An account research kit for enterprise sellers. One command turns an account name into two artifacts: a sourced strategy brief, and a deal model with qualification, bottom-up economics, and a seat-and-ARR ladder.
.\scripts\new-account.ps1 -Account "Lowe's" -Mode bothThere is also a web app: type a company name, get a sectioned dossier with every figure cited.
node web\server.mjs # then open http://localhost:4317Building an account strategy by hand takes a day and produces something that goes stale in a month. Most of that day is not thinking, it is retrieval: earnings, leadership changes, the engineering blog, job postings, who owns security. That work is parallelizable and it is exactly what an agent is good at.
So the kit fans out six read-only research subagents at once, each on a separate slice of the public record, then assembles their sourced findings into a brief. What stays human is the argument: which pattern to lead with, who to call first, and what to ask for.
The risk with generated research is confident fabrication, which is worse than no research because it fails in front of the buyer. Three mechanisms push against it:
- The researcher subagent cannot write files and cannot draft narrative. It returns claims with source URLs and publication dates, or it reports that it found nothing.
- Evidence rules are explicit and shared. Quote figures as published; never annualize, convert, or ratio them. Label estimates as estimates and show their inputs. State what is not public, and what was searched.
- The runner enforces a gate. No artifact, too few sources, or no unknowns section means a non-zero exit. The gate is the contract.
brief.md — the play in five lines, position and pressure, a dated catalyst timeline, Factory's three operating patterns mapped to their actual named systems, the buying center split into economic buyer / champion / gatekeeper / practitioner, an honest competitive read including trade-offs, a multi-threaded reach plan timed to their fiscal calendar, a first-call script, objections, and an unknowns section that converts straight into first-call questions.
deal-model.md — a qualification scorecard where every dimension is Known, Hypothesis, or Gap with a closing action; explicit disqualification criteria; a bottom-up value model built from one defensible unit; and a seat-and-ARR ladder where every rung carries a gate. Pricing appears as a labeled assumption with the arithmetic shown, never as a quote.
I ran the kit on Lowe's, an account nobody had prepped, and audited the output rather than admiring it. Four findings, all of which changed the tool:
1. The gate I wrote was measuring the wrong thing. It counted source links. 67 citations passed cleanly. But counting citations proves nothing, because a fabricated URL and a real one look identical in a word count. So I wrote verify-sources.ps1, which resolves every link. That is now part of the gate.
2. Nothing was fabricated. This was the finding I most expected to go the other way. Every citation I audited pointed to a real page that said what the brief claimed, including the SEC 8-K, which matched all five executive appointments exactly. The evidence discipline held under audit.
3. Link checking is harder than a status code. Three separate traps showed up in one run: corporate newsrooms return HTTP 200 with a "page not found" body; SEC and Reuters answer non-browser clients with 403 or 401; and job postings return 410 within months because postings expire. A naive checker would have called a dozen valid citations broken. The checker now separates dead from blocked from soft-404 and only fails on the first.
4. Once the checker was in the workflow, the agent fixed its own citations. The first Lowe's dossier carried seven soft-404s and two dead links. After the verifier became a step in the run rather than a review afterthought, the agent re-resolved its sources and rewrote them: 28 sources, zero soft-404s, zero dead. Delta and Best Buy came back with zero dead links on the first pass. A measurement placed inside the loop changes the output; the same measurement placed at the end only grades it.
5. Honest gaps are the most useful output. The deal model marked five of eight qualification dimensions as Gap — metrics, decision criteria, decision process, paper process, identified pain. That is correct and it is the point: none of those are knowable from public research. They are exactly what a first call is for, and an account plan that invented them would be worse than one that admits them.
The broader lesson for selling this: an agent will produce a confident artifact whether or not it is right, so the value is in the verification you wire around it. Research is cheap now. Proof is the product.
The markdown artifacts are for reading. The web app is for a live demo: someone names a company they care about, and a dossier assembles in front of them.
node web\server.mjs
# Prospect dossier app -> http://localhost:4317Type a company, press Research, and the server spawns the same headless run the CLI uses. Progress streams into the page; the dossier renders when it lands. Saved dossiers load instantly from the sidebar.
