← series index

# AI for Researchers Webinar Series — Implementation Plan

> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.

**Goal:** Produce the master curriculum plus complete, deeply-researched content (slides, demo script, handout, sources) for all 7 sessions of the "AI for Researchers" webinar series defined in `curriculum-design.md`.

**Architecture:** Shared HTML slide/handout templates first, then the master curriculum, then sessions built one at a time in delivery order. Each session task is self-contained: deep research → `sources.md` → slides → demo script → handout → citation audit. Ehsan reviews after every session task before the next begins.

**Tech Stack:** Markdown content; self-contained HTML slide decks (printable to PDF via browser, consistent visual identity); WebSearch/Exa/WebFetch for research; Playwright (browser screenshot) for render verification.

## Global Constraints

Copied from `curriculum-design.md` — every task implicitly includes these:

- **Audience:** academic researchers, all fields; baseline = casual AI users. No code assumed.
- **Tone/positioning:** fully tool-neutral; SCiNiTO mentioned only where it naturally fits, never pitched.
- **Session format:** 60 min = ~20 min literacy foundation + ~25 min demo + ~15 min Q&A. Every session opens with a 2-minute recap slide of the series so far.
- **Slide decks:** ~25–30 slides, self-contained HTML (no external CDN dependencies), consistent series visual identity, printable to PDF.
- **Demo scripts:** step-by-step, exact prompts verbatim, expected outcomes, fallback plan (screenshots/pre-run outputs) for live failure.
- **Handouts:** 2–4 pages (Markdown; cap raised from 2–3 by Ehsan's 2026-07-28 ruling — ~2,400 words max): prompt templates, tool comparison table, checklist, further reading. Length must be earned by real content (templates, tables, checklists), not padding.
- **Source standard:** every factual claim in slides/handouts traces to a citable source — peer-reviewed papers, official publisher/funder policy pages, primary documentation, or (per Ehsan's 2026-07-27 sign-off) first-party survey reports from major publishers/scholarly organizations when quoted accurately and labeled on the slide as self-reported survey data. Vendor blogs/listicles may inform tool discovery but are NOT citable grounds for claims.
- **`sources.md` per session:** dated, annotated reference list. Slides carry footnote markers (e.g. `[1]`) for statistics and policy claims, resolved in a references slide and in `sources.md`.
- **Freshness:** every deck/handout carries "Landscape as of July 2026" (or the month of authoring) on its title/footer.
- **No git repo:** `/home/ehsan/reports` is not a git repository — skip all commit steps; completion of each task = files written + verification passed + Ehsan's review.

## Deep-Research Procedure (referenced by every session task)

Each session task's Step 1 runs this procedure with the session-specific queries listed in the task:

1. Run every listed query via WebSearch and/or Exa deep search (`exa:search` skill) — at minimum: 2 peer-reviewed/arXiv queries, 2 primary-documentation queries, 1 recent-news/benchmark query.
2. For each candidate claim that will appear on a slide, fetch the primary source (WebFetch) and confirm the claim wording against it. If a claim can't be traced to a primary source, drop the claim.
3. Record every kept source in `sources.md` using the entry format defined in Task 1.
4. Target: ≥ 15 annotated sources per session, ≥ 5 of them peer-reviewed/preprint literature.

---

### Task 1: Shared templates & visual identity

**Files:**
- Create: `assets/slide-template.html`
- Create: `assets/handout-template.md`
- Create: `assets/sources-template.md`

**Interfaces:**
- Produces: `slide-template.html` — self-contained HTML with inline CSS: title slide, section divider, content slide, two-column slide, table slide, "references" slide, footer with series name + "Landscape as of <Month Year>" + slide number. Slides are `<section class="slide">` elements, one per printed page (`@media print { .slide { page-break-after: always; } }`). All session decks copy this file and fill in content.
- Produces: `handout-template.md` — H1 title, "At a glance" box, "Prompt templates" section, "Tool comparison" table skeleton, "Checklist" section, "Further reading" section, footer line with date stamp.
- Produces: `sources-template.md` — entry format all sessions use:
  `- [n] Author/Org (Year). *Title*. URL — accessed YYYY-MM-DD. [type: peer-reviewed | policy | primary-doc | news] — one-line note on what claim(s) it grounds.`

