Terminal
Terminal/Skill
Answers yes or no questions about a list of accounts, citing evidence for each answer and flagging accounts the evidence cannot settle.
By Clay
Works with
npx skills add clay-run/clay-skill-creator --skill account-intelligence-analystThe insight: the other research plays fail by leaving a field blank; this one fails by answering. A brief that cannot find a company's headcount ships an empty field, and the gap is visible. But a question — "are they building an AI team?" — always has a plausible answer available from adjacent evidence: the homepage says "AI-powered", so the answer looks like yes. That is marketing copy, not a hiring signal, and nothing in the output would reveal the substitution. A confident answer from thin evidence is indistinguishable from a confident answer from good evidence, unless the skill reports what it actually observed.
So the question is never answered directly. It is decomposed into proxies — specific
things that could be observed, each with a named arm that observes it and a declared cost —
and the answer is graded by which proxies actually fired, not by how the prose reads. An
account where nothing observable fired returns insufficient evidence, which is a real
answer and the most commonly correct one.
Two consequences that shape every step below:
creditCost: 0.8 on an addresses arm means 0.8 per
address, 8 at its default and 80 at its cap, with the multiplier stated only in a
parameter description. So cheapest-first ordering plus early exit is not tidiness, and
neither is reading the parameter docs before quoting: both decide whether the play is
affordable across 300 accounts or only across 30.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 question | a question they could answer per account by naming what would count as evidence | no default — a vague question answered anyway is the failure this play exists to prevent. Offer the tightened rewrite instead |
| The bar for a yes | what evidence counts: hiring, a shipped product, a named leader | ask — "building an AI team" has no verdict without it |
| The account list | accounts with domains; resolve names to domains first | no default. A wrong domain makes every proxy answer about a different company |
| Budget ceiling | credits available for paid proxies | state cost and wait for approval; step 3 exists for exactly this |
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?"
Run clay whoami; echo "exit_code=$?". If it fails or Clay tools are missing, run the Clay
plugin's setup skill, restart if it says to, and re-run this skill. Tell the user which
workspace they're in.
Resolve the arms you plan to use against the live catalog before promising any of them — provider availability and price both change:
clay workflows actions list > /tmp/catalog.json # cache once per run (~1.4 MB, 650+ actions)
references/proxy-catalog.md maps common question types to arms with their declared credit
costs, read from that catalog. Treat it as a starting point that must be re-resolved, never
as a price list to quote from memory.
A question this play can answer is one where you could say, per account, what would count as evidence. Test it out loud before spending anything:
| Asked | Verdict | Why |
|---|---|---|
| "Which accounts are building an AI team?" | usable after tightening | needs a bar: hiring? shipped product? named leader? |
| "Which accounts are expanding into EMEA?" | usable | offices, EMEA job posts, localized site |
| "Which accounts are a good fit?" | reject | that is ICP scoring — hand to account-tier-scoring |
| "Which accounts have budget this quarter?" | reject | no observable proxy exists at all |
If the question cannot be tightened into observable terms, say so and offer the rewrite. Do not accept a vague question and answer it anyway — that is the failure this play exists to prevent, arriving before any code runs.
Also collect: the account list (with domains — resolve names to domains first, a wrong domain makes every proxy answer about the wrong company), the bar for a yes, and the budget ceiling in credits.
Write the proxy set out and show it to the user before spending. For each proxy: what would be observed, which arm observes it, its declared per-account cost, and its weight.
Question: "Which of these accounts are building an AI team?"
proxy arm cost weight observable?
1. open AI/ML job postings cpj-find-lists-of-jobs 1 cr 3 yes
2. eng headcount growth cpj-get-company-employee-growth 1 cr 2 yes
3. AI/ML named in own site copy site fetch (free) 0 cr 1 yes
4. named AI leader hired news arm, date-windowed 1 cr 2 yes
5. internal roadmap approval — — 2 NO — declared unobservable
Three rules here, and each one exists because skipping it corrupts the arithmetic:
observable: no and its weight stays in
the denominator. It is why an honest maximum confidence can be below 100%.references/proxy-catalog.md has the price spreads that make this a
real decision.per account = Σ (cost of each proxy the user kept)
total = per account × accounts, before early exit
Summing creditCost is only valid for flat-priced arms. Before pricing any arm, read its
parameter descriptions for a per-unit rate ("credits for each …"). If there is one, price it
at the cap you intend to pass, and pass one — a per-unit arm with an unset cap is an unbounded
per-account cost. Where a proxy only needs a number, prefer a flat counting arm over a
per-unit listing arm; the "cheaper" per-unit arm is dearer past a handful of items.
Worked: the four observable proxies above are 1 + 1 + 0 + 1 = 3 credits per account; across 120 accounts that is 360 credits before early exit. Give the number, get a yes, then spend. Early exit (step 5) typically lands the real spend below the ceiling, but quote the ceiling — a quote that assumes best-case exit is not a quote.
A wrong entity makes every proxy answer a question about a different company, and it does it
invisibly. Per account: pre-gate the domain with a free DNS/status probe, then fetch the
company's own site. Confirm the site is the company the user means — name collisions,
holdings vs operating entities, franchises and rebrands are the standard traps. A dead or
parked domain is a finding, reported as anchor failed, not a silent skip and not a
no.
