# Enrich Contacts > Find recent, dated, sourced signals for contacts you already have, and turn each one into an opening angle: how your offer connects to what just happened at their company. Optionally writes the personalization copy for each step of an email sequence. Takes the output of find-contacts unchanged. Use when you have people and need a reason to reach out, when a contact list has no talking points, or when asked to enrich contacts, find trigger events, research signals, or write personalization for a sequence. - Skill: `latentcast/enrich-contacts` (Agent Skill, multi-file: 8 files) - Install (CLI): `npx skillmds@latest add latentcast/enrich-contacts` - Raw SKILL.md: https://api.skillmd.com/api/skills/latentcast/enrich-contacts/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: LatentCast (https://skillmd.com/u/latentcast) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/latentcast/enrich-contacts --- # Enrich contacts **Contacts in. Dated, sourced signals out, each with an angle.** Optionally the personalization copy for your sequence. Takes `find-contacts` output unchanged. If you have a list from somewhere else, you need `recipient_id`, `full_name`, `company` and ideally `domain`. ## Two outputs, and only one is copy **Angles are research.** One line per signal: how your offer connects to this event. Always produced. No booking link, no call to action, no greeting. It is the note you would write to yourself before drafting. > Congratulate the second depot, then offer to hold cost-per-drop flat as the German routes come > online. **`perso_1` to `perso_4` are send-ready email copy.** Only produced if you ask, and only if the outreach profile has a `sequence` block. These need the voice rules; angles do not. > Congratulations on the Rotterdam site. We'd love to help Ostvale hold cost-per-drop flat as the > German routes come online. Neither is the canvas. `welcome_message` and `cta_message` are **viewing-page** copy, read beside the video player, and they belong to `build-personalization-canvas`. Three surfaces, three registers. Do not reuse one for another. ## What you need - An agent that can search the web and fetch pages. Prefer an entity-typed index; see [`references/providers.md`](references/providers.md). - An outreach profile for the signal ladder and, if you want copy, the sequence and voice blocks. - Contacts. This skill does not find people. ## Before spending anything **One research task per company, not per contact.** Two people at the same company share its signals, so researching per contact pays twice for the same answer. Group first, then fan out. Say what you understood and wait: > 43 companies behind 62 contacts, so 43 research tasks. Signals from the last 6 months, ranked > expansion, funding, milestone. Copy for all four sequence steps as well. Proceed? Never refuse a number the user asked for. Price it and let them decide. ## Signals Every signal passes all four gates or it does not ship. 1. **The source states it.** Not implied, not inferred from a headline. 2. **It is this company**, not a same-named one elsewhere. 3. **It is inside the window**, with the actual year established. Undated pages read as current and are often years old. 4. **The reading is correct.** A pilot is not a rollout. A funding round is not a valuation. "Named to a list" is not "won the award". Date everything to `YYYY-MM` at minimum with a `date_confidence`. `source_url` is the page that states it, never a search results page. **Keep one row per signal.** Do not flatten to `signal1/2/3` — that caps you at three and loses the fourth silently. Flattening is an export concern. **Rejecting a signal is a result.** Say which one and why. A run that returns no signal for a company is an honest answer, and roughly two-thirds of rows on a low-publishing segment will come back thin. That is the finding, not a failure. ## The angle One line per signal. What connects the event to what the seller does. - **Ground it in the signal.** If you cannot say how the event connects, the angle is "no clear connection", which is useful information. - **Never restate their strategy back to them.** They know what they do. - **No call to action, no link, no greeting.** That is copy, and it comes later if at all. ## Two more kinds of row: direction, and stack and market Events are not enough for the canvas. `build-personalization-canvas` writes **Strategic Priorities (L)** and **Relevance Signals (M)** only from the research notes it is handed, and it is told to invent nothing. A signals file that carries only dated events leaves those two cells with nothing to draw on, and the canvas build ends up re-researching them under time pressure. So this skill produces two more kinds of row, one research task per company, after the events: | `type` | What it is | Feeds | |---|---|---| | `direction` | Where the organisation says it is going, in its own words: strategy pages, annual reports, policy plans, leader interviews. Goals, growth direction, public commitments. | L, Strategic Priorities | | `stack & market` | What it runs on and sells to: platforms, hosting model, named vendors, AI in production or pilot, compliance regime, who the customers are, scale. | M, Relevance Signals | **They are sourced differently from events.