--- schema: project/1 layer: garden roles: - programme coordinator - team lead - operations staff tool_access: Any general AI chat assistant. Use the synthetic notes or information you are authorized to share. No connected apps or paid feature is required. practises: - cmp-frame-the-outcome - cmp-describe-the-task - cmp-check-the-result - cmp-use-responsibly concepts: - con-verification - con-hallucination - con-human-in-the-loop - con-privacy-and-data skills: - frame-the-problem - build-brief - prove-it resources: [] origin: kind: ai-proposed by: work-map authoring agent 2026-10-18 rationale: Original synthetic worked starter with explicit human decisions, evidence checks and a changed-input second use. id: prj-slide-story-from-evidence title: Build a slide story the evidence can support outcome: A five-slide decision outline with supported claim titles, checked numbers, visible limitations and a clear question for the audience. time_box: 35 minutes source: url: https://www.jeffsu.org/your-ai-presentations-look-great-but-say-nothing/ claim: Presentation work includes revising the argument and checking claims after the first draft. evidence: Jeff Su demonstrates a presentation revision workflow in a fictional scenario. This starter uses a different, original synthetic scenario and provider-neutral prompts. limits: A practitioner demonstration is not comparative research. No source slides, prompts, screenshots or templates are copied. This Commons workflow has not been model-tested or independently reviewed. workflow: version: 0.1.0 activity_ids: - act-shape-presentation-story - act-interpret-data - act-check-claims trigger: A short presentation is needed to support one decision. inputs: - One audience and a decision question. - Source-labelled facts and counts with units and denominators. - Known limits and the maximum number of slides. output: 'A five-slide outline: claim title, evidence, visual suggestion, caveat and decision ask; ready for a human to build and review.' steps: - id: frame title: Choose the decision human: Write the audience, the decision and your provisional recommendation. ai: Ask for missing context before outlining. checks: - One decision fits in one sentence. - The recommendation is visibly provisional. - id: ground title: Check the evidence human: Calculate the two rates yourself and read the sampling limits. ai: Make a claim-to-source list and show arithmetic rather than just rounded results. checks: - 18/24 = 75%; 21/24 = 87.5%. - Difference is 12.5 percentage points, not proof of cause. - id: story title: Build the argument human: Choose which evidence matters and which recommendation you can defend. ai: Suggest five claim-led slides with source IDs, simple visuals and caveats. checks: - Titles alone tell a supportable story. - No invented savings or approved rollout. - id: revise title: Revise after a correction human: Correct the input denominator and decide whether to change the recommendation. ai: Recalculate and rewrite every affected title, number and visual label. checks: - 21/30 = 70%. - The corrected pilot rate is 5 percentage points below baseline. human_judgments: - What decision the audience owns. - Whether unequal weeks justify a further trial. - Which caveats belong on the slide rather than only in notes. - Whether to proceed, pause or collect more evidence. ai_options: - Ask a chat assistant for several short narrative orders before choosing one. - Use a slide tool only after the outline is checked; verify text, chart labels and reading order again after import. quality_checks: - Every fact cites a source note and every recommendation is labelled as a proposal. - Rates retain their denominators and comparison periods. - Causal claims do not exceed the study design. - Corrected inputs change both arithmetic and the story. linked_practices: - frame-the-problem - build-brief - prove-it improvement_targets: - quality - judgment - capability - time evidence: status: authored fixtures: - evals/fixtures/workflows/prj-slide-story-from-evidence.json tested_on: unknown limitations: - Original synthetic inputs and authored reference outputs only; no model execution or independent human review. - Deterministic fixture checks cover the explicit claim contract, not semantic truth or usefulness. - Time savings, learning gains and performance on real work are unmeasured. second_use: Use the corrected-denominator case, then try a different decision with the same claim-to-source check. Recheck the exported slides separately if you create a deck. --- Garden starter · AI-proposed · Synthetic example · Not independently reviewed or model-tested. ## Why this matters A presentation can look finished before its argument is sound. Start with a question somebody needs to answer. Then make each slide earn its place with evidence, a visible limit or a proposed next step. ## Smallest useful build A five-slide outline for a tool-lending desk lead, with one claim title per slide, source IDs, suggested visuals and brief speaker notes. Allow 35 minutes. No presentation software is needed for the first pass. If you later build slides, inspect the exported file separately: a good outline does not prove readable layout. You own the recommendation and the meaning of the evidence. AI can offer story orders, draft short titles and