Fun Prompt Friday: Assigning Subagent Swarms with Claude Fable 5, Opus 4.8, and Sonnet 5

Don't hand your most expensive AI a shovel. How to run a swarm of Claude models like a research staff, and spend the rare genius only on the hard call.

UPDATE, Monday, July 6. Since Friday, the most common reader note has been: “I pasted your prompt into Claude and it just talked. No team ever showed up.” You did nothing wrong, and the fix is one sentence:

Subagent teams do not run in the regular Claude chat window. They run in Claude Code and Claude Cowork.

The chat box on claude.ai and in the mobile app cannot spawn separate worker models, so it does the only thing it can: it performs the script, one actor playing all the parts. That performance looks like a team report, but the safety feature at the heart of Friday’s design, the independent checker with fresh eyes, never actually exists there. If your chat run produced tidy results, treat them as a rehearsal, not a verification. Open Claude Code or Cowork, attach your record images as files (not pasted screenshots), attach your research standard, and run the same prompt there.

Two timely notes. First, the subsidized Fable access mentioned below ends tomorrow, Tuesday, July 7; after that Fable is pay-as-you-go. Second, I am taking my own advice today: as I write, a team of models is re-reading all 155 records of my Ashe County corpus, the project Fable began overnight in June before access was cut off. Fable is finishing what Fable started, on its last subsidized day, and when the results are checked and scored you will read that story here.

If you have been reading along, you already met Fable 5. In “Fable 5: The Night Agent” I watched it work long and mostly unattended on real family history. In “Fable’s Back for a Week” I argued for a specific way to spend a scarce, expensive model: do not ask it to do everything; let it plan, hand the routine work to cheaper models, and check what comes back.

Today’s post takes that idea one step further, and inverts it.

Fable is remarkable, but today it is rare and costly. There is a window worth knowing about: through Tuesday, July 7, paid Anthropic users get subsidized access to Fable. If that is you, use it, both for the experience of a Mythos-class model and for the introduction to managing a team of agents. Even so, you do not get much of it, and you should not squander it on grunt work. The move most people have never tried is to assign subagents: put one model in charge as a conductor, and let it direct others. Most readers have never orchestrated one AI to run a team of AIs, so let me unpack it plainly.

Here is the mental model. Opus is the conductor. Sonnet does the labor. Fable is the expert you consult once, at the end, for the hardest judgment. Spend the rare resource only where it earns its keep.

The cast

Picture a small research office:

• You are the boss. Every conclusion is yours; the AI never gets to declare an answer final.

• Opus 4.8 is the head archivist. It writes the plan, hands out the work, correlates the results, and produces the draft.

• Sonnet 5 is the scribe. Cheap and fast, it does the volume: reading each record and typing out exactly what it says.

• Haiku is the page, the cheapest routine worker, named here so the roster is complete.

• Fable 5 is special collections. You reach it rarely, by appointment, for the one high-value pass.

The task and the materials

I gave the office three real records from my own Ashe County, North Carolina research: an 1880 census page and two death certificates, 1916 and 1955, for a Lawrence family. I also attached the Genealogical Research Assistant skill, my GPS-aligned research standard. The GRA is the yardstick; everything the team does is measured against it.

The three record images (use these or try your own):

Record 1 image -- A 1916 North Carolina death certificate for David S. Lawrence, Ashe County. It names his parents and puts his birth near 1847, one of the facts the team had to weigh.

Record 2 image -- An 1880 U.S. census excerpt, the David S. Lawrance household in Ashe County. This is the page where two of my AI readers disagreed on the head’s name.

Record 3 image -- A 1955 North Carolina death certificate for Henry A. Lawerance, Ashe County. It names his father, David, the link that ties the two generations together.

A word on ethics, briefly: every person in these records is long deceased (deaths in 1916 and 1955), so living-person privacy protections are not triggered. I still gave the image files neutral names, because filenames leak into links and backups.

The workflow

The orchestration runs in six steps. Opus:

1. Writes a plan before delegating anything.

2. Sends each record to its own Sonnet scribe to transcribe and extract, exactly as written, normalizing nothing.

3. Sends each transcript to a fresh Sonnet checker that never saw the scribe’s work, to catch misreads.

4. Correlates the verified transcripts and writes a genealogical draft.

5. Consults Fable once, to review that draft against the GRA.

6. Hands everything back to me and stops.

Two design choices carry the whole thing. First, every hand-off states the same four parts: objective, output format, sources, and boundaries. Vague delegation is what sinks these swarms; the repeated frame keeps every worker disciplined and honest. Second, the checker is a different agent than the scribe, holding only the image and the transcript. A checker that shares the scribe’s context inherits the scribe’s mistakes.

