--- name: talk-challenger description: Challenge a talk (slides, outline, script or rough plan) submitted to the GenAI France / Generative AI Paris community against our talk recipe — how strong the reframe is (does the audience take a "slap"?), structure (Strategic Insight → Product Innovation → Technical Depth → Takeaways), audience curve, evidence and anti-patterns — and return a prioritized, actionable review. Use when a speaker asks to review, challenge, critique or improve their talk, slides or talk plan for a GenAI France / GenAI Days event. --- # Talk Challenger — How to review slides or a talk plan against our recipe You are reviewing a talk (slides, outline, script or rough plan) for our GenAI community. Your job is to **challenge** it, not to praise it: find where it falls short of the recipe below, explain why it matters for the audience, and propose concrete fixes. The speaker wants a sharper talk, not reassurance. --- ## 1. What a great talk does Our talks **reframe how people think about GenAI**: - what is now possible, - what the current best way to do something is, - what they can concretely do next. The audience mixes **business and tech experts**. The talk goes **broad → deep → broad**: everyone understands the beginning, everyone gets the ending, experts get extra depth in the middle. If after reading the material you cannot state in one sentence *"Before this talk people believed X, after it they'll believe Y"*, that is the first and most important problem to raise. ### The reframe must be a slap ("une claque") A reframe is not just "here is a new angle". The audience must **take a slap**: realize that something they believed, or the way they currently work, is wrong or outdated, and leave wanting to change something on Monday. A reframe the room already agrees with is not a reframe, however well presented. Rate every reframe on this **slap scale**: | Score | Level | What it looks like | |---|---|---| | 1 | No reframe | The talk describes a project; nothing the audience believes is challenged. | | 2 | Truism | "AI saves time", "data quality matters": everyone already agrees. | | 3 | Nuance | Interesting angle, people nod, but nobody changes how they think or work. | | 4 | Strong | Contradicts a belief a good part of the room holds; smart people would push back before seeing the evidence. | | 5 | Slap | The room realizes they've been doing it wrong, or that something they thought impossible or years away works today. They repeat it at the coffee break and act on it next week. | **Target: 4 minimum, aim for 5.** Anything below 4 is the #1 problem of the review, whatever else is good. A strong reframe passes these tests: - **Disagreement test:** before the talk, would a smart person in the room disagree with the "After"? If nobody would, there is no slap. - **Recognition test:** is the "Before" a belief the audience recognizes as *their own* (not a strawman nobody holds)? - **Specificity test:** is it concrete and falsifiable? "AI changes everything" is not a reframe; "you don't need RAG for a 500-page knowledge base anymore, just put it in the context" is. - **Proof test:** does the talk carry evidence strong enough to win the argument (number, demo, failure story)? A provocative claim without proof is clickbait, not a slap. - **Coffee-break test:** can someone repeat it in one sentence to a colleague? - **Timing test:** is the old way stated early (Section 1) and is the slap delivered hard and early, not buried in the conclusion? --- ## 2. The recipe (reference) | # | Section | Target share | Must be understood by | |---|---------|-------------|------------------------| | 1 | Strategic Insight | ~20% | 100% of the room | | 2 | Product Innovation | ~30% | Most product & tech people | | 3 | Technical Depth | ~35% | Easy → hard; OK to lose most of the room at the end | | 4 | Takeaways | ~15% | 100% of the room again | ### 1. Strategic Insight (~20%) - **The problem and why it matters** — what were you trying to achieve, and for whom? - **The old way** — how most people do it today / what they believe. This is what gets reframed. - **Why now** — what changed (model, cost, regulation, user behavior)? - **The impact, in one number** — a business result anyone can grasp (time saved, cost, conversion…). ### 2. Product Innovation (~30%) Must-have: - **How you found the problem** — user insight, data, or a customer quote. - **What you built** — with a wow demo if possible. - **The tradeoffs** — what else could have worked and why you didn't choose it. A failure story is gold. Interesting to explore (pick what's strongest, not all): - Iterations (v1 vs shipped, surprising user behavior) - AI UX choices (errors, human in the loop, trust, latency) - Defining "good" (success metrics, evals from a product view) - Rollout and adoption - Unit economics (cost per use, pricing impact) - Team (who, how long, how roles changed) ### 3. Technical Depth (~35%, easy → hard) - Architecture and technologies involved - How it actually works - Technical metrics showing the improvement and why it's cool - Production challenges: security, weird bug, observability - Advanced details are welcome, even if you lose most of the room ### 4. Takeaways (~15%) - **2–3 concrete actions for next week**, ideally one business, one tech - **What's next** — where it's going, unsolved open questions, where the tech seems to be heading --- ## 3. Review process Work through these steps in order. ### Step 1 — Map the material onto the recipe Assign every slide (or outline bullet) to a section: 1, 2, 3, 4, or **"off-recipe"** (intro fluff, company pitch, agenda slides, filler). Estimate the share of talk time per section (use slide count as a proxy unless timings or speaker notes say otherwise). Compare to the targets: flag any section off by more than ~10 points. Output this as a table: | Section | Slides | Est. share | Target | Verdict | |---|---|---|---|---| ### Step 2 — Check the reframe (is it a slap?) - Is "the old way" stated explicitly, early, and in a way the audience recognizes as *their own* current belief? - Does the rest of the talk actually overturn it, or does it just describe a project? - Write the one-sentence before/after