flowchart LR
  s1(["1 · Foundations
today"]):::today --> s2(["2 · Literature
discovery"]):::next
  s2 --> s3(["3 · Reading
& notes"]):::rest
  s3 --> s4(["4 · Writing
& integrity"]):::rest
  s4 --> s5(["5 · Data
& code"]):::rest
  s5 --> s6(["6 · Multi-
disciplinary"]):::rest
  s6 --> s7(["7 · Discipline
deep-dives"]):::rest
  classDef today fill:#ab7d22,stroke:#d9b36c,color:#14132b,font-size:19px
  classDef next fill:#23204c,stroke:#d9b36c,color:#eceaf8,font-size:19px
  classDef rest fill:#23204c,stroke:#8f8cb8,color:#b9b7d6,font-size:19px
        

~20 min foundation · ~25 min live demo · ~15 min Q&A — every session.
Today builds the two artefacts the whole series reuses: the capability/failure map and the 2026 tool taxonomy.

Objectives 4–6 next · full verbatim wording in the reference deck and curriculum

Full verbatim wording in the reference deck and curriculum

Wiley ExplanAItions, fielded Aug 2025 (latest wave; vendor self-reported) ^[16] · peer-reviewed corroboration ^[15]

"Comparative genomics of penguins" —

outside the window: your other chats · yesterday's upload · the rest of the literature —
it exists only if the product deliberately re-injects it (saved memory, project files)

separate limit: reliable knowledge can end months before today's date ^{[2] [3]}

Put what matters at the top or bottom — and five well-chosen papers beat fifty dumped in. Measured on the 2023-era cohort ^[9]; no current-generation replication, so assume it still applies.

flowchart TB
  M(("statistical patterns
of plausible text")):::model --> R["Gorman et al. (2014)
real"]:::real
  M --> F["Sørensen et al. (2019)
invented — same font, same confidence"]:::fake
  classDef model fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:19px
  classDef real fill:#1c1a3f,stroke:#2fae74,color:#7fd6a4,font-size:19px
  classDef fake fill:#1c1a3f,stroke:#c14b58,color:#ef9a9a,font-size:19px
          

Confident, well-formed and wrong is the expected failure mode — not an anomaly ^[4]

What you can do today: give the model explicit permission to abstain, in your prompt ^[4]

Raise the dial for hard multi-step work; verify facts exactly as before.

"Safe" means safe to draft with — never safe to skip reading. Fabrication is structural; vendors document it in their own cards ^{[5] [7] [27]}

One caveat: Tier-2 tools such as Consensus and Scite now ship MCP servers, so a Tier-1 chatbot can call them directly ^{[30] [31]} — read the tiers as what you are reaching for, not walls. Examples illustrative, not endorsements.

flowchart LR
  P["plan"]:::n --> S["search"]:::n --> R["read"]:::n --> I["iterate…"]:::n
  I --> S
  I --> REP["cited report
minutes, not seconds"]:::hot
  classDef n fill:#23204c,stroke:#8f8cb8,color:#eceaf8,font-size:19px
  classDef hot fill:#1c1a3f,stroke:#d9b36c,color:#d9b36c,font-size:19px
          

More citations per query — and worse odds each ^[8]. Some let you edit the plan before it runs ^[23]. Treat the report as a well-organised lead list, never a verified synthesis.

Then iterate: a prompt is a first draft, not an incantation.

— Context  ·  — Role & register  ·  — Instructions & constraints  ·  — Task

Prompting is a documented literature — 58 techniques catalogued ^[13]. Context, constraints and retrieval buy correctness.

flowchart LR
  A["1 · Grounding
where did this come from?
nothing retrieved = nothing checked"]:::n -->
  B["2 · Citations
does every reference resolve
to a record you can open?"]:::n -->
  C["3 · Claim-to-source
does the source say
what the output claims?"]:::n -->
  D["4 · Omission
what was left out?
summaries lose caveats first"]:::n
  classDef n fill:#23204c,stroke:#d9b36c,color:#eceaf8,font-size:19px
        

Checks 2–4: ^{[5] [7] [6] [18]} — a real link can still carry a false claim; open the record, don't trust the format.