** - **Ask for durable facts, not news.** The prompt says so explicitly, or the same index hands back the events you already have. - **The six-month window does not apply.** Direction is durable; a strategy plan from 2023 that still runs to 2027 is the right answer. Date every row anyway and set `date_confidence`, so a reader can see how old it is. - **Own words first.** A strategy page or an annual report beats a trade-press summary of it. Leader interviews are good because they are dated and quoted. - **The four gates still hold.** The source states it, it is this organisation, the date is established, the reading is correct. "Considering cloud for AI compute" is not "moved to cloud". - **Both kinds still get an `angle`.** For these it is a note on what scene 2 should lean on, not a reason to reach out. Two to four rows of each per company is plenty. Every organisation should end up with both, including the ones that returned no event: a company with nothing to congratulate still has a direction and a stack, and that is what the video talks about instead. Say so before you spend it. Two research tasks per company, not one: > 24 companies. 24 event tasks in the six-month window, then 24 direction-and-stack tasks with > no window. Proceed? An open item should say that direction and stack rows are not events and must not be used as the triggering event. The workbook builder tints them so they read differently on the sheet. ## Personalization copy, if asked Requires `sequence` in the outreach profile. **Ask how many emails they actually send** and what each step is for; do not assume. Four is a common shape and the example uses it, but a two-step sequence and a seven-step one are both normal and the copy has to know which it is writing for. One block per step, each with its own stated purpose. Rules that hold across all of them: - **Congratulate the event. Do not describe their business.** The personalization is in what you noticed, not in reciting facts they already know. - **Never criticise how they do it today.** - **Vary the outcome per recipient.** Swap the company name between two blocks; if both still read fine, neither is specific enough. - **Obey `voice.banned` and `voice.required`** from the profile. - **A step with nothing true to say is left blank**, and named in Open items. A generic block is worse than none, because it teaches the recipient the rest is generic too. ## Open items Raise one whenever a human needs to decide something before a send. They collect into their own sheet and are what gets walked through on a call. Real examples: two sources disagreeing on a date; a triggering event outside the campaign's agreed definition; a signal that would be strong if someone confirmed it; a contact whose title does not resolve. ## Output Records matching [`references/pipeline-contract.md`](references/pipeline-contract.md). Assemble with the workbook builder that ships with `find-contacts`: ```bash python3 ../find-contacts/scripts/build_workbook.py records.json enriched.xlsx --csv flat.csv ``` Four sheets: Companies, Contacts, Signals, Open items. The CSV is a lossy export for spreadsheet work. **Check your own links before you hand it over.** A source that does not open is an assertion, not a sourced fact, and clipped URLs are invisible to anything that only reads the records: ```bash python3 ../find-contacts/scripts/validate_sources.py records.json ``` It catches truncated links, search results pages, and per-seat links that resolve only for the account that exported them. ## Quality rules Shared with `find-contacts`, in [`references/quality-rules.md`](references/quality-rules.md). The two that matter most here: - **Six months is the default window, and widening it rarely helps.** Ranking does the work the window is meant to do. Widen it for a genuinely longer cycle, never because a run came back thin. - **A fetch failure is not evidence of non-existence.** Run a known-good control through the same tool before concluding a source is fake. ## What this is not A copywriter. Even with `sequence` configured, what comes back is a **draft that needs reading** before it goes anywhere near a send. Read ten at random and check the outcome actually differs between them.