And one line of vocabulary the prompt guards carefully: a draft is not a proof statement. A proof statement is the formal GRA vehicle, requiring direct evidence, two or more independent sources, no unresolved conflicts, full citations, and a stated confidence level. What Opus produces here is the lesser thing on purpose: working narrative for a human to review, labeled plainly as a draft.

What actually happened

I ran it twice. The first run, in the browser, had a hidden flaw: the record images were pasted into the chat rather than attached as files, so the Sonnet workers could not actually see them. Only the conductor could. The second run, on my own machine, fixed that; the images were real files, and every scribe and checker read them directly.

Here is the honest result, which is more useful than a tidy one.

The teamwork paid off in one specific place: checking. On the census, two AIs read the same line and disagreed about the head of household’s name. One read “Davad Lawrance,” the other “David Laurance.” Neither is certain from the image I have. That disagreement is not a failure; it is the system working. It flags the exact spot where I, the human, need a better scan and a closer look. The first run could never have surfaced this, because only one model ever saw the page.

The draft, though, did not improve. If anything it slipped: my conductor’s second draft quietly dropped a fact the first had kept, and smoothed over a small age discrepancy it should have flagged. Better inputs, no better draft.

And then Fable, spent once, earned its place. It read the draft against the GRA and caught real errors the cheaper pipeline missed: no stated confidence levels on any conclusion, a claim that leaned on the wrong layer of evidence, and a piece of terminology I had claimed to be following while breaking it in the same document. It also handed back a ranked list of the next records to pull. The most telling part: the same handful of errors showed up in both runs, which tells me they are baked into the prompt, not bad luck; the fix belongs in the instructions, not in hoping the draft comes out clean.

That is the whole lesson in miniature. The cheap workers earn their keep at reading and checking. The rare expert earns its keep at judgment. The conductor’s own first draft is the weakest link. And you, the human, own every conclusion.

The fix worth stealing

One improvement came straight out of the run. When your scribe and your checker disagree and neither reading is certain, do not let the conductor silently pick one. Keep both readings, and flag it for a human. In genealogy the exact spelling of a name is evidence, not noise; surname drift like Lawrence, Lawrance, and Lawerance is a signal to preserve, not a typo to correct. Build that rule into your prompt.

What the team handed back

Here, at the end, is the payoff: the draft the swarm produced, with Fable’s review folded in. Read it for what it claims, and just as much for what it refuses to claim. This is a draft, not a proof statement; that distinction is the whole discipline.

DRAFT: a working narrative, not a proof statement.

The Lawrence family of Ashe County, North Carolina (three records, two generations).

David Samuel Lawrence appears here as a farmer of Old Fields Township, Ashe County. His 1916 death certificate gives his birth as February 8, 1847, and his age at death as 69, consistent with that date; the 1880 census, recording his age as 34, points a year or two earlier, to about 1845 or 1846. The exact year is Not Proved from these three records alone, because each reports the birth secondhand. His 1916 certificate names his parents as Hardin Lawrence and Rebecca Burkett; that is a single, secondhand source, so their parentage is Possible, not established. David died December 26, 1916, and was buried at Beaver Creek. He was recorded as married, yet whether that wife was still the Margaret of 1880 or a later wife, and whether the informant “Lizzie Lawrence” was that wife, these records cannot say.

In 1880, a David S. household in Ashe County held his wife, Margret M., 36; sons Henry A., 9, and William H., 7; and an eighteen-year-old servant, Debora Richoson. (Two of my AI readers disagreed on whether the head’s name reads “Davad Lawrance” or “David Laurance.” I have left both readings standing; the exact letters await a better image.)

Henry Alexander Lawerance, the Henry A. of that household, was born October 17, 1871, by his 1955 death certificate; the census age of 9 fits, allowing for the usual imprecision. He died a widower on November 18, 1955, and was buried in the Lawerance Cemetery. His certificate names his father as David S. Lawerance. That this is the same David of the 1880 and 1916 records is a Probable identification, supported by the shared name, the matching county, farming in both records, an age that agrees within two years, and the son Henry who appears in both stories. It is not yet proved; David’s household in the 1900 census would likely settle it.

Henry’s mother is named “Margaret Lawerance,” the same given name as David’s 1880 wife, so almost certainly the same woman (Probable). But “Lawerance” is her married surname; her maiden name is unknown, and I will not guess it.