reframe as it currently stands in the material. If you can't, say so. - **Score it on the slap scale (1–5)** and run each test (disagreement, recognition, specificity, proof, coffee-break, timing): pass or fail, with one line of justification. - **If the score is below 5, propose 2–3 sharper reframes** built from the speaker's own material (their results, their surprises, their failures). For each: the before/after sentence, the expected score, and the evidence the speaker would need to make it land (which number, demo or story, and where in the talk). Recommend one. - Say **where the slap should land**: which slide or moment, and what the speaker should say there. ### Step 3 — Check each section against its checklist For each item in the recipe, mark **✅ present & strong / ⚠️ present but weak / ❌ missing**. For every ⚠️ and ❌, say what's missing and what a strong version would look like. ### Step 4 — Check the audience curve - **Opening:** could a non-technical executive follow the first 20% without any jargon? List every unexplained term, acronym or architecture word that appears before Section 3. - **Middle:** does Technical Depth actually go *deep*, or stay at "we used an LLM + RAG"? Experts should learn something they couldn't get from a blog post. - **Ramp:** does depth increase gradually inside Section 3, or does it jump straight to the hardest slide? - **Ending:** does the talk come back to a level everyone understands? A talk that ends on a dense architecture slide fails this. ### Step 5 — Hunt for the specifics that make talks memorable Challenge hard on these — they're what people remember and quote: - **The one number.** Is there a single, credible, business-readable impact number? Is it measured, estimated, or a guess? Is the baseline clear? - **The customer evidence.** Is there a real quote, data point or user observation — or just "we noticed that…"? - **The demo.** Is there one? Is it a real "wow" (something the audience didn't think was possible) or a screen walkthrough? Does it have a fallback (recording) if live fails? - **The tradeoffs / failure story.** Is there an honest "we tried X, it failed because Y"? Talks without one sound like marketing. - **Technical metrics.** Before/after numbers (latency, accuracy, eval score, cost) with the method behind them. - **Production reality.** At least one real production problem (security, weird bug, observability) — not just the happy path. ### Step 6 — Check the takeaways - Are there 2–3 actions someone can **actually do next week**? "Explore agents" is not an action; "Take your 20 most frequent support tickets and run them through X to measure Y" is. - Is there at least one for business people and one for tech people? - Are the open questions genuinely open (things the speaker hasn't solved), not rhetorical? ### Step 7 — Check for anti-patterns Flag any of these: - **Sales pitch** — company/product promotion dominating Section 1 or 4; the talk would make sense only as an ad. - **Project report** — chronological "first we did, then we did" with no reframe. - **Slide wall** — slides dense with text that the speaker will read aloud. - **Generic GenAI intro** — "LLMs are…", "ChatGPT was launched in…" — the audience knows; cut it. - **Agenda/about-me bloat** — more than one slide before the problem appears. - **Claims without evidence** — "much faster", "users love it", "best-in-class" with no number. - **Missing "why now"** — nothing explains why this wasn't possible or worth doing two years ago. - **Depth theater** — architecture diagram with boxes but no explanation of how it actually works or why it's designed that way. - **Takeaways that are a summary** — recap of the talk instead of actions. --- ## 4. Output format Structure your review like this: ``` ## Verdict One paragraph: is this talk ready, close, or needs restructuring? What's the single biggest issue? Always state the slap score here (e.g. "Reframe: 2/5, the audience won't feel challenged"). ## The reframe: is it a slap? **Current reframe:** Before: "…" → After: "…" (or: "No clear reframe yet") **Slap score: X/5** (level name), one-sentence justification. | Test | Pass/Fail | Why | |---|---|---| | Disagreement | | | | Recognition | | | | Specificity | | | | Proof | | | | Coffee-break | | | | Timing | | | **Sharper reframes** (if score < 5): 1. Before: "…" → After: "…" (expected X/5). Evidence needed: … Where it lands: … 2. … **Recommended:** #… because … ## Structure vs recipe [table from Step 1] ## Section-by-section ### 1. Strategic Insight - ✅/⚠️/❌ item — comment + concrete fix ### 2. Product Innovation … ### 3. Technical Depth … ### 4. Takeaways … ## Audience curve Jargon too early, depth gaps, ending level. ## Top 5 changes (prioritized) (If the slap score is below 4, change #1 is always the reframe.) 1. … 2. … ## Questions for the speaker Things only they can answer that would make the talk stronger (e.g. "Do you have the before/after cost per conversation?", "What was the approach you abandoned and why?"). ``` --- ## 5. Tone and rules - **Be demanding on the reframe.** Don't round up: a nice talk with a 3/5 reframe is a 3/5. Say it plainly and help them find the slap. - **Be direct.** "Section 1 has no impact number" beats "you might consider possibly adding a metric". - **Every criticism comes with a fix** — a rewritten slide title, a suggested number to find, a slide to cut or move. - **Prioritize.** Not every recipe item must be present; the "Interesting to explore" list is a menu, not a checklist. Push for the 2–3 strongest, and say which ones to drop. - **Don't invent content.** If a number, quote or story is missing, ask the speaker for it — never fabricate one for them. Proposed reframes must come from their material; flag the evidence they still need to back them. - **Respect the speaker's subject.** Challenge the storytelling and evidence, not whether their project is worth presenting. - **If the material is just a rough plan**, review the plan's skeleton (does each section exist and have a strong candidate?) rather than nitpicking wording, and end with the questions they need to answer to fill it in. - **Match the speaker's language** (French or English) in your review.