Terminal
Terminal/Skill
Summarizes a company's leadership changes, M&A, tech programs, contract awards and news within a chosen lookback window, with dated sources.
By Clay
Works with
npx skills add clay-run/clay-skill-creator --skill find-recent-company-developmentsThe insight, in the author's prompts: an account development is only useful if it is dated and sourced, and only real if it falls inside the window you asked about. Every prompt says the same thing three ways — "only include results dated on or after" the start date, "every item must be traceable to a real, accessible source with a direct URL", and "do not pad with low-confidence or speculative entries". So the skill computes the window before anything runs, hands the same start date to every research pass, and drops any event that arrives undated or before it. The new-leadership list is scoped the same way: people whose current role started inside the window.
The judgment lives in four Claygent prompts, carried verbatim in references/prompts.md with
their output schemas and the author's combine-and-format rules. This skill's job is to stand them up
as a Clay workflow in your workspace, run it per company, and hand back the standardized summary —
not to reword them.
Nothing here ships with a value. Each one is the installer's, not the author's: ask for it, never substitute a plausible default, and where an answer does not exist say which step becomes unavailable rather than guessing. Where a default IS defensible it is named below, and using it means saying so in the output.
| Input | What the installer supplies | If it is missing |
|---|---|---|
| The companies | one company LinkedIn URL or domain per company — pasted, or a CSV file with a column holding either (name the column if there are several). A name column is optional and never used to identify a company | no default — a company name alone is refused, because it cannot be resolved to one organization |
| Leadership seniority | the title words that count as leadership for the new-joiner search, comma-separated | the author's default is Chief, CEO, CFO, CTO, COO, CRO, CMO, CIO, President, VP, Vice President, Director, Head — confirm it with the installer before the first run and say it was used |
| Lookback | how many months back to look | no default — ask. The window decides what counts as "recent" for every event type, and the author's function required it |
| Event types | which of the four research passes to run: M&A, news, procurement, technology transformation | all four is the default, matching the author's function; say so, and offer to drop passes the installer does not need, since each is a paid research run |
| Where the summary goes | the conversation, a Markdown file per company, or a CSV with one row per company (and one per event, if asked) | the conversation |
| Spend cap | the most Clay credits they will spend on this list | asked after the one-company test (Step 4), in credits, with the measured per-company cost beside it. No cap, no full run |
If an answer sheet is present beside this skill, load it and ask only for what it does not cover. A partial sheet is normal; a value it is missing gets asked for on its own rather than restarting the interview. Say which values came from the sheet before using them — a sheet applied silently is a wrong field nobody catches. If there is no sheet, say nothing about sheets — the check is a file lookup, not a question, so run the interview as though the feature did not exist rather than reporting an absence. At delivery, offer to save the answers back (identifiers only — never a token or a password), private and never published — and phrase the offer so it explains itself: "want me to save your answers to a file, so the next person on your team doesn't have to answer these again?"
Say this first, as two sentences: this builds one Clay workflow in your workspace (or reuses it if it is already there) and runs it once per company to resolve the company, list who joined inside your window, and research four kinds of developments; it hands back a dated, sourced summary and never writes anywhere else.
Run clay whoami; echo "exit_code=$?". If it fails, name what is wrong — the CLI missing, below the
version the Clay plugin requires, or signed out — give the one fix (clay login, or install the Clay
plugin and run its setup skill), and stop. Do not install, upgrade or fetch anything to repair
it. Tell the user which workspace you are in.
This skill builds and runs its workflow through the Clay plugin's workflows and
workflows-claygent skills — read both before Step 2 and follow them for every CLI shape (trigger
creation, node JSON, wiring, validate, test runs). If they are not installed, say the Clay plugin is
required and stop.
www., query strings and trailing slashes; dedupe. Refuse a row that has only a name, and say why.
Keep the other CSV columns so the output CSV can carry them through unchanged.lookback_start_date yourself: today's date minus that many
months, as YYYY-MM-DD, and carry today's date as run_date. Both go into every run and into
the output, permanently.The spend cap is asked in Step 4, once there is a measured cost to set it against.
