Groundwork

Groundwork implements Samantha McKenna's Show Me You Know Me® method, which has one almost uncomfortably literal requirement: open with a fact so specific to the recipient that it would mean nothing to anyone else — and make every such fact traceable, source URL and all, to something they actually posted, said, or wrote.

The pipeline is 31 n8n nodes in five stages, with deterministic control between every model call. Intake collects the sender's company data and targeting criteria once, then an Apollo-style sourcing step returns leads keyed on LinkedIn URL. A checkpoint stage reads back every lead already finalized in a previous run, so a batch that dies partway through never re-researches or re-writes a finished lead. Then, one lead at a time: a Research Agent works under a hard budget of seven tool calls and a strict source-priority order — personal LinkedIn activity first, company news dead last — and a Synthesis Agent condenses the findings into a single Personalization Brief, discarding anything without a source URL and committing to one primary angle the rest of the pipeline is not permitted to abandon.

The core is a bounded draft-and-critique loop. A Generator (Claude Sonnet) drafts subject and body from the Brief and nothing else; a Critic (Gemini Flash) scores the draft against a six-dimension rubric and checks every claim in it against the verified research. The loop runs at most two rounds. A draft that passes is written to Airtable as ready_to_send; one that fails twice — or includes a fabricated personal fact, an instant knockout — is escalated to needs_manual_review. Never silently either one.

State lives in three Airtable tables: a durable identity record, a checkpoint table the dedup logic reads on every run, and a polymorphic audit log where each agent writes its own events — the record that lets any claimed fact be traced back to where it came from. The build is validated end-to-end through this persistence layer; the actual send is deliberately not wired, documented on the canvas itself as the operator's decision to make.

Five specialized agents instead of one.

Research and writing are different problems that fail in different ways, and a single model asked to find something true and write convincingly will reliably do the second at the expense of the first. Splitting the work lets each seat run the model suited to it — agentic tool-use where the job is deciding what to look at next, structured extraction where the job is condensing, prose reserved for the one seat where prose quality is decisive, judgment on a different model entirely. That discipline, not any single clever prompt, is why five model calls produce one coherent, source-traceable email instead of five disconnected opinions.

A deterministic workflow instead of an autonomous orchestrator.

The first architecture draft had a supervisor agent watching the whole sequence — research, synthesize, write, critique, finalize — deciding at each step whether to continue or retry. It would have demoed beautifully. It was also the wrong pattern: orchestrator-workers is for workflows whose subtasks can't be predicted in advance, and this sequence is identical for lead #1 and lead #500. Worse, it puts a language model in charge of counting its own iteration cap — across a fifty-lead run, fifty independent chances to miscount — at roughly fifteen times the token cost of the deterministic version.

The supervisor was removed entirely, not scaled down. Dynamic tool selection survives in exactly one place, the Research Agent's toolbelt, because which platform holds a usable personal signal for a given human genuinely can't be known before looking. The instinct to keep the more impressive-looking layer was real; “will this look good in the case study” is a bad reason to add architecture.

The judge comes from a different reasoning family than the writer.

Frontier models judging their own family rate it favorably 75–84% of the time in pairwise tests, and similarity testing clusters Claude with Mistral, Gemini with GPT. So Claude writes and Gemini judges — a cross-cluster check that behaves like a second opinion rather than a reflection. The two-round revision cap follows the published Self-Refine data: most of the measurable gain lands in the first pass or two, and everything after buys shrinking improvement at real token cost.

One gap, shipped on purpose.

The research behind this build recommended enforcing that two-round cap with a deterministic loop-counter gate — code, not instructions. The live system currently enforces it through the orchestrator's own prompt. Testing so far, including a live run that reached round two, shows the self-counting holding — but “holding so far” and “structurally guaranteed” are different claims, and only one is true today. The simpler version shipped first, observed in the open, with the hardened gate named as the known next step.