- [ ] **Step 1: Write `assets/slide-template.html`** — clean academic look: dark-indigo/white palette, large readable type (min 24px body on slides), Georgia/serif headings + system sans body, no external fonts/CDNs. Include one example of each slide type with placeholder text clearly marked `{{...}}`.
- [ ] **Step 2: Render-verify the template** — open via Playwright (`browser_navigate` to `file://` path), screenshot, confirm: no overflow, footer on every slide, print CSS present (check `@media print` in source).
- [ ] **Step 3: Write `assets/handout-template.md` and `assets/sources-template.md`** with the structures above, placeholders marked `{{...}}`.
- [ ] **Step 4: Verify** — templates contain all required sections listed in Interfaces; no unmarked placeholders.

### Task 2: Master curriculum document

**Files:**
- Create: `curriculum.md`

**Interfaces:**
- Consumes: `curriculum-design.md` (sections 1–3).
- Produces: `curriculum.md` — the attendee/promo-facing master document. Later session tasks copy their session's learning objectives from here verbatim onto their objectives slide.

- [ ] **Step 1: Write `curriculum.md`** containing: series title + one-paragraph pitch; audience & prerequisites; format (7 × 60 min); then for EACH of the 7 sessions: title, 2–3 sentence promotional abstract, 4–6 learning objectives (measurable verbs: explain, run, evaluate, apply…), topic outline (from design §3), demo description, and "you will leave with" (the handout artifacts).
- [ ] **Step 2: Verify against spec** — cross-check `curriculum.md` against `curriculum-design.md` §3: every session topic bullet in the design appears in the curriculum outline; no session missing; abstracts are tool-neutral (search the file for promotional SCiNiTO language — mentions allowed, superlatives not).
- [ ] **Step 3: Ehsan review checkpoint** — present the curriculum summary; wait for approval before Task 3.

### Task 3: Session 1 — AI Foundations for Researchers

**Files:**
- Create: `session-1-foundations/sources.md`, `session-1-foundations/slides.html`, `session-1-foundations/demo-script.md`, `session-1-foundations/handout.md`

**Interfaces:**
- Consumes: `assets/*` templates; Session 1 objectives from `curriculum.md`.
- Produces: the series' canonical "capability/failure map" table and "tool landscape taxonomy" diagram — Sessions 2–7 recap slides reference these by name.

- [ ] **Step 1: Deep research** (run the Deep-Research Procedure) with queries: "how large language models work explainer peer-reviewed 2025 2026"; "LLM hallucination rates citations benchmark study"; "reasoning models o-series claude gemini capabilities limitations research tasks"; "generative AI adoption survey researchers 2026 (Nature/Wiley/Ithaka)"; "prompt engineering effectiveness empirical study"; official model documentation (OpenAI, Anthropic, Google) for current model names/capabilities as of July 2026.
- [ ] **Step 2: Write `sources.md`** from template; ≥15 entries, ≥5 peer-reviewed.
- [ ] **Step 3: Write `slides.html`** (~25–30 slides) from template covering design §3/S1: LLM mental model (training→tokens→context; why hallucination is structural); capability/failure map for research tasks (reliable / unreliable / dangerous — as a table slide); 2026 tool taxonomy (chatbots vs research tools vs deep-research agents); prompting basics (context, role, constraints, iteration, evaluation); every statistic footnoted `[n]`; references slide last.
- [ ] **Step 4: Render-verify slides** — Playwright screenshot pass: no overflow, footers present, count slides (25–30).
- [ ] **Step 5: Write `demo-script.md`** — Demo A: same research question ("What does the literature say about <X>?" — pick a real question from Ehsan's domain, e.g. AI in scholarly publishing) asked naively vs. with structured prompt in two tools, side-by-side comparison table of outputs; Demo B: elicit a hallucinated citation live (ask a chatbot for niche references without grounding), then verify in Google Scholar/SCiNiTO to show fabrication. Exact prompts verbatim, expected outcomes, fallback: pre-captured screenshots checklist.
- [ ] **Step 6: Write `handout.md`** — prompt-structure template (CRIT: Context-Role-Instructions-Task or similar, as taught on slides); capability/failure one-page table; "spot a hallucination" checklist; further reading from `sources.md`.
- [ ] **Step 7: Citation audit** — for every number/claim on slides & handout, confirm a `[n]` footnote exists and resolves in `sources.md`; confirm date stamp present on deck + handout.
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 4: Session 2 — Literature Discovery & Synthesis