Never run a paid proxy against an unanchored account.
Order the kept proxies by cost ascending, free arms first. Per account, observe in that order and record for each proxy exactly one verdict:
| Verdict | Meaning |
|---|---|
supports | the observation was made and it supports a yes |
contradicts | the observation was made and it argues against |
not_observed | the arm ran and returned nothing usable, or the arm was not run |
Three traps verified on live payloads, each of which manufactures a verdict out of nothing:
null
and the other eight populated. Reading that null as 0% produces contradicts — a negative
verdict assembled from a data gap. A null horizon is not_observed for that horizon.jobCount of 33 available. "33 openings", "10 postings" and "4
roles" are all true and answer different questions. State the unit in the proxy definition, or
the same evidence supports either verdict.supports. Require the
observation to be about the claim, not about the company.not_observed is one bucket on purpose: "the arm found nothing" and "we stopped early" are
both absence of evidence, and splitting them invites treating the first as a soft
contradicts. An arm returning empty is never a contradiction. A company with no AI job
postings visible to one provider has not been shown to be hiring nobody.
Early exit: stop observing an account once the remaining unobserved weight cannot change
the answer — if |direction| already exceeds the total weight left to observe, further calls
cannot flip it. Record which proxies were skipped by early exit and why; the coverage figure
in step 6 is computed on what was observed either way, so early exit lowers coverage
honestly rather than hiding.
Same arithmetic family as account-tier-scoring on purpose — a house pattern, applied to a
different unit. There it is weighted ICP dimensions to a tier; here it is weighted proxies to
an answer.
coverage = Σ weight(proxies with supports or contradicts) / Σ weight(ALL proxies, including unobservable)
direction = Σ weight(supports) − Σ weight(contradicts)
| Condition | Answer |
|---|---|
coverage ≥ 0.50 and direction > 0 | yes |
coverage ≥ 0.50 and direction < 0 | no |
coverage < 0.50, or direction = 0 | insufficient evidence |
| Condition | Confidence |
|---|---|
answered, coverage ≥ 0.75, no contradicts | high |
| answered, otherwise | medium |
insufficient evidence | not reported — a confidence on a non-answer is theatre |
The three answer conditions are exhaustive and mutually exclusive: direction = 0 routes to
insufficient regardless of coverage, and coverage below 0.50 routes there regardless of
direction, so no account satisfies two rows and every account satisfies one. A tie is never
broken toward no — a split verdict is unsettled, and calling it no converts ignorance
into a negative claim about a real company.
Because unobservable proxies stay in the denominator, a question whose unobservable weight exceeds half the total cannot reach an answer for any account. That is not a bug to route around: it means the question is not answerable from observable evidence, and the honest output is to say so once, at the top, rather than 120 times.
Per account: domain · answer · confidence · coverage · direction then the proxies, each
with its verdict and — for anything that fired — a quote or a dated source link. Anchor
failures listed separately from insufficient-evidence, because they are different problems
with different fixes.
Roll up: counts by answer, the mean coverage, the proxies that fired most and least often, and the credits actually spent against the quote. Then the two lists that make this play trustworthy rather than impressive:
yes and can see exactly which two things fired, with links.insufficient evidence pile is large and nobody is embarrassed by it.contradicts, turning silence into no.not_observed as
contradicts.insufficient evidence when coverage < 0.50 or direction = 0; NEVER break a
tie toward no.creditCost across arms without checking which are flat.not_observed for that field; NEVER read a null as a zero and
never let one become a contradicts.Question: "Which of these 40 accounts are building an AI team?" Tightened bar, agreed with the user: an open AI/ML engineering role or a named AI leader hired in the last 180 days.
Proxy table as approved: AI/ML job postings (w 3, 1 cr) · engineering headcount growth (w 2, 1 cr) · AI/ML in own site copy (w 1, free) · named AI leader in dated news (w 2, 1 cr) · internal roadmap approval (w 2, unobservable). Total weight 10, observable weight 8, so maximum achievable coverage is 0.80 — stated up front.
Cost: 3 credits per account × 40 = 120 credits ceiling, approved. Two accounts fail the
free anchor (one parked domain, one holdings company whose operating entity is on a different
domain) and are reported as anchor failed, spending nothing.
One account: site copy mentions AI (supports, w 1) → job postings return two ML engineer
roles (supports, w 3) → direction is now 4 with 4 observable weight left; a full contradiction
of both remaining proxies would land at 0, which is insufficient, so the answer is not yet
settled and observation continues. Headcount growth is flat (contradicts, w 2), news finds
nothing (not_observed). coverage = (1+3+2)/10 = 0.60, direction = 4−2 = +2 → yes,
medium — medium rather than high because coverage is under 0.75 and one proxy contradicts.
Another: site copy silent, no postings, headcount flat, no news. Every arm ran and returned
nothing usable → coverage = 0.20 → insufficient evidence, not no. It appears on the
unsettled list with the note that a headcount-by-department arm at 8 credits would settle it.
Across the 40: 11 yes, 6 no, 21 insufficient, 2 anchor failed. Actual spend 94 credits against the 120 ceiling, the difference being early exit. The site-copy proxy fired on 34 of 38 anchored accounts and is flagged in the roll-up as near-useless discrimination — it says almost everyone mentions AI, which is a finding about the proxy, not about the market.