Sharing a build from the last few days that genuinely surprised me at the workflow level.
I work on K-12 education funding policy. I had an Airtable base called Adequacy that was supposed to model public school funding gaps state by state, but the v1 build was rough — Oregon residue in field descriptions, hardcoded formula constants, an "ADMw" field name baked in from one state's terminology, no real per-district adequacy targets.
The goal was to point Hyperagent at a clean version of this base and get it to produce publication-grade adequacy analysis.
Step one: terraforming with Claude + Airtable MCP
This is the part of the workflow I want to flag for this audience. I connected Airtable to Claude through the MCP connector, then worked with Claude in conversation to rebuild the base. Claude made the edits — updating field descriptions to state-neutral language, renaming ADMw to Weighted Enrollment, adding a new Adequacy Target Per Student input field, rewriting the Adequacy Target formula from a hardcoded IF({Weighted Enrollment} < 10000, 13000, 12000) to {Adequacy Target Per Student} * {Weighted Enrollment} then rewriting downstream formulas (Adequacy Gap, % of Adequacy Met, Funding Status, the categorical flags) so the unit math stayed consistent.
What I did: described the problem, made architectural calls (use real per-district adequacy targets as inputs vs. hardcode a constant, split total-dollar gap from per-pupil gap into two fields), verified the output against ISBE's published numbers.
What Claude did: the actual base edits, the formula rewrites, the schema changes, loading 30 verified Illinois districts as a first pass, then producing a CSV for the remaining 820 when the API loading would have been slow.
About 90 minutes of working session and the base was state-neutral, formulas verified, 851 Illinois districts loaded with FY26 EBF data.
Step two: Hyperagent Role
Used Hyperagent's Custom AI Agent Development flow to spin up EBF Analyst: a specialist agent with an identity prompt and a separate Skill called EBF Adequacy Methodology holding schema references, redistribution math, and scenario archetypes. Saved both.
Step three: two analytical prompts
The first prompt asked for a state-of-the-state adequacy briefing for the Illinois State Board of Education. The second asked the same agent for a philanthropic-funder strategic landscape using the same data, different audience.
Both outputs came back as designed publication artifacts — Stripe Press / Linear aesthetic, structural finding as a typographic event, histogram of all 851 districts colored by adequacy tier, named district callouts, scenario comparison tables, methodology colophon.
Attached Artifacts:
https://hyperagent.com/s/rSnVaT3Xdn0DwcBNTy_hSg
https://hyperagent.com/s/Z-G3i4ryNPL--ddzezdKNQ
https://hyperagent.com/s/0rqYNGaNRUQwsI3Te0ipyQ
What I learned
Three things worth flagging:
One: The Claude + Airtable MCP workflow changed what counts as a one-person job. Two days ago this was a base I would have had to either manually overhaul or hire someone to clean up. The data-semantics work that Hyperagent's own docs say matters most (rewriting field descriptions, fixing formula unit math, renaming state-specific fields to state-neutral ones) is exactly the work the MCP connector makes tractable. I did not write field descriptions one at a time. I described what was wrong with the schema and Claude rebuilt it.
Two: The architectural finding worth knowing: my base has five pre-existing Omni Agents (Field Agents) producing AI text outputs at the record level. EBF Analyst, in its own methodology note, explicitly disclosed that it did not use those fields and worked only from the numerical columns. Hyperagent treats Field Agent outputs as data, not as authoritative reasoning. Worth knowing if you have built Field Agents into your base and were assuming Hyperagent would adopt them as workers.
Three: The second analysis was sharper than the first. Hyperagent's Memories system carried context forward, so the second invocation produced strategic reasoning that built on the analytical foundation of the first. The compounding context worked as advertised.
Happy to answer questions about the workflow, the prompts, or what the Claude / Airtable / Hyperagent combination unlocks in practice.