**Files:**
- Create: `session-2-literature/sources.md`, `slides.html`, `demo-script.md`, `handout.md` (same structure as Task 3's files)

**Interfaces:**
- Consumes: templates; Session 2 objectives from `curriculum.md`; recap slide references Session 1's tool taxonomy.
- Produces: the series' canonical tool-comparison table for discovery tools (reused in Session 6).

- [ ] **Step 1: Deep research** — queries: "semantic search vs keyword search academic literature retrieval evaluation study"; "Elicit Consensus Semantic Scholar Scite coverage accuracy evaluation peer-reviewed"; "deep research agents evaluation citation accuracy DeepTRACE audit"; "systematic review automation AI Cochrane guidance 2026"; "OpenAlex coverage bias study"; primary docs for Elicit, Consensus, Semantic Scholar, Scite, ResearchRabbit, SCiNiTO, and ChatGPT/Gemini/Claude deep research features (current as of July 2026).
- [ ] **Step 2: Write `sources.md`** (≥15 entries, ≥5 peer-reviewed).
- [ ] **Step 3: Write `slides.html`** covering design §3/S2: semantic vs keyword (what each misses); citation-grounded tool tour with comparison table (corpus size, grounding, citation export, cost); deep-research agents — strengths + documented failure modes (cite audit studies); caution zones (systematic reviews, coverage bias, paywalls); footnoted; references slide.
- [ ] **Step 4: Render-verify slides** (Playwright pass as in Task 3 Step 4).
- [ ] **Step 5: Write `demo-script.md`** — one real research question through (a) a citation-grounded discovery tool and (b) a deep-research agent; verification segment: pick 3 citations from the agent report and check each exists and supports the claim. Exact prompts, expected outcomes, fallback screenshots checklist.
- [ ] **Step 6: Write `handout.md`** — discovery-tool comparison table; deep-research prompt template; "verify an AI literature review" checklist; further reading.
- [ ] **Step 7: Citation audit** (as Task 3 Step 7).
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 5: Session 3 — Reading, Notes & Knowledge Management

**Files:**
- Create: `session-3-reading-km/sources.md`, `slides.html`, `demo-script.md`, `handout.md`

**Interfaces:**
- Consumes: templates; Session 3 objectives from `curriculum.md`.
- Produces: the 5-paper demo corpus list (also used in Session 6's cross-field demo if suitable).