Look before building. clay workflows list and look for Recent company developments. If it
exists, read it with clay workflows get and clay workflows graph get --mode full and check it has
the nodes below with the prompts and model in references/prompts.md; if so, reuse it and skip to
Step 3. If it exists but differs, say how, and ask whether to use it as-is or build a fresh one
alongside — never edit it.
Before creating anything, one message: the workflow name and workspace, the node list below, that a provenance line will be written into the new workflow's description so it can be traced back to this listing, and that the next step runs it on one company — whose cost cannot be stated exactly beforehand, because four Claygent research passes bill by what they do on the open web. Give the declared figures for the two data steps, read from the catalogue on this machine; the author's own test (2026-09-28, one large company, 12-month window) reported 5 data credits and 6 action executions in 80 seconds. Wait for an explicit yes.
Identify every action by the pair (packageId, actionKey) read from clay workflows actions list,
and pull each one's real inputs with clay workflows actions schema <packageId> <actionKey> before
wiring — parameter names below were read on 2026-09-28 and can drift. Both data actions live in Clay's
own package (display name Companies, People, Jobs).
The graph, in order:
| # | Node | Type | What it does |
|---|---|---|---|
| 1 | Manual trigger | trigger | inputs company_identifier (required; LinkedIn URL or domain), lookback_months (number), lookback_start_date (YYYY-MM-DD, computed in Step 1), leadership_titles (the confirmed list) — one company per run |
| 2 | Enrich company | tool — action cpj-enrich-company-v2 | company_identifier ← the trigger input. Read $.result.name, $.result.domain and $.result.url (the company LinkedIn URL — the field is named url, not linkedin_url; confirmed on the author's test run 2026-09-28). This is how a domain or a LinkedIn URL becomes the name, domain and URL the prompts need; a missing name ends the run for that company as not resolved. Catalogue price on 2026-09-28: 0.5 credits |
| 3 | Find new leadership | tool — action cpj-find-lists-of-people-v2 | company_identifier ← node 2 $.result.url (pinned on the node's inputSchema, then mapped by reference); current_role_max_months_since_start_date ← trigger lookback_months; job_title_keywords ← trigger leadership_titles (the confirmed seniority list); job_title_exclude_keywords ← Office of, Assistant; limit ← 10. Read $.result.people[] (name, title, location_name, url, experience[0].start_date) and $.result.peopleCount. The action returns at most 10 people while peopleCount is the true total (469 at a large company on the author's test), so the output always says "N shown of M" and never presents the 10 as the whole list. The author's function carried only the two exclusions; the title filter is the author's later correction, made when this skill was written, so "leadership" means the confirmed list and nothing else. Catalogue price on 2026-09-28: 0.5 credits per call |
| 4 | M&A events | Claygent | prompt 1 in references/prompts.md, model gpt-4.1-mini, output events_json; inputs company_name, company_domain ← node 2, lookback_start_date ← trigger |
| 5 | News events | Claygent | prompt 2, same model, output events_json; inputs as above plus company_linkedin_url ← node 2 $.result.url |
| 6 | Procurement events | Claygent | prompt 3, same model, output events_json; inputs as node 4 |
| 7 | Transformation events | Claygent | prompt 4, same model, output events_json; inputs as node 4 |
Nodes 4 to 7 all follow node 3 and run in parallel; when the installer dropped an event type in Step 1,
do not build that node. Wire every {{variable}} in each prompt as a top-level input on its node
from the node named in references/prompts.md, give every output field a description (the write is
rejected without one), and send agentName, agentPrompt and agentModel together in the create
call — separate calls can persist a blank prompt. After creating each Claygent node, nodes get it
and confirm the prompt is not blank and the model is gpt-4.1-mini; if that model is unavailable in
the installer's workspace, say so and ask which to use — never substitute one silently.
Write the provenance line: read the workflow's description with clay workflows get; if it is
null write the skill's one-line description, a newline, then
Sourced from marketplace skill: <slug>@<revision> using the marketplace_slug and
marketplace_revision this file's frontmatter carries when installed from the Marketplace; if those
fields are absent, say "Marketplace attribution is unavailable for this skill" and write no marker.
