Skip to content

JasonColapietro/suede-creator-skills / skills/suede-prospecting

suede-prospecting

Verified

Build a source-backed, scored prospect list across SaaS, B2B, local business, or early demand signals. Use when asked to "find our first design partners", define an ICP, enrich leads, run gap plays or warm-path network mining, or document disqualification evidence. Includes source dates, confidence, licensed-data checks, and a qualification handoff. NOT FOR: sending outreach (use suede-cold-email), CRM routing (use suede-revops), or competitor profiles (use suede-competitor-profiling).

FileKindSize
CARD.mdmarkdown4.4 KB
SKILL.mdskill md19.3 KB
agents/openai.yamlconfig551 B
evals/evals.jsonconfig13.8 KB
references/b2b-prospecting.mdmarkdown5.5 KB
references/collector.mdmarkdown9.5 KB
references/compliance.mdmarkdown7.1 KB
references/data-sources.mdmarkdown14.5 KB
references/demand-signals.mdmarkdown9.4 KB
references/local-prospecting.mdmarkdown8.1 KB
references/saas-prospecting.mdmarkdown11.0 KB
references/worked-handoff.mdmarkdown1.6 KB
scripts/collect.mjsscript67.7 KB

references/collector.md

View on GitHub
# Free Account Signals (`collect.mjs`)

`scripts/collect.mjs` pulls free, dated public evidence about a target account: its technographics, hiring, new properties, developer adoption, and ad-library links. Node 22 or later, no dependencies, no API keys. Use it before paying for a tech-lookup or enrichment credit, and keep the paid or manual routes in `data-sources.md` as the fallback whenever a command is blocked or a signal is not covered.

It studies companies only. It never collects people: reviewer names are never read, and names, emails, and handles in job posts are never written. Contacts still come from the Phase 3 contact rules in SKILL.md and `compliance.md`, never from this script.

## Commands for prospecting

Run these per account after the ICP is set (Phase 1) and the candidate list exists (Phase 2). They qualify known accounts; they do not discover new ones.

```bash
node <skill-dir>/scripts/collect.mjs dns <domain> --out prospects/<run>/<domain>/dns.csv --moves prospects/<run>/<domain>/signals.csv
node <skill-dir>/scripts/collect.mjs stack <domain> --months 12 --out prospects/<run>/<domain>/stack.csv --moves prospects/<run>/<domain>/signals.csv
node <skill-dir>/scripts/collect.mjs jobs <board-slug> --out prospects/<run>/<domain>/jobs.csv
node <skill-dir>/scripts/collect.mjs subdomains <domain> --out prospects/<run>/<domain>/subdomains.csv --moves prospects/<run>/<domain>/signals.csv
node <skill-dir>/scripts/collect.mjs adoption --npm <pkg> --github <owner/repo> --months 12 --out prospects/<run>/<domain>/adoption.csv
node <skill-dir>/scripts/collect.mjs ads "<Company>" --domain <domain> --out prospects/<run>/<domain>/ads.csv
```

Every command takes `--json` and `--help`. `--moves` appends dated rows (`date,channel,kind,move,reach,link,evidence`) to a per-account `signals.csv`, creating it with a header when absent and skipping rows already present by date, channel, kind, and move, so a rerun does not duplicate signals. Bad input exits 1 with `collect: <message>`.

| Command | Source | Prospecting signal | Evidence label |
|---|---|---|---|
| `dns` | The system DNS resolver | Email senders (MX, SPF), verified tools (TXT verification records), and marketing subdomain platforms (CNAME). A free technographic read: which CRM, email, or support vendor the domain is set up for | Observed, dated the pull day; what a record implies about use is Inferred |
| `stack` | Monthly Wayback captures of the homepage | First and last month each of ~64 tools appears (ad pixels, analytics, testing, referral, lifecycle, ecommerce, payments, call tracking). A recent `first_seen` is a stack change; a `last_seen` well before the latest capture can mean a tool was removed, a possible vendor switch | Ad pixels Inferred (`ad-start`), other tools Observed (`tool-added`) |
| `jobs` | Public Greenhouse, Lever, or Ashby boards | Open roles by function, posted date, and marketing channels named in each post (seo, paid, lifecycle, content, partnerships, and others). Roles in your buyer's function or channel are a buying trigger | Observed; the strategy they imply is Inferred |
| `subdomains` | crt.sh certificate logs | Earliest certificate per subdomain: a new `docs.`, `app.`, `shop.`, or country subdomain points to a launch or expansion | Observed; what a subdomain is for stays Inferred until a page confirms it |
| `adoption` | npm, PyPI Stats, GitHub (unauthenticated) | Monthly downloads and star counts for a developer-tool account's own package or repo: a rising curve is a growth signal | Observed (third-party counts) |
| `ads` | None: prints ad-library search links | Links for Meta, Google, LinkedIn, and TikTok libraries to check whether the account is spending on paid | Record the library, filter, and date with each finding |

The board slug for `jobs` is not always the domain. Find it in the careers page's link to `boards.greenhouse.io/<slug>`, `jobs.lever.co/<slug>`, or `jobs.ashbyhq.com/<slug>`. When `jobs` reports no board, the company may use another applicant system: check the careers page before recording "no open roles."