- [ ] **Step 1: Deep research** — queries: "LLM summarization accuracy scientific papers evaluation study"; "NotebookLM grounded generation source-bound accuracy"; "structured data extraction from papers LLM systematic review accuracy"; "Zotero AI integration plugins documentation"; "reading comprehension AI summaries flatten nuance study"; primary docs: NotebookLM, SciSpace, Paperpal, SCiNiTO PDF chat, Zotero (July 2026 state).
- [ ] **Step 2: Write `sources.md`** (≥15, ≥5 peer-reviewed).
- [ ] **Step 3: Write `slides.html`** covering design §3/S3: paper Q&A / PDF chat; critical-literacy segment: where summaries flatten nuance (cite evaluation studies); grounded notebooks concept; structured extraction across papers → comparison tables; reference-manager integration; footnoted; references slide.
- [ ] **Step 4: Render-verify slides.**
- [ ] **Step 5: Write `demo-script.md`** — pick 5 open-access papers on one topic (choose during research; list DOIs in script); build a grounded notebook (NotebookLM or equivalent); ask 3 planned questions incl. one the corpus can't answer (show grounded refusal); extract a methods/effect-size comparison table. Exact prompts, expected outcomes, fallbacks.
- [ ] **Step 6: Write `handout.md`** — grounded-notebook setup checklist; extraction prompt templates; "trust but verify a summary" checklist; further reading.
- [ ] **Step 7: Citation audit.**
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 6: Session 4 — Writing, Publishing & Integrity

**Files:**
- Create: `session-4-writing-integrity/sources.md`, `slides.html`, `demo-script.md`, `handout.md`

**Interfaces:**
- Consumes: templates; Session 4 objectives from `curriculum.md`.
- Produces: the publisher-policy comparison table (canonical for the series; referenced by Session 7's medical segment for ICMJE context).

- [ ] **Step 1: Deep research** — queries: "ICMJE recommendations AI 2026 current text"; official policy pages: Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE, Science/AAAS, JAMA, Nature (fetch each, quote exact wording); "journal AI policy adoption study 70% disclosure 0.1%" (find the primary study behind the disclosure-gap statistic — REQUIRED before the stat may appear on a slide); "AI text detector accuracy false positives non-native English speakers study"; "peer review confidentiality generative AI NIH funder policy"; "AI writing assistance non-native English speakers equity study".
- [ ] **Step 2: Write `sources.md`** (≥15; policy pages count as `policy` type; ≥5 peer-reviewed incl. detector-accuracy and disclosure-gap studies).
- [ ] **Step 3: Write `slides.html`** covering design §3/S4: editor-vs-ghostwriter spectrum; legitimate uses incl. language polishing for non-native speakers (cite equity studies); policy landscape table (publisher → policy category → disclosure requirement → exact-source footnote); universal no-AI-authorship rule; disclosure gap stat (only if grounded in Step 1); detector unreliability (cite studies, incl. bias against non-native speakers); peer-review confidentiality + funder rules; footnoted; references slide.
- [ ] **Step 4: Render-verify slides.**
- [ ] **Step 5: Write `demo-script.md`** — Demo A: policy-compliant editing pass on a provided sample paragraph (include the paragraph in the script — write a deliberately rough 150-word methods paragraph, no real author's text), showing edit-suggestion prompting vs rewrite prompting; Demo B: write a disclosure statement for that edit matching one real publisher's required format (quote the requirement). Exact prompts, expected outcomes, fallbacks.
- [ ] **Step 6: Write `handout.md`** — policy quick-reference table (with "check your target journal" caveat + links); disclosure-statement templates (2–3 publisher formats); "before you submit" checklist; further reading.
- [ ] **Step 7: Citation audit** — extra check: every policy row in the table links the publisher's own page, not secondary reporting.
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 7: Session 5 — Data, Code & Your AI Workflow

**Files:**
- Create: `session-5-data-code/sources.md`, `slides.html`, `demo-script.md`, `handout.md`

**Interfaces:**
- Consumes: templates; Session 5 objectives from `curriculum.md`.
- Produces: the "personal AI workflow" capstone canvas (one-page framework; reused as closing frame in Sessions 6 and 7).

