horizontwo/marymary PR #23 — Prospect brief work product — PR #23
horizontwo/marymary · pull request #23 ·
Transcript
PlainEnglish
Welcome. We're walking through pull request twenty-three for Mary Mary — a new work product type called prospect-brief. This change introduces internal research dossiers on individual candidates, with a strict firewall between observed facts and AI interpretation. Let's see how it works.
PlainEnglish
This PR has two primary goals. First, enforce a strict fact-versus-interpretation firewall using error-severity validators. Second, enable the Research Gnome to produce prospect briefs. It also adds comprehensive validator tests and registers the new type in the work product system. Importantly, it preserves internal-only behavior — no delivery adapter, no autonomous outreach.
PlainEnglish
Before this PR, the Research Gnome could produce trend briefs and metrics, but had no way to research individual prospects. This change introduces the prospect-brief work product type — an internal-only research dossier with a structural firewall that keeps observed facts separate from AI interpretation. Every fit-score claim must reference underlying evidence, not just be flagged as a warning.
Architecture
Let's zoom out and see where the change lives. A new prospect-brief directory under workproducts holds the definition, test suite, and registration module. The main index adds a side-effect import to trigger registration, and the Research Gnome defaults file gains a producible type entry plus new system prompt rules. Six files total.
Architecture
Here's the definition itself. The schema enforces six required top-level fields — prospect, sourceSegment, observed facts, interpretation, fit score, and suggested ask. Four validators attach here: two error-severity checkers that enforce the fact-interpretation firewall, and two warning-severity hygiene nudges. The delivery field is explicitly undefined — this is internal-only, review-gated. No autonomous outreach.
CodeQuality
The two error-severity validators are the heart of the firewall. Evidence ref check ensures every fit-score point references a valid observed fact by index — out-of-bounds or missing references block delivery. Inference without facts check prevents the agent from writing interpretation when the observed array is empty. These aren't warnings — they're hard stops.
CodeQuality
This is the first work-product test suite in the repo. Ten test cases cover the happy path, all four validators, and registration. We check that out-of-bounds evidence refs and missing facts trigger errors, while contact hygiene and provenance issues stay at warning level. This establishes a testing pattern for future work products.
Architecture
Over in the Research Gnome defaults, we add prospect-brief to the producible work products array. This single line enables the agent to produce prospect briefs alongside trend briefs and metric proposals. The system prompt gains five rules that align with the schema and validators — facts stay separate from interpretation, every score point cites evidence, no email scraping.
Architecture
The registration flow is clean. The workproducts index adds a side-effect import for prospect-brief. That import triggers the module's registration file, which calls register work product and adds the definition to the global registry. Node module caching ensures registration happens exactly once, even if imported from multiple paths.
PlainEnglish
After this PR lands, the Research Gnome can produce prospect briefs — internal-only dossiers on individual candidates. Each brief enforces a firewall: observed facts stay structurally separate from AI interpretation, and every fit-score point must reference the evidence backing it. No delivery adapter, no autonomous outreach. Humans approve or reject each brief. This is the first work product with a test suite, establishing a pattern for type constraints and validator testing.
PlainEnglish
That's pull request twenty-three. A research workflow that enforces epistemic hygiene — facts separate from interpretation, every claim backed by evidence. The validators live in the definition, the test suite covers all four, and the Research Gnome's system prompt aligns with the constraints. Clean, internal-only, and review-gated. Thanks for walking through this with me.
How this was made
Lenzon read horizontwo/marymary at pull request #23 and generated this walkthrough automatically. The narration above is the transcript of what it says.
Explain a pull request from your own repo
Point Lenzon at a repo or a pull request and get a narrated walkthrough like this one.
Try it