whetstone-signal
Skill8 starsChanged 3 days agoMITView on GitHub
Use this when ad campaigns that let the platform pick the audience (Meta Advantage+, Google Performance Max, Google Demand Gen, or similar) bring plenty of cheap leads, signups, or clicks that rarely turn into revenue. Helps you choose the conversion event the algorithm should chase, score and send back higher quality conversions, and check the results against your own backend numbers. Also use it when setting up a new automated campaign and deciding what to optimize for.
In this repository28 other files it gives its agents.
- design-a-grok-botSkill
- make-bot-uiSkill
- overheard-setupSkill
- overheardSkill
- routine-healthcheckSkill
- transcript-healthcheckSkill
- design-taste-frontendSkill
- diagram-designSkill
- motion-pickerSkill
- redesign-existing-projectsSkill
- engineering-playbookSkill
- show-meSkill
- chatgpt-images-browser-playbookSkill
- aeo-geo-siteSkill
- ai-citation-sourcesSkill
- ai-search-gap-loopSkill
- exact-match-intent-seoSkill
- markdown-twinSkill
- shorts-reframeSkill
- coachSkill
- draftSkill
- no-ai-slopSkill
- outlineSkill
- social-copywritingSkill
- source-checkSkill
- top-editSkill
- x-articleSkill
- youtube-niche-reportSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
21 rules. About 0 are common, 21 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈21
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 40
Ask for these before you recommend anything.
- Distinctive: line 60
Write each stage with its weekly count and the median days from click to that stage.
- Distinctive: line 80
Write one sentence that a salesperson and an engineer would both read the same way.
- Distinctive: line 88
Use behavior and stated answers.
- Distinctive: line 93
If the form is the main event, add one short question whose answer splits buyers from browsers.
- Distinctive: line 110
Start with 4 to 6 signals.
- Distinctive: line 123
Values: if you bid on value, send a number that tracks real worth.
- Distinctive: line 127
If the owner has paused all campaigns, do not start or re-enable any without asking them first.
- Distinctive: line 137
If people scoring 50 to 60 almost never buy, raise the cutoff.
- Distinctive: line 138
If the qualified event drops under the volume bar, lower the cutoff or widen the date window.
- Distinctive: line 142
Avoid changing the cutoff more than once a week.
- Distinctive: line 153
Keep a holdout: a geo, a time block, or a platform conversion lift test where the new setup is off.
- Distinctive: line 181
Add one form question: "How many people will use it?" Answers of 5 or more mark a fit lead.
- Distinctive: line 210
Add one question: postcode.
- Distinctive: line 222
Never buy data from data brokers or enrich leads with bought traits such as net worth, income, credit scores, marital status, or ethnicity.
- Distinctive: line 224
Never scrape, de-anonymize, or try to identify people who saw, liked, commented on, or clicked an ad beyond what the platform shares with …
- Distinctive: line 226
Never score, target, or exclude people using sensitive or protected traits: race, ethnicity, religion, health, sexual orientation, gender …
- Distinctive: line 230
Never use personal data for ad measurement or targeting on content made for children.
- Distinctive: line 232
Use only first-party data that people gave you with consent, for a purpose they were told about.
- Distinctive: line 235
Follow each platform's terms, including special ad categories for housing, jobs, credit, and social issues, where many targeting and …
- Distinctive: line 238
Never send a fake or inflated value to push the bidder.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Whetstone Signal
Automated ad campaigns pick the audience for you. They look for people similar to the ones who already converted. The setting still in your hands is what counts as a conversion. If that is "filled in a form," the system gets very good at finding form fillers, including the ones who never pick up the phone.
This skill sharpens that target. You pick an event closer to revenue, make sure it happens often enough for the platform to learn, and send it back with a value attached. Then you check your own sales numbers to confirm the change helped.
When to use
- Cost per lead looks good but sales says most leads are junk.
- A new Advantage+, Performance Max, or Demand Gen campaign needs an optimization goal.
- The platform reports more conversions than your CRM or database can find.
- You want value-based bidding but have no honest values to send.
- Retargeting or prospecting keeps reaching existing customers or staff.
Skip it when the account has fewer than about 10 conversions a week of any kind. Fix traffic or offer first.
Inputs to ask for
Ask for these before you recommend anything. If the user cannot answer one, note the gap and continue with the safest default.
- Platforms and campaign types in use, and current bid strategy.
- Funnel stages with weekly counts for the last 4 to 8 weeks (for example: visit, lead, booked call, showed up, paid).
- Where truth lives: CRM, billing system, database, spreadsheet.
- How long it takes a typical lead to reach a paid or high value outcome.
- Average deal or customer value, and how much it varies.
- Tracking already in place: pixel or tag, server-side events (Meta Conversions API), Google enhanced conversions, offline conversion imports.
- Consent setup: cookie banner, privacy policy, and whether form users agree to their data being used for ad measurement.
- Who the product is for, and whether any of it is aimed at children.
Procedure
1. Map the funnel and count it
Write each stage with its weekly count and the median days from click to that stage. Ask the user to confirm the counts from their own system, not from the ad platform.
2. Run the volume check
The platform needs roughly 50 optimization events per week per campaign (or per ad set on Meta) to leave learning and stay stable. Find the deepest stage that clears that bar, and that happens within about 7 days of the click for Meta or about 30 to 90 days for Google offline imports (check the platform's current window).
- Deepest stage clears 50 a week: optimize for it.
- It clears 20 to 50: use it if the campaign budget can be consolidated, or combine it with a scored version of a shallower stage (step 4).
- Under 20: optimize for a shallower stage, but send a value with it so the system can still tell good from weak.
3. Define a qualified event
Write one sentence that a salesperson and an engineer would both read the same way. It needs:
- a name (for example
qualified_lead) - the exact rule that fires it, based on things the person did or told you
- where it is recorded and who owns it
- how long after the click it usually fires
Use behavior and stated answers. Never use guessed traits about the person (see Guardrails).
4. Add an optional self-qualify question
If the form is the main event, add one short question whose answer splits buyers from browsers. Examples: team size, timeline, budget band, service area postcode, "Which of these describes you?" Keep it to one question. Each extra field costs volume, so recheck step 2 after adding it.
5. Build a first-party engagement score
Score only actions you observed on your own properties with consent:
| Signal | Example points |
| Answered the qualifying question with a fit answer | +30 |
| Opened or clicked the follow-up email | +10 |
| Visited pricing or booking page twice or more | +15 |
| Booked a call or appointment | +40 |
| Used a disposable email or obvious fake details | -50 |
Start with 4 to 6 signals. Set the points so a score of 50 or more means "the team would happily take this call." Tune after two weeks of data.
6. Send the qualified result back
Pick the route that fits the stack:
- Meta: send the qualified event or a value through the Conversions API from your server, deduplicated with the browser pixel by event ID. Use hashed contact fields only where the person gave consent.
- Google: upload offline conversions or use enhanced conversions for leads, keyed by click ID or hashed email. Value rules can raise or lower value by location, device, or audience list.
- Values: if you bid on value, send a number that tracks real worth. Expected value works: score band multiplied by average deal value and by that band's historical close rate.
If the owner has paused all campaigns, do not start or re-enable any without asking them first.
Then switch the campaign goal to the new event or to value once it has about a week of clean data. Change one campaign first.
7. Move the threshold each week
Once a week, compare score bands with real outcomes:
- If people scoring 50 to 60 almost never buy, raise the cutoff.
- If the qualified event drops under the volume bar, lower the cutoff or widen the date window.
- Log each change with the date so you can match it to performance shifts.
Avoid changing the cutoff more than once a week. Each change resets some learning.
8. Add first-party exclusions
Upload consented lists of existing customers, current trial users, staff, and known spam submitters as exclusions where the campaign type allows it. Base exclusions on what people did with you, never on sensitive traits.
9. Hold out and compare
- Keep a holdout: a geo, a time block, or a platform conversion lift test where the new setup is off.
- Each week, compare platform-reported conversions with the same events in your backend. Track the ratio. A ratio that keeps climbing means the platform is claiming credit you cannot find.
