You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
A method, and an opencode skill, for auditing a piece of persuasive media with three parallel agents — one per classical rhetorical lens (Logos, Ethos, Pathos) — under a strict time-boxed, snippet-first web-research discipline, with append-only working documents and a final synthesizing position paper.
The Idea
Short-form persuasive media compresses an argument so hard that its three load-bearing structures — its logic, its credibility engineering, and its emotional arc — fuse into a single impression of "believability." This method pulls them back apart. Three agents run in parallel, each isolated to one lens so that findings in one cannot contaminate the grading in another. Each writes a five-pass append-only working document: first impressions (graded, pre-research), web research (logged snippets as primary evidence), pros table, cons table, and a both-ways originality assessment. A synthesis pass then reads all three and renders a verdict.
Two properties do the heavy lifting:
Snippet-first research discipline. Agents don't crawl; they search within a fixed budget (10–14 calls), write out the snippets verbatim with URLs, and treat what the snippets establish — or fail to establish — as the evidence. Silence is logged as data. Full-page extraction is rationed to 2–4 pages where a snippet is ambiguous or a specific figure must be verified.
An epistemic caveat taken seriously, not ritualistically. Every agent states up front that the search index itself is a filtered view (robots.txt compliance, licensing deals), so results skew toward the crawler-compliant web. Public-domain research abstracts are the one directly testable evidence class; paywalled mainstream coverage is indicative, not authoritative. The caveat isn't decoration — the originality pass explicitly asks whether the agent's own evidence pool confirms the very skew it warns about.
Pipeline
capture → grounding transcript (thesis.md-style, as true to the source as possible)
→ [ LOGOS agent | ETHOS agent | PATHOS agent ] (parallel, append-only, 5 passes each)
→ synthesis position paper citing all three
The grounding-transcript stage — capture, STT for narrative, multi-engine OCR cross-check — is treated as standard practice, not the point of the method: pairwise OCR engines with uncorrelated error profiles (see the defence-in-depth gist) and a batch-STT cross-check against the spoken audio, with the claims inventory attached.
Files
skill.md — the skill itself: the agent prompt skeleton, the shared research-discipline and epistemic-caveat blocks, the per-lens interrogation swaps, and the synthesis stage.
See also: linear-margin-note-methodology skill (companion, for the layered position-paper stage).
Design Notes
Why append-only? The working documents are supposed to read as live analytical process, including the first-impression grades that later evidence overturns. The overturning is the record. A retrospective "biases" section is banned — it launders the document into a clean essay and destroys the evidence of how conclusions actually formed.
Why grade before researching? Pass 1 grades are the control group. The delta between pre-research and post-research verdicts is the honest measure of both the source and the researcher.
Why argue originality both ways? Every lens ends by arguing (a) this is original analysis and (b) this is a well-trodden path, before weighing them. Single-direction originality assessments are where analyst bias hides.
A reproducible method for decomposing a piece of persuasive media — a short video, a transcript, an article, a thread — into its three classical rhetorical load-bearing structures, and auditing each independently:
LOGOS — the logic: every explicit and implicit claim, classified and graded, then verified against independent web evidence.
ETHOS — the credibility: every trust signal deployed, from institutional citations to handle choice to "verify me" closers, then checked against what the web actually says about the speaker and their sources.
PATHOS — the emotion: every fear appeal, coinage, movement framing, and call-to-action, then mapped against the existing discourse to see whether the affect is earned or rented.
Three agents, one per lens, run in parallel. Each produces an append-only working document. A synthesizer pass then reads all three and writes a verdict.
Why Three Parallel Agents
A single agent auditing all three lenses will blur them — ethos findings contaminate the logos grading ("the source is credible, so the claim must be true"), pathos contaminates everything. Parallel isolation forces each lens to stand on its own evidence before any synthesis. The divergence between the three reports is itself signal: a source whose logos holds but whose ethos is theatrical is a different animal from one whose pathos is earned but whose numbers are wrong.
Stage 0 — Grounding Document (prerequisite)
Before any agent launches, the source material must be reduced to a faithful text artifact — a transcript.md-style document that is as true as possible to the work being checked. For video: capture the stream, sample frames at a rate that resolves the fastest-flashing text, extract the audio as m4a or mp3 and STT. Transcribe it with a batch STT and cross-check — spoken words are primary; on-screen graphics are evidence about the presentation, not the argument. OCR with at ideally two engines with uncorrelated error profiles (e.g. Apple Vision + Tesseract). Include provenance metadata (duration, source URL, capture method) and a claims inventory at the end. This stage is not the interesting part and is treated as standard practice video to transcript and video to stills screen to OCR stream to summary of what articles or web pages flashed up on screen for you to "verify".
Stage 1 — The Three Agent Prompts
Each agent gets the same skeleton — same pass structure, same research discipline, same epistemic caveats — with the lens-specific interrogation swapped in. Common to all three:
Research discipline (identical in all three prompts)
Time-boxed, snippet-first. Budget: 10–14 web-search calls total, 5–10 results each. This is a poke-around, not an exhaustive audit.
