Open this page whenever the user gives feedback on a draft, wants outreach to sound better or perform better, announces a strategy or positioning shift, or asks to set up or refresh the org's voice. This is the playbook for making the skill the org's — everything about improving outreach gets recorded through here.
Recording feedback — the procedure
When the user reacts to a draft, run these four steps in order:
1. Fix the draft first. The feedback applies to the message in front of you before anything else.
2. Turn it into a principle. Rewrite the reaction as a reusable one-line do/don't, keeping any condition the user gave. "still cringe" about a bonding line → don't name the relationship — no "fellow founder", no "just two builders". "keep it under 60 words for CFO types" → to CFOs: under 60 words. The principle, never the incident — the rule must work on next month's draft to a different company.
3. Pick the home — one of exactly three places:
| Where it applies | Home | Edit to make |
|---|---|---|
| One motion only ("on cold lists…", "for webinar follow-ups…") | that motion's reference page | add the line under its ## This org's rules heading |
| Everywhere this org does outreach — tone, bans, audience-conditioned rules | the Reach out page itself | add the line under its ### This org's voice and rules heading |
| One person only ("my sign-off", "my emails specifically") | that sender's instructions | update the sender instructions |
Team-wide is the default — a user teaching Swan to write better is almost always teaching it for everyone. Personal is only for explicit "me/my" feedback. If genuinely unsure between one motion and everywhere, ask in one line: "save that for cold lists only, or for all outreach?"
4. Confirm what was written and where — one line back to the user, so wrong routing gets caught immediately.
Also treat silent edits as feedback: when the user rewrites a draft before approving instead of commenting, diff their version against yours, generalize each delta the same way, and propose saving the ones that look like standing preferences.
Notes: skill pages only get edited in live conversation — a play running in the background never writes to them. Message copy itself is never pasted into pages; approved sends become style examples automatically, and unapproved drafts never get pre-saved.
Set up or refresh the voice
Never a prerequisite for drafting — a first draft can ship provisional; this flow is how it stops being provisional. Branch on what's connected and degrade down the list:
- CRM + email connected: pull threads from the last 90 days where the contact replied or a meeting was booked. Extract subject patterns, opener style, length, CTA style, sign-off. Three winning threads sharing a pattern is a pattern; one is noise. Record loser patterns too — "tried 'quick question' 40 times, zero replies" is as valuable as a winner.
- Email only: search the sent folder for outbound threads that got replies; same extraction.
- Neither: ask the user to paste 3–5 cold emails that booked meetings, or to write one now for a real target.
- Nothing at all: derive a starting voice from the website and case studies, warmed toward 1:1 outreach — shorter sentences, less corporate, more direct. Flag it clearly as a soft starting point.
Save the results: sender instructions per sender (a founder and an SDR don't sound the same — don't homogenize); 3–5 strong examples per persona and play type as message examples; and a compact note in org memory — each sender with a one-line voice fingerprint, example counts by play type, cadence preference, open loops, a provisional flag when voice came from positioning rather than real threads, and a last-refreshed date.
Course-correct when the strategy changes
Saved examples don't expire, can't be edited, and can't be deleted — the levers are the skill and fresh examples. When the user rejects a draft's whole direction or announces a shift:
- Write the new direction into the skill first — the parent's org voice section for org-wide shifts, the motion page's org rules for play-level ones. Rules outrank examples, so this alone stops old examples from steering strategy.
- Refresh the pool: draft one or two messages in the new direction, get the user's explicit approval in the conversation, and save them as examples with reasoning that names the new strategy. Retrieval matches on reasoning, so future drafts surface the new examples ahead of the old.
- Voice-level shift → update the sender instructions too.
Old examples will still surface sometimes; that's fine — take their rhythm, never a choice the current rules forbid.
Improve from results
When the user asks why outreach isn't working or wants to raise reply rates: pull recent sequences for the play, compare the slices and angles that got replies against the ones that died, and propose one change at a time — a narrower slice, a retired angle, a different persona — written into the relevant motion page so the next run inherits it.