Sections: company overview with scale figures, trends (reported results, guidance, industry and macro pressure), a dated catalyst timeline, the engineering estate mapped to Factory's operating patterns, a scored fit assessment arguing both sides, the top-ten contact list ranked by who to call first, the competitive field, deal shape, objections, unknowns, and the full source list.
A sourced dossier tells you what is true. It does not tell you whether you have been here before, what to say, how to get in, or whether any of it is still current. Five additions close that gap.
Account history. Everything in the dossier is public. The private half is what we already know: prior opportunities, why they were lost, who championed us, who blocked us, and how long ago anyone last spoke. A rep who walks in without that is walking into an ambush. web/crm/provider.mjs is a seam, not an integration: it reads synthetic fixtures today and a real Salesforce adapter drops into the same shape. The panel is labeled INTERNAL DATA and, while the fixtures are in use, carries a loud sample warning. Every person in the fixture is fictional on purpose, because inventing deal history about a named real executive is the same sin as fabricating a citation.
The useful part is not the table, it is the overlap. The server cross-references who we have already talked to against who the research says matters now, and reports the honest result — usually that none of the prior contacts are in the current buying center, which means rapport does not transfer and the account needs re-earning rather than resuming.
Point of view. Why anything, why Factory, why now, and the ask. Each pillar carries cited evidence and a required risk field stating how a sharp buyer would push back, because a pillar with no stated counter-argument has not been thought through. "Why anything" is not allowed to mention Factory. "Why now" has to rest on a dated catalyst or openly admit there is no urgency; manufactured urgency is the fastest way to lose a senior buyer.
The section that earns its place is Do not say: the claims a rep would be tempted to make under pressure that the evidence does not support. On the first real run it caught the tool quoting its own illustrative 30% planning assumption back as if it were the prospect's baseline. That is the exact failure that ends a deal in the room, and it is now printed in red above the outreach.
Warm paths. Ranked routes in that beat a cold email, with the provenance of each one marked: CRM for a prior relationship, public for a cited customer reference, alumnus, or board overlap. Speculative paths are labeled speculative rather than dressed up.
Outreach. A send-ready sequence for the top two or three targets only, because a ten-person sequence never gets sent. No placeholders and no merge fields, and every factual sentence in a body carries the source id behind it — which is what stops a sequence from confidently misquoting a prospect's own earnings back at them.
Freshness and what changed. Every dossier states its age and turns amber, then red. Re-running archives the previous copy first, so the second visit opens with a diff: fit score moved, catalysts added, people added or gone, source count changed. Research that does not advertise its age invites someone to quote a figure that stopped being true two quarters ago.
The rule holding all of this together: public research, internal CRM data, and generated argument are three different kinds of claim and are never styled as one. Blending them is how a rep ends up presenting an internal guess to a buyer as a sourced fact.
Three design decisions carry the weight:
The agent writes JSON, not prose. A dossier is validated against dossier.schema.json before it renders. Scraped narrative would have been faster to build and impossible to lay out honestly — you cannot put a fit score in a progress bar if the score only exists inside a paragraph. Structure is also what makes the output checkable.
Citations are referential, not decorative. Every claim carries a sourceId that must resolve against the sources array. The server checks for dangling references on load and the UI banners them. A citation that points at nothing is a bug, and it surfaces as one.
Link verification is in the UI. The Sources panel has a button that resolves all of them and reports OK / bot-blocked / soft-404 / dead. It is the same verify-sources.ps1 the CLI gate runs. Being able to press that button in front of a buyer is the entire point.
# zero dependencies: Node built-ins only, so it always starts
web/server.mjs # jobs, schema validation, source-ref integrity, verifier endpoint
web/public/ # index.html, styles.css, app.js
web/data/<slug>.json # generated dossiers# 1. show the inputs: two skills, one subagent, one runner
Get-ChildItem -Recurse .factory
# 2. show the prompt without spending a run
.\scripts\new-account.ps1 -Account "Kroger" -DryRun
# 3. run it live on an account nobody has prepped
.\scripts\new-account.ps1 -Account "Lowe's" -Mode both
# 4. read the artifact and the gate result
Get-Content .\briefs\lowe-s\brief.md
# 5. or do the same thing in a browser, and let them pick the company
node web\server.mjsDroid CLI (found via DROID_BIN, ~/bin/droid.exe, or PATH), PowerShell 5.1 or later, and network access for research.