Each micro-workflow is tied to one or more named releases, and is meant to be something you can test in a 15–30 minute session with real records.
“Binder-in-one” brick wall review (GPT-5.5, 1M context)
Load a full case file—census extracts, deeds, probate abstracts, correspondence, and your current research log—into GPT-5.5 and ask: “Identify every explicit relationship, implied relationship, and geographic clue, then list three plausible hypotheses for X’s parents with the supporting evidence and gaps.”
Use its output as a working hypothesis list, then annotate each point with your own reliability assessment before accepting anything.
Multi-county land chain reconstruction (GPT-5.5 “super app” behavior)
Upload a sequence of deeds across counties and states, then ask GPT-5.5 to build a property chain-of-title table with grantor/grantee, metes-and-bounds descriptions, neighbors, dates, and inferred geographic moves.
Follow up by asking it to draft a narrative of the land migration and identify which missing deed books or tax lists you should target next.
Long-form research narrative polish (Claude 4.7)
Paste a messy, 8–12 page draft report into Claude 4.7 and ask it to reorganize into sections (Research Question, Background, Evidence Summary, Analysis, Conclusion) while preserving every factual statement.
Then ask for a side-by-side table listing your original sentences vs. the revised version so you can check for meaning drift before adoption.
Source-citation sanity pass (Claude 4.7 + Claude Code stability)
Provide Claude with a report plus its footnotes and ask it to: (a) identify inconsistent citation formats, and (b) flag any citations where the narrative claim seems stronger than the source could support.
Let it propose a harmonized citation pattern (e.g., Evidence Explained-inspired) while you remain the final authority.
Persistent “project steward” agent (Claude Managed Agents with Memory)
Configure a Claude-managed agent labeled “Courtright Project Steward” and feed it your research objective, key ancestors, and a link or upload of your current log.
Each week, ask: “What did I do last time? What are the next three concrete research tasks, with record types and repositories?” and let the memory feature keep track of progress so you don’t have to re-explain every time.
Gemini-driven research log generator (Gemini file generation)
Paste a set of records you’ve consulted for a family and ask Gemini to generate a CSV or Sheets file with fields like date, jurisdiction, record type, repository, call number, person(s) mentioned, and key findings.
Download the CSV/Sheet and store it with your project; repeat whenever you finish a new research session.
Citation-ready timeline in one click (Gemini → DOCX/PDF)
Provide transcriptions or abstracts for a single ancestor and ask Gemini to create a chronological narrative with inline citation placeholders, then export as DOCX or PDF.
Open the file locally and fill in the precise citation formats using your preferred style, turning Gemini into a drafting tool rather than a final authority.
On-the-road research companion (Gemini on devices/cars)
While traveling to a courthouse or archive, use Gemini on your phone or car interface to review your research question and have it read a short briefing aloud: “Remind me of the exact place, time frame, and open questions for the Smith line in X County.”
Capture follow-up tasks verbally (“Create a to-do list for deeds 1870–1900 X County”) and later download the resulting file from Gemini when you’re back at your desk.
Parallel evidence check with Model Council (Perplexity)
Run a focused query like “Where are digitized land records for Logan County, Oklahoma Territory, 1890–1907?” using Perplexity’s Deep Research/Model Council mode so it uses GPT-5.x, Claude, and Gemini in parallel.
Compare where the models agree and diverge about repositories and collections, then verify each suggested collection directly in the linked catalog.
Automated locality background sheet (Perplexity Deep Research + Computer)
“Two AI readers, one human judge” document review (GPT-5.5 + Claude 4.7)
Open-weight local assistant for sensitive data (DeepSeek V4 / Qwen / Kimi K2.6)
For material you’re uncomfortable uploading to cloud tools (DNA notes, living relatives’ information), set up a local or private-cloud instance of an open-weight model like DeepSeek V4 or Qwen3.6-27B and fine-tune or prompt-tune it on generic genealogical tasks (e.g., log summarization, neutral language rewriting).
Use it to clean and standardize your internal notes without exposing sensitive or living-person data externally.
Bulk research-log clean-up (open-source model + CSV)
Export a messy research log as CSV, then locally run an open-weight reasoning model (Kimi K2.6, Qwen, etc.) to normalize locality spellings, ensure consistent date formats, and tag each row with a “next action” (follow-up, completed, needs citation).
Review those automated tags in your spreadsheet and adjust where needed; this is especially useful for long-term projects with hundreds of entries.
Audio companion for long narratives (xAI Text-to-Speech)
Take a dense research report or family narrative and send the text through xAI’s Text-to-Speech API to generate an audio version for private listening while commuting or doing chores.
As you listen, jot down moments where the story drags or feels unclear and mark those spots back in the text for revision.
Multi-agent task split for big projects (Grok 4.20 Multi-agent Beta)
In xAI’s Enterprise API environment, configure one Grok agent for “record extraction” and another for “analysis.”
Have the first agent extract structured facts (names, dates, places, relationships) from a packet of documents, then hand that structured data to the second agent to build timelines and identify conflicts.
“Last session recap plus next steps” assistant (Claude Managed Agents memory)
For a long-running research problem, use a Claude managed agent with memory to store your summaries after each session: “Today I searched X, found Y, didn’t find Z.”
At the start of the next session, ask: “What did I do last time and what three follow-up actions do you recommend with specific record types and repositories?”
Notebook-to-outline converter (Gemini → Markdown / DOCX)
Repository scouting with Deep Research (Perplexity)
Ask Perplexity Deep Research for “All major online and offline repositories holding records for [county/state] during [time period] relevant to land, probate, and vital records,” and let it search across multiple sources for catalogs and guides.
Use the output as a starting point to build your own locality guide in your note-taking system, checking each repository’s catalog directly.
“What am I missing?” pattern hunt (GPT-5.5 extended context)
Upload a large Excel sheet of extracted census entries, tax lists, and city directory entries for one surname and locality, and ask GPT-5.5 to: (a) cluster entries into likely family groups, and (b) propose hypotheses for identity merges/splits you should investigate.
Follow up with targeted queries: “Which entries for ‘J. W. Smith’ are most likely the same man, and why?” and then verify each claim in the original records.
AI-assisted syllabus and handout drafting (Claude 4.7 + Gemini file export)
For an upcoming talk or class, feed Claude 4.7 a rough session idea and ask it to propose a 1-page outline and a list of concrete, exercise-based learning objectives for genealogists.
Paste that outline into Gemini and ask it to generate a formatted handout as a DOCX or PDF with headings and blank spaces for participant notes, ready for minor editing and branding.
Ethics and verification checklist companion (any major model + RootsTech guidance)
Using a model of your choice, paste the recently highlighted RootsTech principles for responsible AI in family history and ask the model to turn them into a short checklist you apply to each AI-assisted project (e.g., “Did I verify all claims in original records?”).
Save that checklist as a reusable template in your research binder and run through it before finalizing any AI-assisted narrative or conclusion.
No comments:
Post a Comment