## Batch use across a candidate list

Run the cheap commands across the whole list and the slow ones only on the shortlist:

1. **Every candidate:** `dns` (local resolver, seconds per domain) and `jobs` (three board lookups per slug).
2. **Shortlist only:** `stack` and `subdomains`. `stack` spaces archive.org requests at least 1.5 seconds apart and caps each run at `--max-requests` (default 40). A 24-month run takes about 25 requests and 40 seconds; `--months 12` needs fewer and usually covers a buying-signal window. crt.sh is often slow.
3. **Developer-tool accounts only:** `adoption`, when the account publishes a package or public repo.
4. **Accounts where paid spend matters to the offer:** `ads`, then open each link by hand.

A shell loop over a `domains.txt` file works for the first step, writing each account's files under `prospects/<run>/<domain>/`. Run accounts one after another, not in parallel, so the pacing holds.

## Where the output goes in this workflow

- **Phase 1 (ICP):** use `dns` and `stack` columns to write technographic pass or fail criteria the collector can check, such as "uses a complementary CRM" or "runs a Meta pixel."
- **Phase 3 (qualify):** each collector value is one cited source. Put the capture or record link from the output in the row's `source_urls` and the pull date in `verified_date`. `dns` and `stack` read different sources (live DNS records versus archived page code), so agreement between them on the same tool can count as two independent sources for High confidence. A collector value alone, with no corroboration, is Medium at most.
- **Phase 4 (score):** a dated `jobs` role in your buyer's function, a recent `stack` change, or a new `subdomains` launch can serve as the buying signal. Older than the ICP's signal window, it supports Warm, not Hot.
- **Phase 5 (lead sheet):** stamp every collector value in the `signal` or `why_prospect` column with its command, source, and pull date, for example `Hiring 2 lifecycle marketers (jobs: greenhouse, pulled 2026-11-12)`. Keep the per-account `signals.csv` as the lineage file for the sheet.

## Evidence rules

- **Dated public observation.** Collector output is what a public source showed on the pull date. It is not licensed data, not a verified fact about the company's intent, and not a contact record. No license check is needed to run it, but the fixed hosts' own terms still apply and the output is for the user's own prospecting, never for resale.
- **Blocks mean "not collected."** When a command prints `blocked by <host> (<status>): use an archived copy or a manual export`, record that signal as `not collected (blocked by <host>, <date>)` and fall back to a manual or licensed source. Never retry around a block or read it as a negative result.
- **Absence is not disqualification.** No pixel in `stack`, no board in `jobs`, or no ads found is not evidence on its own. Disqualify only on a positive finding (for example, `dns` shows a direct competitor's verification record and `stack` shows its script) or on a manual check that confirms the absence, and cite both in the Skip row.
- **Partial runs say so.** A run that hits `--max-requests` prints that its output is partial. Carry that note into the open questions.

## How it behaves

- **Identifies itself.** User-Agent `suede-collect/1.0 (+https://skills.suedeai.ai)`. It never imitates a browser and reads no environment variables or tokens.
- **Contacts only fixed public hosts:** web.archive.org, crt.sh, boards-api.greenhouse.io, api.lever.co, api.ashbyhq.com, itunes.apple.com, api.npmjs.org, pypistats.org, api.github.com (unauthenticated), and www.youtube.com, plus the target domain for `site` (robots.txt and sitemaps only) and the exact host of a podcast feed you pass. Each redirect hop is checked, and an off-list hop stops with a message.
- **Stops at blocks.** A 401, 403, or 429, or a robots.txt Disallow, stops that source with `blocked by <host> (<status>): use an archived copy or a manual export`. One retry on a network error or 5xx, never more.
- **Paces the archive.** At least 1.5 seconds between archive.org requests and `--max-requests` (default 40) per command; a capped run says its output is partial.
- **Keeps people out.** Reviewer names are never read, and names, emails, and handles in job posts are never written. Star history reads only star dates, never who starred.
- **Writes only where told.** Files go only to the `--out` and `--moves` paths.

The script also has `site`, `history`, `hero`, `reviews`, `youtube`, and `podcast` commands built for growth teardowns. They are rarely needed for qualification; run `--help` for their usage.

## Known gaps to state in the open questions

- `stack` reads only the homepage and cannot see tools loaded through a tag manager or behind a consent banner. Tools present in the first capture are listed "at start" with no move, and `first_seen` is the first monthly capture containing the script, so the tool may predate it.
- `dns` findings are undated; the row carries the pull date, not the date the tool was adopted. A verification record can outlive the tool.
- `jobs` covers only public Greenhouse, Lever, and Ashby boards. Greenhouse posted dates fall back to the last update.
- `ads` fetches nothing. Open each link, and write "no ads found in <library, filter, date>" when empty.
- PyPI Stats keeps about 180 days. Unauthenticated GitHub allows 60 requests an hour and lists only the first 40,000 stargazers, so stars over time are interpolated from sampled pages.
- Download counts include CI and bots. Use them as a curve shape, not a user count.
- crt.sh is often slow; a timeout prints a retry message.
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 323kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall