Terminal
Terminal/Skill
Sorts a list of target accounts into weekly work buckets, with a sourced account context and a draft message for each one worth working.
By Clay
Works with
npx skills add clay-run/clay-skill-creator --skill account-priority-bucketsThe insight: PAIN signals require a verified buyer quote. If you can't quote them, you don't have a PAIN signal. You have a MOMENT signal or a firmographic guess. Most account scoring blends firmographics, activity and inferred intent into one number, and reps then personalize outreach around pain nobody said out loud. Buyers answer with "how did you get that idea?" This skill keeps the four questions apart: FIT decides whether to pursue, MOMENT and CONFIDENCE decide how urgently, and PAIN decides readiness. Priority = FIT × MOMENT × CONFIDENCE. PAIN determines readiness state.
What follows from it: an account with a new executive, fresh funding and a strong fit but no quote goes to Research Priority, never to Action. The job there is to find the pain before anyone sends. And every Account Context carries a written reason the account might not buy, because a hypothesis without one is a sales pitch.
AI proposes. Evidence supports. Humans govern. Outcomes teach. This skill proposes and shows its evidence. A human decides whether to act on any of it.
Nothing here ships with a value except where a default is named. Each input is the installer'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 used, say so in the output.
| Input | What the installer supplies | If it is missing |
|---|---|---|
| Account list | CSV, table or pasted list; minimum one company domain per row | no default; there is nothing to grade |
| ICP band | employee band (floor and ceiling), revenue band, top industries, regions, replacement targets (tools you displace) | FIT cannot be scored; stop and ask. If they have no evidence-based ICP yet, say that building one comes first and do not grade |
| Firmographic gate | which ICP dimensions must be true for an account to count at all (for example employee band and region) | treat no dimension as a gate and say so in the output |
| FIT weights | relative weight of each ICP dimension | propose equal weights, get approval, and label them as unapproved defaults if the installer declines to set them |
| High FIT cutoff | the Fit Score (0 to 10) at or above which an account counts as High FIT | no default; ask. Without it no account can leave Deprioritize |
| Personas | the champion and economic-buyer titles per motion (for example VP Analytics, CTO) | the new-executive signal cannot run; its weight stays in the output as unobserved |
| MOMENT window | how many days back a change still counts as "recent" | no default; ask. Undated events never count, whatever the window |
| MOMENT weights | points per signal type | default is the GTM AI OS signal library: public pain quote 10, competitor dissatisfaction 8, new executive 6, job posting 5, tech installed 4, funding 3, engagement 3. Say the default is in use |
| First-party evidence | anything the installer already holds per account: call quotes, CRM notes, review excerpts, web visits, content downloads, webinar attendance | PAIN can only come from public statements found in Step 5; engagement scores as unobserved |
| Where the evidence lives | which system holds call notes and engagement (CRM, call recorder, marketing automation, a spreadsheet) and how to export it | ask; never assume a CRM |
| Credit ceiling | maximum Clay credits this run may spend | ask before Step 4; stop at the ceiling |
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. Run the interview as though the feature did not exist. At delivery, offer to save the answers to a file beside the skill: it is not part of the skill, it holds their ICP band, cutoffs, windows and weights, it means they never answer these questions again, and a teammate given the file is asked only what it does not cover. It stays with them, is never submitted or published, and holds no passwords or API keys.
Say this to the installer before anything else: "This reads your account list and Clay enrichment. It writes a priority board and Account Context drafts here in the session. It never touches your CRM and never sends anything."
Run clay whoami; echo "exit=$?" and clay --version. If either fails, name the component, the
version required, and the one command that fixes it (clay login, or the Clay plugin's setup
skill), then stop. Do not install, upgrade or fetch anything to repair the platform. Name the
workspace clay whoami reports.
Where the work runs decides what it costs: the scoring, bucketing, hypotheses and message drafts run in this agent, at zero Clay credits. Only Step 5's enrichment calls spend credits.
Do not start a step before the steps above it have their answers. If a declared input is missing, ask for it. Never assume a default and continue.
www. and paths) and dedupe.
Report how many rows collapsed.FIT answers "should we pursue them?" Score each ICP dimension 0 or 1 (a band match is 1 inside the band, 0 outside), weight it, and renormalize over observed dimensions:
Fit Score = 10 × Σ(wᵢ × sᵢ) / Σ(observed wᵢ)
0 or false is evidence of failure and stays in both sums. Never test presence by
truthiness.PAIN answers "do they have a problem you solve?" It is Verified only when there is a verbatim statement from a named person at the account, with a source and a date: a call quote, a review, a post, a public statement. A paraphrase, a model's summary, or an article about the account is not PAIN.
MOMENT answers "why now?" Sum the weights of the signals that fired inside the window. Each signal type counts once per account however many times it fired; the evidence line lists every instance.
CONFIDENCE answers "how sure are we?" Every signal carries one tag:
| Level | Meaning |
|---|---|
| Verified | direct source, primary record: a recorded call quote, a review, a tech-stack detection |
| Supported | reliable secondary source: a news report, analytics, a provider's structured record |
| Inferred | pattern-matched or AI-generated: a title guess, a stack guess, a model's reading of a page |
| Unknown | no confidence claim possible |
Every signal counts toward MOMENT at full weight, whatever its tag. CONFIDENCE controls what a human may act on, not the score: Inferred and Unknown signals never go into personalized outreach or a message draft without human verification, and the board shows how many of each account's MOMENT points rest on Inferred or Unknown evidence, so a reader can see when a bucket depends on unverified signals.
Buckets. Evaluate in this order; the first match wins.
| Bucket | Definition | SLA |
|---|---|---|
| Deprioritize | Fit Score below the High FIT cutoff, or a failed gate | remove from the working list |
| Action Priority, high | High FIT + Verified PAIN + MOMENT 15 or more | same business day outreach |
| Action Priority, standard | High FIT + Verified PAIN + MOMENT 10 to 14 | within 48 hours |
| Research Priority | High FIT + MOMENT 10 or more + no Verified PAIN | investigate to find PAIN, then reclassify |
| Monitor | High FIT + MOMENT below 10, with or without Verified PAIN | nurture, watch for MOMENT; the board shows which Monitor accounts already have Verified PAIN |
Within a bucket, sort by Fit Score × MOMENT, highest first, ties broken by the most recent signal date.
http-api-v2 with a User-Agent) fails on a domain, flag the account "liveness
doubt" and exclude it from paid calls.Report what the free pass eliminated before any credit is spent.
Every call in this skill reads; nothing it does can mutate the installer's systems, so a real batch is safe.
clay routines list --limit 100,
clay routines get <id> for estimatedCreditCost (the list call omits it), and
clay workflows actions schema <packageId> <actionKey> for real inputs and any per-unit billing in
the parameter descriptions. Record the (packageId, actionKey) pair you resolved. If a named
function is absent, say so and leave that signal unobserved. Never substitute the nearest thing.metadata.upfrontCreditUsage.totalCost), never the balance delta.For every call: gate on payload values, never on run status. A complete run with an empty or absent
field is a miss, and a miss can still bill. Enforce the MOMENT window on every dated fact; an undated
event does not count.
| Signal | What runs | What goes in | What to verify | Confidence |
|---|---|---|---|---|
| FIT firmographics | managed Enrich Company function | domain | exact headcount where present, else the band string compared as a band; industry; HQ region; read website, not domain, for identity | Supported |
| Tech installed | managed Website Technology Stack function | domain | one of the installer's replacement targets appears in the returned string; drop metadata pseudo-entries; flag an exactly 8,192-character result as truncated | Verified, noted "may be historical" |
| New executive | people-index new-hire lookup, filtered to the installer's persona seniority and titles | domain, persona titles | a start date inside the window, title matches a persona | Supported |
| Job posting | a job-postings action that filters by seniority or department at the source and returns a true total (jobCount) | domain, persona function, window | at least one posting inside the window matching the persona function; record it as yes/no with the posting as evidence, never as a count | Supported |
| Funding | funding waterfall normalized through the typed company/fundingRound output | domain | a dated round inside the window where the account raised the money; an investment the account made in another company is not its funding | Supported |
| Public pain or competitor dissatisfaction | no Clay function discovers these. Two sources only: the installer's first-party evidence, and a public URL the installer names, fetched with scrape-website (outputFields: bodyText) after a free http-api-v2 status probe | the URL | a verbatim quote from a named person at the account, with a date; a 404 or soft-404 is a miss | Verified if the page is the speaker's own post or review, otherwise Supported |
| Engagement | installer's first-party data only; Clay has no identified-visitor function | the export from Step 1 | a dated event inside the window | Verified |
News results are topical, not evidence. Before any news item counts, check that it names the account as the subject (not a same-name company, a former executive's new venture, or a supplier), that the event date sits inside the window (a crawl date is not an event date), and dedupe one event reported by several outlets.
If a signal's function is unavailable, keep its weight visible as unobserved in the output. Dropping it quietly makes every MOMENT score look complete.
After the run, report credits actually spent against the estimate. If they differ, say by how much and why.
Apply Step 2 exactly. Each account ends in exactly one of: the five buckets, cannot score, or liveness doubt.
Check before delivering: any account whose evidence line names a mismatch with the ICP or gate is rescored or moved, never left in Action with a caveat. A reader acts on the order of the board. If the top account needs an explanation for why it does not fit, that is a scoring bug.
For every Action Priority and Research Priority account, write an Account Context with these fields: company and Fit Score with component scores; Verified PAIN in the buyer's words with speaker, source and date (or "none found" for Research); MOMENT changes and intent with dates and points; the confidence tag on every item; the hypothesis (why this account should care, which motion, which persona is champion and which is economic buyer); why we might be wrong (at least two concrete reasons: incumbent contract, no exec sponsor, recent budget cut, prior failed implementation, a signal that may be diligence rather than intent); the next action with its SLA.
For Action Priority accounts only, add a message draft in the architecture Signal → Pain → Consequence → Value → Proof → Action, built only from Verified and Supported evidence. The Value and Proof lines use the installer's own claims and case studies; if they supplied none, leave those lines as marked gaps rather than inventing a stat or a customer. One ask per message. Label every draft "for human review; not sent".
Deliver: the priority board, the Account Context files, the cannot-score list, the liveness-doubt list, the weights, windows and cutoffs used, and actual credits spent.
| Bucket | Account | Fit | MOMENT (Inferred/Unknown share) | PAIN | Top evidence | Next action |
|---|---|---|---|---|---|---|
| Action, high | northwind.example | 8.7 | 19 (3 of 19) | Verified | VP Analytics, 12 days ago: "spending 30% of our data team's time on reconciliations" | outreach today |
| Research | contoso.example | 8.1 | 11 (0 of 11) | none found | new Head of Data 38 days ago; Director of Data Engineering posting 9 days ago | find the pain, then reclassify |
| Deprioritize | fabrikam.example | 3.2 | n/a | n/a | 140 employees, below the 500 floor (gate) | remove from list |
Northwind Fintech (northwind.example). Fit Score 8.7/10 on weights band 35, industry 30, region 22, replacement target 13: employee band 1 (1,200, exact count), industry 1 (fintech), region 1 (Canada), replacement target 0 (none detected). PAIN, Verified: VP Analytics, LinkedIn post, 12 days ago: "We're spending 30% of our data team's time on spreadsheet reconciliations. That is not why I hired them." MOMENT 19 (3 points Inferred): public pain quote 10 (Verified); new Head of Data 44 days ago, joined from a data-warehouse vendor, 6 (Supported); Series C 97 days ago, 3, not counted, outside the 90-day window. Two pricing-page visits from the CTO, 3, Inferred from an IP match: counted, kept out of the draft until a human verifies it. Hypothesis: the new Head of Data brings modern infrastructure experience and the VP Analytics has named the pain in public; Analytics Modernization motion, VP Analytics as champion, CTO as economic buyer. Why we might be wrong: a new data leader may already have a preferred vendor from their last company; post-funding scaling often freezes new evaluations; the pricing visits are unverified and may be diligence. Next action: personal outreach to the VP Analytics today, multi-thread to the CTO and Head of Data within 14 days. Draft (for human review; not sent): Signal: your post about the team spending 30% of its time on reconciliations. Pain: that is the loss data teams describe before they switch. Consequence: board dashboards go stale and trust in the numbers erodes. Value: [installer's claim, not supplied]. Proof: [installer's case study, not supplied]. Action: worth 20 minutes next week to compare notes?
Ask: "Which of our 400 prioritized accounts should the team work this week?" Interview: employee band 500 to 2,500 (gate), regions US, Canada, UK (gate), industries SaaS, fintech, healthtech, replacement targets a legacy data platform and spreadsheet processes; High FIT cutoff 7.5; MOMENT window 90 days; personas VP Analytics and Head of Data (champion), CTO (economic buyer); default MOMENT weights; call quotes exported from their call recorder for 31 accounts; credit ceiling set by the installer.
Free pass: 400 rows, 6 duplicate domains collapsed to 394; 22 fail a gate on supplied headcount and go straight to Deprioritize; 3 fail the status probe and go to liveness doubt. 369 go to the batch. The 10-account batch shows one wrong-entity funding match (a same-prefix real-estate firm), which the entity check caught; the gate message carries the actual credits for 10 and the projection for 359. Approved.
Result: 11 Action high, 19 Action standard, 64 Research, 81 Monitor (23 with Verified PAIN), 188 Deprioritize, 6 cannot score (region unobserved). The 30 Action accounts each get an Account Context and a draft; the 64 Research accounts get an Account Context with "PAIN: none found" and the MOMENT evidence a rep should start from.