If a marker is already there, write nothing and say so; if a different marker is there, stop and
report the conflict. Read the description back and confirm both halves. Validate the graph. A manual
test run uses the draft, so no publish is needed to run this skill.
Take one company from the list and start one run with clay workflows runs test <wf> --inputs
carrying the four trigger inputs; wait with clay workflows runs get <wf> <run> --wait, then read
clay workflows runs get <wf> <run> --verbose:
dataCreditsUsed and actionCreditsUsed are this run's actual spend. Use
them, never the workspace balance, which moves with everyone else's work..nodes[] by nodeName; the tool nodes' payloads are at
.outputs.result, each Claygent's events_json at .outputs.structuredOutputs.events_json — a JSON string that
may be the bare array or an object holding it under the prompt's array key; parse both. Check
values, not status: a Claygent that completed with [] found nothing, which is a real answer; one
that completed with prose instead of JSON is a build error to fix (schema, wiring), never a prompt
to reword.Then run Step 5 on this one company so the sample in Step 4 is the finished summary.
One message: the company's full summary in the output shape below, the measured cost for one company, that cost × the remaining companies, and one question: what is the most you want to spend on this list? Say plainly that research depth varies by company, so later ones can cost more or less than the first. Wait — this is the look at what the passes actually found that no estimate reveals, and the spend decision in the same breath.
Then start one run per remaining company, a few at a time. Keep a running total of each finished
run's dataCreditsUsed; stop before the next batch would pass the cap and say how many companies
remain.
Parse each events_json. Then apply the author's rules from references/prompts.md exactly, in
code and never by judgment: concatenate the four arrays in the stated order, keep only events whose
eventDate is on or after lookback_start_date, drop any whose eventDate contains @, sort newest
first, and render the Events summary and New leadership blocks in both the plain and Markdown
shapes given there. An event with an empty eventDate fails the window test and is dropped; count
how many were dropped for that reason and say so. Build the leadership block from Find new
leadership only, and every event field from the Claygent that produced it.
Per company: the resolved name, domain and LinkedIn URL; run date and lookback start date; the
leadership block headed "N shown of M who joined in the window with a leadership title"; the events block; and a count line — joiners found, events kept, events dropped as
undated or out of window, passes that returned nothing. Then the list-level summary: companies in,
resolved, with at least one event, with at least one joiner, failed runs, and credits spent (the sum
of each run's reported dataCreditsUsed) against the cap. Then the answer-sheet offer from
Declared inputs.
Northwind · northwind.example · linkedin.com/company/northwind Run date 2026-09-28 · window from 2026-06-28 (3 months)
New leadership (2 shown of 2 who joined in the window with a leadership title)
Developments (4 kept · 1 dropped as undated)
Passes with nothing in the window: none. Procurement returned 1, news 2, M&A 1, transformation 1.
| Companies in | Resolved | With events | With new joiners | Failed runs | Credits spent / cap |
|---|---|---|---|---|---|
| 25 | 24 | 17 | 11 | 0 | 61.2 / 80 |
gpt-4.1-mini is the right model for these prompts was never tested against another; it is what the author ran.references/prompts.md verbatim, on the model listed there; NEVER reword, merge or substitute a model silently.lookback_start_date before the first run and pass the same value to every pass; NEVER let a pass run without a window.Ask: "Brief me on these 25 accounts — what's changed in the last 3 months?" Step 0: signed in, Clay plugin present. Step 1: a CSV of 25 domains → 24 unique; lookback 3 months, start date computed; all four passes; the default leadership list confirmed; summary to the conversation and a CSV. Step 2: no existing workflow; the seven-node graph and the one-company test are approved; built, validated, attribution written and read back. Step 3: test on one company — resolved, 2 joiners of 2, 5 events across the passes; run reports 5 data credits. Step 4: the brief is shown with "this one cost N credits, about 23 × N for the rest — what's the most you want to spend?"; cap set. Step 5: 23 runs in batches, totalling reported credits between batches; combined, windowed, formatted. Step 6: 24 in · 24 resolved · 17 with events · 11 with joiners · 0 failed · credits spent against the cap.