The presentation is deliberately linear. This is the reference layer: how each workflow works, what the AI is uniquely doing, what can fail, how a human validates it, and the files needed to reproduce the examples.
Each package includes a manual mode for a weak internal harness and a fuller tool-connected mode. Every output ends in the same triage: verified, directional, or needs human review.
Interrogate a conviction, retrieve the strongest arguments for and against it, and turn the result into a falsifiable strategy.
Translate a strategy into a clickable mock where experience, platform, value, and controls mature together.
Group recurring causes in work-item prose, link every theme to evidence, and compare the record with an anonymous pulse.
The differentiator is not a longer prompt. It is a purpose-built context system that prepares the right evidence before the model argues with the draft.
What AI reads: each post or transcript in a frozen subscriber-archive snapshot.
What it creates: typed frameworks, claims, contexts, applications, caveats, contrarian positions, and dated source attribution.
What AI compares: positions across the corpus, not just within one episode.
What it creates: cross-corpus synthesis plus explicit pairs of people who disagree. This makes tension a retrieval primitive—not an improvised debate.
What the tool reads: the actual claim, decision, and conditions in the new draft.
What it brings back: a relevant defender, critic, evidence snippets, source/date, and the condition that would change the decision.
What the model must say: whether each claim is verified, directional, or needs human review. Generated dialogue is labeled synthesis—not a quote.
What the leader owns: whether the evidence applies, what changes, and what test or decision date follows.
The expensive work happens once. The portable panel then runs locally over compact artifacts, so a new strategy draft does not trigger a fresh archive-wide model call.
A resumable, lower-cost pass extracts eight typed record classes from each post or transcript. This converts long prose into consistent evidence units without asking one giant prompt to remember the corpus.
A second model compares extracted positions across sources, builds guest profiles and topic synthesis, and produces explicit disagreement pairs. The more capable pass is reserved for the smaller, harder comparison job.
The parent system uses Voyage AI embeddings for meaning-based search. The portable version trades vector search for static, grep-able guest, topic, synthesis, and disagreement indexes—and makes no API call at query time.
chunks.jsonldisagreements.jsonguest_profiles.jsonguest_index.jsontopic_index.jsonsynthesis.jsonSource and license note. This personal prototype uses a frozen Lenny’s Newsletter and Lenny’s Podcast subscriber-archive snapshot. The site publishes derived synthesis only—never raw archive files or substantial source text. Source credit: Lenny Rachitsky. Not affiliated with or endorsed by Lenny’s Newsletter. See Lenny’s Data and the archive’s published license terms.
AI performs high-volume interrogation and synthesis. It does not own truth, authorization, customer impact, people decisions, or the consequences of a strategy.
Each workflow package includes its instruction file, complete or generated sample input, expected output, validation checklist, known failure modes, and both manual and tool-connected operating modes.
Ten-minute path, five-minute fallback, exact talk tracks, click path, recovery language, and likely questions.
All three packages, both standalone HTML examples, the brief, and the presenter guide.
Raw notes, answered questions, evidence trace, expected narrative, skill, QA, and failure modes.
Complete strategy input, offline clickable output, stage acceptance tests, accessibility and external-request QA.
Deterministic 412-ticket fixture, pulse responses, calculation method, expected findings, dashboard, and safety gates.
The story, positioning, governance language, and recommended next pilot in a concise handoff.