What Works

The difference between a drafting tool and a copilot is that a copilot learns. Once a week (Friday, after signal-watch's recap), this skill takes real results and converts them into learnings.md — a working-folder file that write-email, draft-sequence, and handle-reply read before drafting. Per-rep conversion data, compounding weekly.

Step 1 — Get the data

There's no direct sequencer API connection yet (Outreach or otherwise), so the data comes from the rep:

Preferred: paste the Outreach report — per sequence/template: delivered, replies, meetings booked. A CSV export or a screenshot-transcribed table both work.

Fallback (5-question interview, if no export):

1. Which sequences/templates ran this week, roughly how many prospects each?
2. Replies per template (even approximate)?
3. Meetings booked — and which touch created each one (call #, email #, LinkedIn)?
4. Which call opener/voicemail actually got conversations?
5. Objections heard more than once this week?

Take approximate numbers without complaint — a rough weekly loop beats a precise quarterly one. Mark low-confidence numbers ~.

Step 2 — Score against benchmarks

Small-number honesty: under ~50 sends per template, rates are noise — say so, compare counts not percentages, and hold the verdict until n grows. Don't kill a template on 2 sends or crown one on 3.

Deliverability tripwire (guardrail #3): reply rate near zero AND no bounces/opens signal → suspect the domain before the copy. Flag to your admin rather than iterating on words nobody is receiving.

Step 3 — Diagnose, then decide

For each template/sequence, one verdict: keep (at/above benchmark), fix (specific element — name it: trigger line, trap question, CTA, persona mismatch), or kill (below benchmark at meaningful n, or two straight "fix" weeks with no movement). Use the funnel logic: no replies = message/list; replies but no meetings = CTA/handling; connects but no conversations = opener.

Cross-reference handle-reply's flagged objections: an objection appearing 3+ times is either a list problem (wrong accounts) or a messaging gap (escalate to your marketing lead per guardrail #7).

Step 4 — Update learnings.md

Create or update learnings.md in the working folder. Structure (keep the whole file under ~60 lines — prune stale entries; a learnings file nobody can scan is a log, not a learning):

# Learnings — updated [date] · week [N] of tracking

## Proven (use by default)
- [pattern + evidence: "renewal-angle trigger lines: 3 meetings from 40 sends (Hook #5, Practitioner)"]

## Killed (do not reuse)
- [template/pattern + why + date]

## Testing (jury out)
- [pattern + current n]

## Objection frequency (for handle-reply + your marketing lead)
- [objection → count → status: approved response holding / needs messaging review]

## Open flags
- [deliverability watch, clearance items, anything for your marketing lead or admin]

Every entry carries its evidence and date. When a "Proven" entry's pattern stops working, move it down rather than deleting — knowing what USED to work matters.

Output format

# What Works · week of [date]

## Scorecard
| Sequence/template | Sent | Replies (%) | Meetings | Verdict |
|---|---|---|---|---|

## This week's read
[2-3 sentences: the single biggest signal in the data]

## Changes
- KEEP: ...
- FIX: [template] — [the specific element] → [proposed rewrite direction]
- KILL: ...

## For your marketing lead / admin
[recurring objections, deliverability flags, anything needing a decision]

learnings.md updated ✓

End with: offer to redraft the "FIX" templates now (write-email picks up the new learnings automatically), and remind the rep the loop only compounds if it runs weekly — same day, every week.