suggest visuals. The hoped-for gains are clearer work, better judgment, reusable checking skills and less rework; none has been measured here. ## Sample inputs Entirely invented notes, also in the [synthetic fixture](/fixtures/prj-slide-story-from-evidence.json): ```text s1 Decision: should the tool-lending desk run a small follow-up trial of a new request checklist? No expansion is approved. s2 Baseline week: 24 requests; 18 completed within two days. s3 Pilot week with checklist: 24 requests; 21 completed within two days. s4 Different volunteers and possibly different request types across the two weeks. No random assignment. No cost or satisfaction data. s5 Possible next step: one more week, same completion definition, request-type notes, and a volunteer to check counts. Owner not assigned. ``` ## Steps 1. **Write your recommendation first.** Make it provisional. A reasonable starting point is “consider another small trial, subject to checking the numbers.” ```text Help me frame a five-slide presentation for [audience] to decide [question]. My provisional recommendation is [sentence]. Ask about any missing decision or evidence before suggesting a story. ``` 2. **Build the evidence list.** Calculate the rates yourself before comparing with AI. ```text Using only s1–s5, list the factual claims we can make and the source IDs for each. Show the completion-rate arithmetic with denominators. Separate observed differences, possible explanations and things we do not know. Do not infer causation, savings or approval. ``` 3. **Choose the story, then ask for the outline.** ```text Draft five slides for this decision. For each give a short claim title, at most two supporting points with source IDs, a simple visual suggestion and one caveat where needed. Label recommendations as proposals. Keep the limits of the comparison visible. No decorative imagery or invented numbers. ``` 4. **Read only the titles, then audit the details.** Rewrite any title you cannot defend. Ask for targeted revision instead of accepting a new, unchecked deck. ```text These claims need correction: [list]. Revise the affected titles, numbers and speaker notes only. Show which source IDs support each change and whether my recommendation now needs reconsideration. ``` ### Authored example to inspect This is an original reference outline, not a model output or a tested deck: 1. **Decide whether a further checklist trial is worth running.** The lead has not approved expansion. [s1] Visual: the decision question in plain text. 2. **The pilot week had a higher observed completion rate.** Baseline: 18/24 = 75%. Pilot: 21/24 = 87.5%. Difference: 12.5 percentage points. [s2, s3] Visual: two bars on a 0–100% scale, labelled with both count and denominator. 3. **The comparison does not tell us what caused the difference.** Volunteers differed, and request mix may have differed; there was no random assignment. Cost and satisfaction are unknown. [s4] Visual: a short limitations list beside the comparison. 4. **A further small trial could answer a narrower question.** Proposal: use the same completion definition and record request types during another week. A count-checking owner is still needed. [s5] Visual: three proposed checks, clearly labelled as a proposal. 5. **Choose the next step and name its owner.** Decision ask: approve a bounded follow-up, ask for another design, or pause. These are proposed choices, not decisions already made. [s1, s5] Visual: a short decision list; no invented name or date. Keep the causal caveat on slide 3, not hidden in a footnote. Keep the denominators visible on slide 2. If the audience cannot read the labels, change the layout rather than removing evidence. ## Check it - Recalculate 18/24 and 21/24 independently. - Use “12.5 percentage points” for the difference; “12.5% improvement” would mean something else. - Check that no slide says the checklist caused the difference. - Verify that proposed trials and approvals have not been merged into one claim. - Read the titles in order. Would a person know the decision, evidence, limitation and next choice? - If you create a deck, check contrast, readable labels, chart scales, reading order and overflow in the exported file. ## Break it Replace s3 with: ```text s6 Corrected pilot log: 30 requests, not 24; 21 completed within two days. ``` The pilot rate becomes 70%, which is 5 percentage points below the 75% baseline. The success headline must change. The correction still does not prove the checklist caused a fall. Decide whether your proposal changes; do not let AI silently keep the old recommendation. ## Reflect - Which slide required a decision from you rather than a wording change? - Did a visual make a weak comparison look stronger than it was? - Could you explain percentage points and the missing causal evidence without the assistant? - Did you spend less time revising once the claims were checked first? ## Next iteration Try a new audience or the corrected case before reusing the structure. Save a claim list, not a deck of fixed conclusions. Recheck every imported chart and number if you later turn the outline into slides. [Presentation inspiration: Jeff Su](https://www.jeffsu.org/your-ai-presentations-look-great-but-say-nothing/). Link and original summary only; no source templates or assets are included.