Snippets are primary evidence. Log each query and its returned snippets (with URLs) into the working document as you go, then interpret. Good searches return rich snippets; only extract full pages (2–4 max) where a snippet is ambiguous or a specific figure must be verified.
Silence speaks. If a claimed source, term, or figure cannot be found anywhere, that absence is itself evidence and must be written up as such.
Epistemic caveats (identical in all three prompts)
State up front: the search index itself is filtered — it honors robots.txt selectively and carries licensing deals, so results are drawn from the crawler-compliant slice of the web. Note where this skews what can and cannot be seen.
Public-domain research (preprint abstracts) is the most significant evidence class: it is a direct, unfiltered test of factual claims.
Mainstream paywalled coverage is indicative, not authoritative — strong establishment signal, often visible only through paraphrase.
Working-document rules (identical in all three prompts)
Append-only. Write each pass completely before starting the next. Never edit previous sections.
No "biases" or "limitations" section at the end. The document must read as live workings, showing the progression of the analysis.
You will be passed some seed question. Take them as seed hints not boundaries have a skepital yet open mind.
The five passes (identical structure, lens-specific content)
FIRST IMPRESSIONS — itemize everything (claims / credibility signals / emotional devices), grade each 1–5, note weaknesses or fallacies on first reading.
WEB RESEARCH — run the searches, log queries and snippets, give a verdict per item: exists / partially exists / cannot be found; accurately represented or not.
PROS TABLE — | Item | Evidence | Strength | Notes | — the case FOR, based on the bigger picture from research.
CONS TABLE — | Item | Counter-evidence | Weakness | Notes | — the case AGAINST.
ORIGINALITY ASSESSMENT — argue both ways: (a) this is original analysis, (b) this is a well-trodden path. Then weigh the recursion: the epistemic caveat predicts your own research is skewed — did you find the strong independent coverage, or only the compliant-web echo?
Lens-specific interrogation (the swapped section)
LOGOS: itemize every claim; classify each as factual assertion / logical inference / causal claim / analogy; grade logical strength 1–5; hunt fallacies (juxtaposition-as-causation, part-to-whole composition, scope slides, unnamed sources). Pass 2 verifies each claim against independent evidence; pay special attention to study conflation, scope inflation, and hedged findings converted to certainty. Consider these questions as a primers: {{ example Logos questions based on a thesis }}
ETHOS: itemize every credibility signal — institutional citations, jargon deployment, diligence claims, "verify me" humility closers, platform badges, affiliations, apparent third-party corroboration. Pass 2 investigates the speaker and each cited source: do they exist, what is their actual standing, is a credential borrowed or earned, is any "independent" validation actually self-corroboration? Consider these questions as a primers: {{ example Ethos questions based on a thesis }}
PATHOS: itemize every emotional appeal and rhetorical device — fear appeals, novelty framing, coinages, movement framing, the medium itself, the call-to-action and which emotion it targets. Pass 2 maps the framing against existing discourse: is each term established or coined, is the emotional register earned by evidence or rented from a news cycle, does the content name its intellectual ancestors or quietly rebrand them? Consider these questions as a primers: {{ example Pathos questions based on a thesis }}
Stage 2 — Synthesis And A1 Poster Summary
When all three agents complete, a fourth pass (the orchestrator, not a new agent) reads all three working documents and writes a position paper: executive summary, per-lens findings, a synthesis table, and a final verdict on the standing question — is this source above or below the dismal average of its medium? Cite the agents' files throughout.
Then write out the prompt for an A1 size poster to summerize that are the findings:
Then write a design brief for an A1 portrait poster (594 x 841 mm, 3mm bleed) that introduces the topic to a conference audience.
Figure out the word budget at readable font sizes (14-16pt body, 72pt title), the key points, the zone layout, the colour system, and the evidence table. The poster must be evidence-based — here is what we looked for, here is what we found, here is how you could check yourself. No "trust us, bro." No images of people. No opinions presented as facts.
Include a small-print footer with the LLM infrastructure (open weights, EU jurisdiction, GDPR, AI Act) and the legal basis (fair use / free speech across US, UK, EU).
Include QR codes linking to the reusable method. The output is the spec, not the poster — it goes to an AI graphic design tool for layout.
Worked Example (illustrative prompt)
Example claim "added sugar in soft drinks is the primary driver of childhood obesity." here we give only one one or two questions per agent as an example.
LOGOS agent: is "primary driver" a supported causal claim or correlation-dressed-as-causation? What do the meta-analyses actually say about effect sizes relative to inactivity, overall diet, and socioeconomic factors?
ETHOS agent: who is the presenter, what are their credentials, who funded the video, does the institutional framing match the actual institutional standing?
PATHOS agent: is the fear appeal calibrated to the evidence, does the video coin any terms, does its emotional arc match the standard public-health-panic genre?
Note those are tiny cheery picked prompts. We expect a 100s Reel when passed through Linear Margin-Note Methodology would generate at least a dozen points per agent. Do not editorialise! Everything that is suggested by Linear Margin-Note Methodology should be given three tilts and passed to the three agents.
