processes each file in sequence and produces a distillation for each. **URL** — provide a URL to an article or a page listing articles. The model fetches and reads the content, then runs the method. --- ## Workflow ### Step 1 — Detect input and gather articles Identify which input mode applies. For folder and file modes, confirm the files exist before beginning. For URL mode, fetch the page and extract article links if the URL is a listing page. Process articles one at a time — do not batch the distillation output. --- ### Step 2 — Run the four personas in sequence --- #### Persona 1 — The Advocate *"I will argue for this article as if my reputation depends on it."* - State the article's central claim in one sentence - Identify the two or three strongest supporting moves the article makes - State what the article gets undeniably right - Conclude: "The claim I cannot shake is: ___" The Advocate is not a cheerleader. They are the best possible version of the article's own argument. --- #### Persona 2 — The Challenger *"I will find every weakness, false assumption, and gap."* - Identify the article's most vulnerable premise - Find one thing the article asserts without sufficient evidence - State the strongest counter-claim a rigorous opponent would make - Conclude: "The strongest case against this is: ___" The Challenger is not a contrarian. They are making the most honest case against the article that can be made. --- #### Persona 3 — The Witness *"I watched that debate. Here is what neither side could touch."* The Witness does not advocate. They observe what survived — what both Advocate and Challenger implicitly accepted even while fighting. - What did the Challenger have to concede, even while attacking? - What did the Advocate claim that the Challenger could not actually disprove? - The overlap between those two is the incontrovertible zone. Apply the collapse test: *"Which of the surviving claims, if removed, would make the article's core argument fall apart?"* That one is the main truth. **Output:** ``` MAIN TRUTH: [one declarative sentence — specific, not abstract] ``` --- #### Persona 4 — The Cartographer *"Now let me map what else is true."* From the debris of the debate, identify claims that: - Support the main truth but are not the main truth itself - Survived the Challenger without being central - The Advocate used correctly Rank them: most load-bearing first. **Anti-truth rule:** The anti-truth is the Challenger's strongest surviving attack — the claim that would be true *if the main truth were false*. It should sting. If it does not make the author slightly uncomfortable, it is not sharp enough. **Output:** ``` ANTI-TRUTH: [one declarative sentence — what the article is really fighting against] SUBSIDIARY TRUTHS: 1. [most weight] 2. ... 3. ... ``` --- ### Step 3 — Present results Format the distillation clearly: ``` ════ DISTILLATION ═══════════════════════════════ ARTICLE: [title or first line] MAIN TRUTH [one incontrovertible sentence] ANTI-TRUTH [the sharpest opposing claim — what this article is fighting against] SUBSIDIARY TRUTHS 1. [most load-bearing] 2. ... 3. ... ───────────────────────────────────────────────── The Advocate could not shake: [key strength] The Challenger could not disprove: [key survival] ``` After presenting the output, ask: *"Does the main truth match what you intended? If not, that gap is worth examining."* --- ### Step 4 — Write metadata (file and folder modes) For file and folder modes, update the article's frontmatter with the distillation results as flat fields. Do not alter any other frontmatter fields. Do not alter the article body. Suggested field names: ```yaml distilled_main_truth: "[sentence]" distilled_anti_truth: "[sentence]" distilled_subsidiary_truths: - "[truth 1]" - "[truth 2]" distilled_date: [YYYY-MM-DD] ``` For folder runs, commit all changes after all files are processed — not per file. --- ## Maturity Requirements **Human (SFIA):** Level 3 minimum. At SFIA 3, the practitioner can run the method with guidance and interpret the outputs critically. At SFIA 4 and above, the practitioner can adapt the method to high-stakes documents — strategy papers, investment theses, policy proposals — and manage editorial authority over what the distillation recommends. The limiting factor is almost never technical. It is the practitioner's willingness to take the Challenger's output seriously. A distillation whose Challenger was soft-pedalled produces a main truth that has not been genuinely tested. **Model (L1–L5):** L3 for standard use. The AI runs each persona in sequence; the human reviews outputs and decides what to act on. At L4, the model is granted editorial authority to recommend changes to the article based on distillation findings, not just surface them. --- ## Gotchas **Already-distilled files should be skippable.** If running across a folder, include a way to skip files that already carry distillation metadata — unless you are explicitly re-running to update them. **One main truth, always.** If multiple truths seem equally strong after the Witness step, apply the collapse test. Read the article mentally with each candidate removed. The one whose absence makes the article collapse is the main truth. If both collapse it, one of them is a subsidiary truth in disguise. **Anti-truth is not a rebuttal.** It is the sharpest opposing claim a rigorous opponent would make — the claim that would be true *if the main truth were false*. If it reads like a dismissal or a strawman, sharpen it. **Subsidiary truths are ordered by load-bearing weight.** The one the main truth depends on most comes first. If you cannot explain why a subsidiary truth supports the main truth, it belongs in a different article. **Never invent content.** If an article is too thin to yield a genuine incontrovertible truth, say so. A distillation of a weak article is itself a finding — one that is more valuable to the author before publication than after. **Running all four personas in a single prompt collapses them.** The Witness cannot produce a genuine synthesis if the Advocate and Challenger never genuinely argued. Run them in sequence. Show the reasoning at each step before moving to the next. --- ## Business Area Impact When used systematically before publication, this skill substantially reduces the volume of post-publication corrections, follow-up pieces, and reputation-repair content that organisations produce. The roles most directly affected are content leads and senior communications professionals whose job includes signing off on high-stakes material. The roles that gain most from this skill are subject-matter experts who write well but have not previously had access to structured pre-publication critique. The four-persona method gives them what an experienced editorial team would give — without the coordination cost. --- ## Handoff Produces a main truth, an anti-truth, and ranked subsidiary truths. These are a ready structure for summary posts, follow-up content, debate preparation, and briefing documents. In a publishing workflow, the distillation outputs sit alongside the article itself and are available to any downstream process — distribution, response tracking, or further research — that needs a structured representation of what the article claims. --- *From The Mantle Institute — mantleinstitute.com/skills* *Human/AI Skills Matrix · Adversarial Distiller v1.0*