# Research > Account Research — point an agent team at a target company and get a sourced dossier. Use when I say "research ", "account research on ", "build a dossier on ", "what's the angle on ", or paste a company name/domain and ask for research. Runs researcher → signal-scanner → composer and writes the dossier to my-work/account-research/. - Skill: `bluerock-io/research` (Agent Skill) - Install (CLI): `npx skillmds@latest add bluerock-io/research` - Raw SKILL.md: https://api.skillmd.com/api/skills/bluerock-io/research/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: bluerock-io (https://skillmd.com/u/bluerock-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/bluerock-io/research --- Run the Account Research team on a target company and produce the dossier — the shareable artifact. You orchestrate three agents; they do the work. ## First — anchor to the project The dossier, its working folder, and the `voice.md` / `objectives.md` the run reads all live in the builder's project — the folder holding `CLAUDE.md` and `design/` side by side. **The session usually starts inside it already**, so check that first. Its folder is normally named `my-workspace`, but identify it by that signature, not by its name (older setups have it one level below the home folder under a name the builder chose). Before you start: run `ls`. See `CLAUDE.md` and `design/` side by side? You're in the project. If not, find it: `ls */CLAUDE.md`, then `ls ~/*/CLAUDE.md`, else `find ~ -maxdepth 3 -path '*/design/dashboard.html'`. `cd` into that folder, capture its **absolute path** with `pwd`, and use that full path for the working folder and the files you read (e.g. `/home/you/maria-hub/my-work/…`). Can't find it? Ask the builder where they cloned their project. Working outside it means the dossier lands where they can't find it and the voice tuning silently skips. ## Setup 1. **Get the target.** A company name (and any hint — sector, region, or domain — to disambiguate). If it's ambiguous and you can't tell which company, ask one question before spending a run. 2. **Make the working folder.** Slugify the name → `/my-work/account-research//`, using the project's **absolute path** from the anchoring step above. Create it. Everything for this run lives here; `my-work/` is builder-owned and never overwritten. ## Run the team, in order **Pass the working folder's ABSOLUTE path to every agent, in the dispatch prompt itself.** A subagent starts wherever the session started — usually the home folder, one level above the project — and it has no shell to find its way back. Handed a relative path, it writes one level too high and the run is silently lost. The path you pass is the handoff. 3. **Dispatch `researcher`** with the company (+ hint) and the working folder's absolute path. It writes `profile.md` (who they are, sourced). Wait for it. 4. **Dispatch `signal-scanner`** with the same absolute folder path. It reads `profile.md` and writes `signals.md` (recent, dated, sourced signals). Wait for it. 5. **Dispatch `composer`** with the same absolute folder path. It reads `profile.md` + `signals.md`, plus `voice.md` + `objectives.md` from the project root if present, and writes the dossier `.md` (overview → signals → strategic angles, in my voice). ## Finish 6. **Render the dossier as a Claude Artifact — the polished work product.** The markdown file (`.md`) is the source of record and what the builder pushes to their repo; the artifact is the *shareable, polished view* — and seeing a clean, finished work product is the aha. Publish the dossier as a **single self-contained HTML Claude Artifact**: a clean report page (company header with the `Verified · N sources` line, the overview, the dated signals, the strategic angles with their hook / why-now / fit, and the `Built with BlueRock · Account Research` footer). CSP-safe, so: - **Inline everything**; **no external requests** — no CDN, no web fonts, no remote images (the artifact CSP blocks them). System-font fallbacks (`Georgia, serif` for headings, `system-ui, sans-serif` for body) in the cool-paper light look. - **Keep every citation** — sources carry through to the rendered report exactly as in the markdown; don't launder sourced facts into unsourced claims. - Read-only. No CTAs — it's a work product to share, not a conversion surface. - Print-friendly (it's the thing they'll want to send someone). If publishing an artifact isn't available in my environment, don't block: the markdown dossier is written and saved — tell me the artifact couldn't be created and give me the path. 7. **Report:** the dossier path, the one-line best angle the composer surfaced, and the artifact (or the fallback note). Don't reprint the whole dossier. ## Rules - **One company per run.** Don't let a run sprawl across companies. - **Honest sourcing.** This is public-web research (the agents use WebSearch / WebFetch), not a paid data feed — the dossier cites sources and marks gaps. Say so if a section came back thin; never present guesses as findings. - **Tune to me.** If `voice.md` / `objectives.md` exist, the dossier should sound like me and rank angles by what I actually care about. If they don't, run anyway and suggest running `/onboard` (or saying "onboard me") to make the next one sharper. - If `researcher` or `signal-scanner` returns almost nothing (private company, no public footprint), tell me plainly rather than padding the dossier.