- [ ] **Step 1: Deep research** — queries: "LLM code generation accuracy data analysis scientific computing evaluation"; "AI statistical analysis errors reproducibility study"; "ChatGPT Advanced Data Analysis / Claude analysis tool / Julius no-code documentation"; "agentic coding tools research computing 2026"; "reproducibility guardrails AI-generated analysis best practices"; primary docs for 2–3 no-code analysis paths current July 2026.
- [ ] **Step 2: Write `sources.md`** (≥15, ≥5 peer-reviewed).
- [ ] **Step 3: Write `slides.html`** covering design §3/S5: AI for analysis (codegen, stats assistance, viz) with documented error modes; reproducibility guardrails (versions, seeds, verify-every-number, keep-the-code); no-code paths for non-programmers; agentic tools and trajectory; capstone: personal AI workflow canvas mapping Sessions 1–5 tools onto the research lifecycle; staying current (concrete venues/newsletters, cited); footnoted; references slide.
- [ ] **Step 4: Render-verify slides.**
- [ ] **Step 5: Write `demo-script.md`** — pick an open dataset (choose during research; include URL + license in script); upload → AI-assisted exploratory analysis → one statistical test → one chart → verification segment (re-run the produced code, check one number by hand). Exact prompts, expected outcomes, fallbacks.
- [ ] **Step 6: Write `handout.md`** — analysis prompt templates; reproducibility checklist; personal-workflow canvas (fillable); further reading.
- [ ] **Step 7: Citation audit.**
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 8: Session 6 — AI for Multidisciplinary Research

**Files:**
- Create: `session-6-multidisciplinary/sources.md`, `slides.html`, `demo-script.md`, `handout.md`

**Interfaces:**
- Consumes: templates; Session 6 objectives from `curriculum.md`; discovery-tool table from Session 2; workflow canvas from Session 5.
- Produces: the cross-field orientation workflow (jargon translation → landscape map → reading list → synthesis memo) — referenced in Session 7's closing.

- [ ] **Step 1: Deep research** — queries: "interdisciplinary research barriers AI literature review study"; "AI cross-domain knowledge synthesis evaluation"; "citation graph tools connected papers cross-field discovery"; "team science AI collaboration tools study"; "LLM jargon translation domain adaptation evaluation"; "interdisciplinarity measurement bibliometrics AI 2025 2026".
- [ ] **Step 2: Write `sources.md`** (≥15, ≥5 peer-reviewed).
- [ ] **Step 3: Write `slides.html`** covering design §3/S6: why crossing fields is hard (cite interdisciplinarity literature); AI as field-translator (jargon, methods, canonical papers); adjacent-literature discovery via citation graphs + semantic search; cross-disciplinary synthesis — handling different evidence standards; team science (shared notebooks/knowledge bases, collaborator discovery); the 4-step cross-field orientation workflow; footnoted; references slide.
- [ ] **Step 4: Render-verify slides.**
- [ ] **Step 5: Write `demo-script.md`** — live cross-field orientation: researcher from field A (pick a concrete pairing during research, e.g. materials scientist → machine-learning-for-materials literature) runs the 4-step workflow: jargon translation prompt → landscape map via citation-graph tool → AI-annotated reading list of 8 papers → one-page cross-field synthesis memo. Exact prompts, expected outcomes, fallbacks.
- [ ] **Step 6: Write `handout.md`** — the 4-step workflow as a checklist with prompt templates; cross-field tool table; pitfalls list (false-friend terminology, differing evidence standards); further reading.
- [ ] **Step 7: Citation audit.**
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 9: Session 7 — Discipline Deep-Dives: Specialized AI by Field

**Files:**
- Create: `session-7-disciplines/sources.md`, `slides.html`, `demo-script.md`, `handout.md`

**Interfaces:**
- Consumes: templates; Session 7 objectives from `curriculum.md`; policy table (Session 4) for the medical segment; workflow canvas (Session 5) and orientation workflow (Session 6) for the closing.
- Produces: series finale; per-discipline "AI frontier map" one-pagers inside the handout.