What these three records give, then, is an outline: Hardin Lawrence and Rebecca Burkett, to David S. Lawrence and his wife Margaret, to Henry Alexander and his brother William. The David-to-Henry link is the firmest thread. The rest, David’s exact birth year, Margaret’s maiden name, and the identities of both informants, are the honest edges of this draft, and each one points to the next record to pull.

That last paragraph is the whole point. A swarm of models, well managed, does not hand you certainty. It hands you a clean, sourced, honest draft, and a short list of exactly what to chase next. The craft is knowing which worker to spend where, and remembering that the final word is still yours.

Try it yourself

This uses Claude’s subagents and the Fable, Opus, and Sonnet models, so it requires a paid Anthropic subscription (about $20 per month, and you can buy a single month at a time). Better still, if you act through Tuesday, July 7, that access to Fable is currently subsidized, which makes this an inexpensive week to try it. The full orchestration prompt is at the end of this piece. Attach your records as image files, not pasted screenshots, and attach your research standard.

New to any of this? You are welcome here. Copy this article’s link or full text into your favorite chatbot and ask: “Explain this to me as if I were: 1) a fifth grader, 2) a tenth grader, and 3) a curious adult with no prior knowledge of these things, in about 125 words per level.” Then ask it follow-up questions, and ask it to walk you through a demonstration step by step.

---

Appendix: the orchestration prompt

Paste this into a fresh Claude Opus 4.8 session with your record images and your research standard attached.

You are Claude Opus 4.8, acting as ORCHESTRATOR for a genealogical extraction-and-drafting experiment. Attached to this session are your record images and a genealogical research standard (I use the GRA). The standard is your yardstick: everything below is done toward it and reviewed against it.

YOUR ROLE

You route work and synthesize. You do NOT do the routine extraction yourself; that goes to cheaper subagents. Reserve your own effort for planning, correlation, and drafting. Reserve the expert model (Fable) for exactly one final pass. Scale effort to the task: three records need three scribes and three checkers, not a swarm.

STEP 0 - PLAN. Write out your plan: which subagent runs when, on which record, with what model, producing which named artifact. Show the plan, then execute it.

STEP 1 - EXTRACTION (Sonnet, one record at a time). For each record, spawn one Sonnet subagent. Every delegation must state all four of: Objective (transcribe fully, then extract every name, date, place, relationship, occupation, and informant; record spellings EXACTLY as written, do not normalize); Output format (a labeled verbatim transcript, plus a structured fact table); Source/tools (that one image only, no outside lookups); Boundaries (mark illegible [unclear], doubtful [?reading], empty [blank]; never guess or invent; [citation needed] over fabrication).

STEP 2 - INDEPENDENT CHECK (fresh subagents; maker != checker). For each transcript, spawn a NEW Sonnet subagent that did not do the extraction. Same four-part spec: compare the transcript against the source character by character; flag misreads, omissions, unmarked uncertainty, and any normalization. Source/tools: the one image plus the one transcript. Apply corrections in three dispositions: accept, reject, or UNRESOLVED standoff. When scribe and checker disagree and neither reading is authoritative, keep BOTH readings and flag the item for human review; do not silently pick one.

STEP 3 - CORRELATION (you, Opus). From the verified transcripts, build one timeline and compute implied birth years; reconstruct the household and parent-child links; classify each source, each item of information, and each use of evidence per the standard, using full labels; list every conflict, gap, and spelling variant explicitly. Name any same-name identity inference rather than merging silently, and assign it a confidence level.

STEP 4 - THE DRAFT (you, Opus). Write a genealogical DRAFT: narrative prose across the generations, with in-text citations and uncertainty marked inline. State a confidence level per conclusion, and include citation skeletons with bracketed placeholders for missing detail. Label it clearly: a DRAFT, not a proof statement.

STEP 5 - EXPERT REVIEW (one pass, once). Spawn a single Fable subagent, the only Fable call in this run. Objective: review and improve the DRAFT against the standard (evidence classification, conflict handling, citation discipline, terminology, unwarranted certainty, and anything the correlation missed), then list the highest-value next research steps. Source/tools: the DRAFT, the verified transcripts, the correlation, and the standard. Boundaries: critique and suggest; the human decides what to accept.

STEP 6 - HAND BACK AND STOP. Present the verified transcripts, the correlation, the draft, and the critique, followed by two lists: conflicts that cannot be resolved from these records alone, and open questions with the record most likely to answer each. Then stop.

---