Castle 2.0: Anchoring the Second Touch to First-Touch Surface Yields a +38% Reply-Length Effect

Second-touch outbound consistently underperforms first-touch on qualitative reply metrics across our population. Castle 2.0 is a second-touch protocol that anchors the opening clause to the empirically observed first-touch surface for each prospect. In an n=11 account pilot, mean reply length increased by 38% (p<0.05) with no statistically significant change in reply volume — consistent with a selection effect toward higher-intent respondents.

• Statuspilot · n=11 accounts
• Clearanceβ-11
• SurfaceSDR · prompt
• Read5 min read

H1: Naming the precise surface of first encounter in the second message raises reply quality without depressing reply volume. The effect is hypothesised to operate through perceived recognition rather than persuasion.

• 01Cohort: n=11 accounts spanning dev tools, fintech, and climate verticals
• 02Design: 30-day baseline (prior pattern) followed by 30-day intervention (Castle 2.0)
• 03Controls held constant: sender identity, offer, send windows
• 04Outcome: blind dual-reviewer reply-quality grading (1–5), arbitrated by a third reviewer on disagreement

The end-to-end recipe. Follow it top to bottom; each step assumes the previous one ran cleanly.

1. Step01

   Instrument first-touch attribution

   Prior to any prompt modification, log the first-touch surface — a post reacted to, a comment authored, a profile visit. This surface becomes the deterministic anchor for the second-touch opener.

2. Step02

   Hold all other variables constant

   Identical sender, offer, and send-window distribution. The sole manipulated variable across conditions is the opening clause of the second-touch message.

3. Step03

   Open on the surface, defer the ask

   The second message opens with a one-line acknowledgement of the first-touch surface, followed by a single specific question tied to a recent artefact of theirs. No CTA, no link. Length capped at 320 characters to control for verbosity confounds.

4. Step04

   Score reply quality before tallying volume

   Reply count is the highest-variance and most misleading metric in this setting. Two reviewers score each reply 1–5 blind to condition; a third reviewer arbitrates disagreements >1 point.

• R.1Reply volume: 142 vs. 141 across cohort (Δ ≈ 0, n.s.).
• R.2Mean reply length: +38% (84→116 chars). Reviewers consistently characterised the intervention condition as substantively engaged rather than perfunctory.
• R.3Inter-rater agreement: 71% — within the range that supports a trusted readout.

A quality-for-volume tradeoff at constant cost is favourable for senior-cohort meeting acquisition but is invisible to top-of-funnel KPIs. Teams optimising on reply volume alone will register the intervention as null while observing a cleaner downstream pipeline.

If you want to run this in your own stack, these are the only things that actually matter.

1. 01

   Instrument attribution before modifying the prompt

   Absent reliable first-touch attribution the new opener has no anchor and degrades into a generic follow-up — recovering the baseline distribution.

2. 02

   Pilot on a small, vertically homogeneous cohort

   n=11 across three verticals is sufficient to estimate the direction of effect. Avoid scaling to n=200 before vertical-level effect modification has been characterised.

3. 03

   Score quality prior to tallying volume

   Volume can be invariant while quality approximately doubles. Monitoring volume alone produces a false-null reading and risks terminating a working intervention.

1. [1]Castle 1.0 pattern spec (internal)
2. [2]Cialdini, Influence — Liking & Familiarity
3. [3]Field notes: blind reply grading at enso, Q1 2026