- Judge success on backend numbers: qualified pipeline, revenue, cost per paying customer. Platform CPA alone does not count.
Choosing the event
| Situation | Optimize for | Send back |
| 50+ paid customers a week from ads, short cycle | Purchase or paid signup | Real revenue |
| Under 50 sales a week, 50+ booked calls | Booked call or qualified lead | Expected value by score band |
| Plenty of leads, few calls, long sales cycle | Lead with fit answer (step 4) | Score band value via CAPI or offline import |
| Leads under 50 a week | Lead, all of them | Value from score so the system still learns quality |
| Free product, revenue later (content, freemium) | An early action that predicts retention | Fixed value per action, adjusted by cohort data |
| Platform count far above backend count | Do not change the goal yet | Fix tracking and deduplication first |
Worked examples
B2B SaaS demo funnel
A project tool gets 400 demo requests a month from Performance Max. Sales says a third are students or solo users. Counts per week: 95 demo requests, 30 demos held, 6 closed deals. Deals close in about 45 days.
- Closed deals (6 a week) are too few. Demos held (30) are short of 50.
- Add one form question: "How many people will use it?" Answers of 5 or more mark a fit lead. About 55 a week qualify.
- Score: fit answer +30, work email domain +15, booked a slot +40.
- Send
qualified_demothrough Google offline conversion import keyed by click ID, value = 0.2 close rate x 6,000 average deal = 1,200. - Hold out two regions for four weeks. Compare demos held and pipeline per euro spent in test versus holdout.
YouTube channel subscriber funnel
A cooking channel runs Demand Gen to grow subscribers. Subscribers are cheap but many never watch a second video.
- The platform cannot see "watched three videos" directly, so define the qualified event on the channel's own site instead: joined the email list from the video description and opened the second email.
- About 70 a week qualify. That clears the bar.
- Send the event with a fixed value through enhanced conversions, using the email the person typed in with consent.
- Weekly guardrail: compare new subscribers who watched 2+ videos in YouTube Studio for test weeks against the prior four weeks.
- Content is aimed at adults. If a channel is made for kids, do not use this method for that audience at all.
Local services lead form
A plumber runs Meta Advantage+ lead ads. Many leads live outside the service area or want a quote for a job the business does not do.
- Add one question: postcode. Add a second only if volume allows: job type.
- Qualified lead = postcode inside the service area and a job type on the list. About 40 a week, under 50.
- Optimize for all leads, but send each lead's value through the Conversions API: 0 for out of area, 50 for in area, 150 for in area with a booked visit. Use value optimization once there is a week of data.
- Exclude existing customers using the consented customer list.
- Guardrail: booked jobs in the job system per week, compared with Meta's lead count.
Guardrails: never do this
- Never buy data from data brokers or enrich leads with bought traits such as net worth, income, credit scores, marital status, or ethnicity.
- Never scrape, de-anonymize, or try to identify people who saw, liked, commented on, or clicked an ad beyond what the platform shares with you.
- Never score, target, or exclude people using sensitive or protected traits: race, ethnicity, religion, health, sexual orientation, gender identity, disability, age where restricted, political views, financial hardship, or close proxies for any of these.
- Never use personal data for ad measurement or targeting on content made for children.
- Use only first-party data that people gave you with consent, for a purpose they were told about. Honor deletion and opt-out requests in every list you upload.
- Follow each platform's terms, including special ad categories for housing, jobs, credit, and social issues, where many targeting and exclusion options are off by design.
- Never send a fake or inflated value to push the bidder. It teaches the system to find the wrong people and breaks your own reporting.
Checklist
- Funnel counts come from the backend, for the last 4 to 8 weeks
- Chosen event clears about 50 a week, or a value is attached
- Qualified event is written as one unambiguous rule
- Score uses only consented first-party behavior and answers
- Events are deduplicated between browser and server
- One campaign changed first, with a holdout in place
- Weekly log of threshold, platform count, backend count, and ratio
- Exclusion lists are consented and current
- No sensitive traits, broker data, or kids content anywhere in the setup
Template
Copy qualified-event-spec.md to record the event definition, scoring, send-back route, and weekly review in one place.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