- [ ] **Step 1: Deep research** — this session needs the widest sweep; queries per field: **Medicine:** "FDA cleared AI medical devices count 2026", "clinical trial matching AI evaluation", "PubMed-grounded clinical AI assistants evaluation"; **Biotech:** "AlphaFold 3 applications 2025 2026 review", "AI protein design enzyme Nobel 2024 aftermath review", "AI drug discovery clinical pipeline 2026 status"; **Engineering:** "ML simulation surrogates engineering review", "generative design CAD AI 2026", "engineering code copilot adoption"; **Physical sciences:** "AI materials discovery GNoME A-Lab review", "machine learning astronomy physics pipelines review", "AI theorem proving AlphaProof Lean 2026"; **Social/humanities:** "LLM qualitative coding evaluation study", "computational text analysis social science LLM review", "digital humanities AI 2026", "LLM agent-based simulation social science". Prefer review articles + primary announcements.
- [ ] **Step 2: Write `sources.md`** (≥20 entries for this session, ≥8 peer-reviewed/preprint, tagged by field).
- [ ] **Step 3: Write `slides.html`** covering design §3/S7: 4–5 slides per field (landmark systems, what changed for practitioners, 2–3 named tools, one caution) for medicine, biotech/life sciences, engineering, physical sciences & math, social sciences & humanities; regulatory context in the medical segment (link Session 4 policy table); closing: "track your field's frontier" (key venues, benchmarks, review articles — cited) + series wrap linking back to the workflow canvas; footnoted; references slide. Cap: 32 slides for this session (7 topics justify exceeding the 30 default slightly).
- [ ] **Step 4: Render-verify slides.**
- [ ] **Step 5: Write `demo-script.md`** — audience-poll format: prepare all five field demos, run the top two voted; each field demo = one 5-minute walkthrough in a field-specific tool (e.g. AlphaFold Server structure prediction; LLM qualitative coding of 10 sample survey responses — include the sample responses, synthetic, in the script). Exact prompts/steps, expected outcomes, fallbacks for all five.
- [ ] **Step 6: Write `handout.md`** — one "AI frontier map" per field (named tools, one seminal review to read, where to follow updates); series-wide closing checklist; further reading. This handout may run 4–5 pages given five fields.
- [ ] **Step 7: Citation audit.**
- [ ] **Step 8: Ehsan review checkpoint.**

### Task 10: Series-wide consistency pass

**Files:**
- Modify: all `session-*/slides.html`, `session-*/handout.md`, `curriculum.md` (fixes only)

**Interfaces:**
- Consumes: everything.

- [ ] **Step 1: Cross-session audit** — recap slides reference the right prior artifacts; terminology identical across sessions (e.g. the tool-taxonomy names from Session 1 used verbatim later); no tool described inconsistently; all date stamps consistent.
- [ ] **Step 2: Curriculum sync** — `curriculum.md` abstracts/objectives still match what each session actually delivers; fix drift in `curriculum.md`.
- [ ] **Step 3: Full render pass (SERIAL, guarded)** — earlier per-session render checks ran while other agents shared the same Playwright browser and were demonstrably hijacked (a Session 4 audit measured Session 5's deck three times, including one screenshot). Re-verify ALL SEVEN decks one at a time, with no other render work in flight, using: (a) a page-identity assertion before every measurement (URL matches this deck's own loopback port AND the document title/unique token matches), and (b) text-node-level measurement via TreeWalker resolving each node's parent computed style — never container-level selectors, which is how a 15px sub-label survived a check reporting 19px. For each deck confirm: slide count, zero overflow at viewport ≥1392px, footers sequential and correct, no external network requests, and minimum content font-size ≥18px (captions and citation superscripts exempt). Verify print-to-PDF on one full deck. Stop every server started.
- [ ] **Step 3b: Series-wide small-text audit** — decide and apply one rule for `.statlab` stat-card labels (currently 16px, inherited from Session 3 into Session 4) and any other sub-18px classes found across decks; apply the decision uniformly so no deck diverges.
- [ ] **Step 4: Final report to Ehsan** — inventory of all deliverables + any open questions (e.g. which demo accounts need setting up before the first webinar).