We built a Claude Code skill that convenes a panel of named, real retail commercial real estate investors to deliberate, argue, and reach a verdict on a real decision. Here is exactly how it was built and run, in order.
Step 1: Scope the panel
We asked Claude for the top 10 names in retail CRE investing specifically, excluding brokerage and development, going back 50 years. Before answering, Claude asked two clarifying questions: should the list include deceased or retired legends, or only people alive and plausibly recruitable as advisors; and how strictly should “retail” be defined. We answered: include iconic names regardless of whether they’re alive, and keep the retail definition strict.
Step 2: Get the roster
Claude returned a 10-name list. Five names surfaced directly in the working session: Edward DeBartolo Sr, Milton Cooper, the Bucksbaum family, David Simon, and Rick Caruso. We accepted the list as-is rather than hand-editing it, noting we don’t have deep personal expertise in retail to second-guess the selections.
Step 3: Spawn parallel research agents
We instructed Claude to spawn one research sub-agent per advisor, 10 agents total, running at the same time. Each agent’s task: build the most comprehensive dossier possible on that individual’s investment philosophy, risk tolerance, strengths, and weaknesses, and write it to a shared research file.
Step 4: Research completes
All 10 agents finished in roughly an hour. Combined output: approximately 47,000 words of sourced, quote-rich profiles, one per advisor, saved to a single research file.
Step 5: Define the skill’s behavior
We told Claude to turn the research into a triggerable skill and to ask clarifying questions before building. It asked four:
- What range of decisions should the board weigh in on? Answer: anything.
- Who from the board should speak on a given question? Answer: a “convener” selects the 3 most relevant voices per question, and that selection is overridable.
- How should the deliberation be presented? Answer: two modes, quick and deep, defaulting to a verdict plus the key tensions.
- How should a deliberation end? Answer: recommend a path and preserve dissent rather than forcing consensus.
Step 6: Build the skill
Claude built the skill with this fixed sequence: convene the relevant advisors, read each one’s research file, state opening positions, run the argument, run a mandatory pre-mortem, synthesize a recommendation, and log any dissent separately from the synthesis.
Step 7: Run a live test
We gave the skill one prompt: “I’m considering investing in a new deal in a new market. Let’s say the market is in Phoenix, Arizona. Shopping center acquisition. How should I be thinking about this?”
Step 8: The board convenes and seats advisors
The skill read the relevant advisor files and seated DeBartolo Sr, Milton Cooper, the Bucksbaum family, David Simon, and Rick Caruso for this question, with the rest of the panel benched as not relevant to a Phoenix retail acquisition.
Step 9: Opening positions
DeBartolo took the bull position on the Phoenix market itself, citing population growth and the logic of getting ahead of demand. Cooper and Simon framed the deal as a spread and basis problem rather than a market-growth story. Caruso named what the system called the outsider problem: no local broker, no trusted property manager, no read on whether the anchor tenant is actually healthy.
Step 10: Argument
DeBartolo directly challenged Cooper and Simon’s caution, arguing that over-caution is its own way to lose and that he built through the worst recession since the Depression on the same suburban-growth thesis.
Step 11: Mandatory pre-mortem
The board ran a fixed exercise: assume it’s 18 months later and the deal failed, then state exactly how. The advisors named specific failure modes: overpaying relative to basis, buying the wrong corner, the anchor tenant going dark with no way to backfill it, and leverage that couldn’t survive a downturn.
Step 12: Synthesis
Final recommendation: pursue the deal only if it is a necessity-anchored, grocery-backed asset; bought below replacement cost rather than at a market cap rate; financed conservatively with long fixed-rate debt and real equity; and backed by genuine local market knowledge, either partnered in or independently verified. Stated confidence: medium to high.
Step 13: Dissent logged
DeBartolo’s dissent was recorded separately rather than folded into the synthesis: he stated for the record that he disagreed with the caution in the room, that fortunes in retail real estate are made by betting ahead of growth, and that waiting for a perfectly priced, fully de-risked deal risks watching the trend get captured by someone else.
Result
A working, triggerable Claude Code skill that takes any real decision as input and returns a structured deliberation: seated advisors, opening positions, an argument phase, a mandatory pre-mortem, a synthesized recommendation, and any dissent preserved on the record rather than smoothed away.
Caveats
The five advisors named above are the ones referenced directly in this build session; the remaining names on the 10-person roster were not specified in the available record. The research dossiers are built from public sourcing, not live interviews, so deceased advisors’ positions are reconstructions from documented philosophy and quotes, not new statements. The Phoenix scenario was a hypothetical test, not a real deal being screened. The skill has so far been run on one test question.
The skill is live and can be triggered for any decision going forward.
