--- name: recruiting-funnel-metrics source: https://app.decimal.ai/s/recruiting-funnel-metrics@1/SKILL.md source_sha256: cc968039a202 --- # Compute recruiting funnel metrics from their standard definitions Handed a hiring pipeline — 400 applied, 80 screened, 30 onsite, 10 offered, 7 hired, with a mix of dates and sources — the base model produces numbers that look right and often aren't. It divides by whatever count is nearest, starts the clock on whatever date is handy, and reports two different recruiting metrics with the same formula. Each of these headline numbers has *one* standard definition with a *specific* denominator and a *specific* date anchor. Get the denominator wrong and the number is not a rougher version of the right answer — it is a different metric wearing the right name. The traps are almost always the same: picking the wrong denominator, anchoring a duration on the wrong date, and collapsing two near-twin metrics (time-to-fill vs time-to-hire, source-of-hire vs source yield, stage pass-through vs top-of-funnel conversion) into one. Compute each metric to its canonical definition, name which of a twin pair you are reporting, and never quote a duration without saying which two events it spans. ## When to activate - "Here are our pipeline stage counts — what are our pass-through / conversion rates?" - "What's our time-to-fill / time-to-hire for the roles we closed?" - "What share of our hires came from each channel?" or "which source is most productive?" - "We extended N offers and M were accepted — what's the acceptance rate?" - Any request to turn requisition dates, stage counts, source tags, or offer outcomes into a metric. Not this skill: drafting rejection or outreach emails, checking an offer against a comp band, forecasting next quarter's req volume, or a generic web/product conversion funnel with no hiring stages (that is a plain funnel-decomposition question). ## The canonical definitions Each metric below lists its **numerator ÷ denominator** and, for durations, the **two dated events** the clock spans. These are the definitions to compute to; the sections after cover the traps. ### Durations - **Time-to-fill** — days from the **requisition being opened/approved** to the **offer being accepted**. It measures how long the *role* was open, so its clock starts before any particular candidate exists. The start anchor is the req, not the hire. - **Time-to-hire** — days from the **eventual hire first entering the pipeline** (their application or the day they were sourced) to **their offer being accepted**. It measures the *hired candidate's* journey, so its clock starts when that person appeared, which is usually well after the req opened. The start anchor is the candidate, not the req. - **Time-to-start** — from offer accepted to first day worked. A separate downstream duration; do not fold the notice-period gap into time-to-fill. Report durations as a **median**, not a mean. Filled-role durations are right-skewed — one req that sat open for months drags the average far above what a typical role takes. ### Rates and shares - **Stage pass-through rate** — candidates who advanced ÷ candidates who **entered that same stage**. It is conditional and stage-to-stage. The denominator is the prior stage's count, never the top of the funnel. - **Overall (applicant-to-hire) conversion** — hires ÷ **total applicants** at the top. Cumulative, end to end. Measure it directly on one cohort; do not assume it equals the product of stage rates unless every stage was measured on that same sequential cohort. - **Offer-accept rate** (offer acceptance rate) — offers **accepted** ÷ offers **extended**. The denominator is offers put out, not candidates interviewed and not hires made. Unresolved (still pending) offers are not accepts and not declines — exclude them from the denominator until they resolve, rather than counting them as rejections. - **Source-of-hire** — hires from a source ÷ **total hires**. It is a *composition*: the shares across all sources sum to 100%. It answers "where did our hires come from?" - **Source yield** (source efficiency) — hires from a source ÷ **candidates from that source**. It is a *rate per source*; the sources' yields do not sum to anything. It answers "how productive is this source per candidate it sends?" ## The near-twin traps Three pairs share a name-fragment but not a formula. Naming which one you report is half the work. **Time-to-fill vs time-to-hire.** Same endpoint (offer accepted), different start. Fill starts at the req; hire starts when the person who was hired entered. A candidate who applies two weeks before accepting has a two-week time-to-hire even if the req had been open for two months (a two-month time-to-fill). Quoting the req-anchored number as the candidate's experience, or vice versa, is the most common error here. **Source-of-hire vs source yield.** Both are "per source," but source-of-hire divides by *total hires* (a mix that sums to 100%) and yield divides by *that source's own candidate count* (a productivity rate). A channel that sends few candidates but converts them well has a *high yield* and often a *low share of hires* at the same time — both true, neither interchangeable. Dividing a source's hires by that source's own applicants and calling the result "source-of-hire" reports yield under the wrong label. **Stage pass-through vs overall conversion.** Pass-through is one stage's exit ÷ that stage's entry; overall conversion is hires ÷ all applicants. Computing a stage's rate against the top-of-funnel applicant count instead of the prior stage produces neither metric. ## Cohort hygiene A stage rate is only meaningful once the cohort has *cleared* the stage. If half of last week's applicants are still sitting in screening, their outcome is undecided — computing screen-to-interview rate now counts them as failures they haven't yet become and understates the rate. Measure pass- through on a cohort old enough that the stage has resolved for (nearly) all of them, or state that in-flight candidates are censoring the number. ## The output For a funnel question, return each metric to its standard definition: the exact numerator and denominator you used (and, for a duration, the two dated events and that it is a median), which twin of an ambiguous pair you are reporting, and a flag on any number computed from an unresolved or mixed-cohort population. When the same word could mean two metrics, give both or ask which — do not silently pick one.