Signal-Based Outreach Engine

Purpose. Review the *shape* of the sequence, not the words, step count, channel mix, timing, caps. (Copy lives in *Copy Review*.)

Trigger: monthly. Fed by *Weekly Performance Review*.

1. Read the sequence via Gojiberry `get_campaign` (`steps`): step types (invitation / message / email / visitProfile), order, `delayAfterLastStep`, `launchHour`, `activeDays`, send caps, `isManual`.
2. Map the funnel drop-off by step from `campaignStatus` (found → contacted → accepted → replied → interested) to see where prospects fall out.
3. Recommend structural changes: add/remove/reorder steps, adjust inter-step delays, rebalance LinkedIn vs email, change daily caps or launch window, connection-first vs email-first.
4. Human approval before changing a live campaign's structure.

Output: structure recommendations + a change log. Feeds *Campaign Setup & Launch*.

How to read it honestly.

- *Credit the first answered step.* A reply marks the whole thread as answered, so counting answered states per step over-credits later steps. Sort each replier's events by time and credit only the first step they answered.

- *Split by connection state.* Compare reply rates for accepted connections against people who never connected; if the gap is large, connection-first beats email volume.

- *Check skips.* Read the share of steps skipped (invites not sent, DMs skipped for non-connections), a long sequence often collapses into one cold email for most leads.

- *Check timing.* Share of sends landing at weekends or outside working hours.

- Recommend one structural change per cycle, with the result that would prove it right.

Purpose. Review the *words* of the outreach and improve them, openers, subject lines, message bodies, A/B variants. (Structure lives in *Campaign Structure Review*.)

Trigger: weekly (Thu). Fed by *Weekly Performance Review*.

1. Pull per-step and per-A/B-variant engagement from Gojiberry: `list_campaigns` → `get_campaign` (`steps`) → `list_contacts({ listId })`, reading `campaignStatus`. For each step/variant compute: invitation acceptance, email open rate, message/email reply rate.
2. Rank openers, subject lines and messages best→worst. Flag any variant materially below the campaign benchmark (e.g. open below 60%, reply below target).
3. Draft improved copy for the weak variants, keep the winning angle, rewrite the loser. Maintain a running copy-test log (what was tested, what won).
4. Human approval before pushing any copy change to a live campaign.

Output: revised copy for the campaign + an updated copy-test log. Feeds *Campaign Setup & Launch*.

The reply gate, score every first touch before it ships.

A touch needs at least two of these, a first touch three:

- *Bring something*, a concrete asset made for this account (a teardown, a list, a worked example), named in the ask.

- *Observe, don't pitch*, open on a cost or blind spot they may not have priced, as an observation or genuine question.

- *One fresh, specific line*, something dated within the last three months that only a person who looked could write.

- *Easy to forward*, makes sense to a colleague with no context; one small ask.

Then check: delete the opener, does it still work? (Then it was decoration.) Could it go to another company unchanged? Is there one ask? Is the proof a number or a name? Score pull 0–10 on its own; below 7 goes back for a rewrite, naming the missing move.

Size the ask to the relationship. Cold → permission to send the asset. Connected or recent history → the asset plus a soft "worth a look?". Already replied → only now ask for the meeting.

Signal-anchored openers. The first line should only make sense because the trigger happened. Use public facts only and never reveal how the lead was found.

Never: a statistic as the opener, a direct competitor of the recipient as proof, "just following up" or "circling back", or reopening a quiet thread with nothing new. Log the winning move on every positive reply.

Purpose. Maintain the Gojiberry lead-finding *Agent* that sources prospects matching your ICP and intent signals. In Gojiberry, Agents find leads; Campaigns do the outreach.

Define your ICP (set on the Agent, or via `update_agent`):

- Job titles: `{{target job titles}}`

- Industries: `{{target industries}}`

- Company size: `{{company sizes, e.g. 11–50, 51–200}}`

- Locations: `{{target locations}}`

- Exclude: agencies / service providers (`excludeServiceProviders: true`); named competitors + existing customers via `ignoredCompanies`

- Keep `minLeadScore` at a floor that protects list quality

Signals (configured in the Gojiberry UI). Keyword engagement, competitor engagement, brand awareness (profile visitors / post engagers), buying events / recent job changes. Pick the signals that map to how your buyers show intent.

Cadence, fortnightly, or whenever the weekly review flags a weak signal.

1. `get_agent`, read current ICP + config.
2. `get_agent_logs`, which signals produced leads, volumes, failed runs.
3. `get_intent_type_counts`, intent distribution across sourced contacts.
4. Refine with `update_agent` (titles / industries / sizes / locations / mandatoryKeywords / ignoredCompanies / minLeadScore). Human approval before materially broadening the ICP or disabling the service-provider / score filters.
5. Log what changed and why.

Output: a tuned Agent feeding a clean, on-ICP lead list into *Campaign Setup & Launch*.

Read each signal before it becomes a lead.

- *What changed, and why now?* One line on what the trigger means for this account, plus how strong and how certain it is.

- *Decay profile.* Act-now (short, sharp intent) or ramp (a window opening over weeks)? Stamp a reach-out-by date and a cold-after date; an expired act-now signal is just a cold lead again. Never invent urgency the trigger doesn't support.

- *Verify at the source.* Hiring, funding and leadership claims count only when confirmed at the company's own site, or at a listing matched on company name and domain. Research summaries alone are never enough. If it doesn't verify, drop it entirely; don't soften it into the copy.

- *Prune by output.* Signals that produce no on-ICP leads over a full cycle are candidates to cut. Treat a run that didn't execute differently from a run that found nothing.

Purpose. Catch every reply, alert the team, and log it to your CRM, daily. Drafts replies; never auto-sends.

CRM home: a dedicated outbound deals pipeline, kept separate from your main sales pipeline. Suggested stage mapping:

- New lead sourced → *New Lead*

- Sequence sent → *Contacted*

- Positive reply → *Positive Reply*

- Call agreed → *Meeting Booked* → *Qualified Opportunity*

- Not interested / unsubscribe → *Closed Lost*

1. Gojiberry `list_unibox_threads`, threads updated since the last run (`dateFrom`), all seats. Prioritise `interested: true` and unseen.
2. For each, `get_unibox_thread_messages` to read the latest exchange.
3. Classify: positive, neutral (later / referral), negative (not interested / unsubscribe), auto (OOO / bounce).
4. Positive → post to your outbound Slack channel (name, company, sourcing signal, reply snippet, LinkedIn link); create/update the CRM contact + a deal at *Positive Reply* (move to *Meeting Booked* when a call is agreed) + attach a note with the thread; set the Gojiberry contact `fit: "qualified"`, `state: "answered"` if a call is agreed. Draft a suggested reply in the Slack alert.
5. Negative / unsubscribe → deal to *Closed Lost*; Gojiberry contact `state: "excluded"`.
6. Neutral → CRM note; if "later", `state: "paused"`.
7. escalate anything ambiguous or needing a human reply. Never `send_unibox_linkedin_message` without sign-off, draft only.

Output: a routed, CRM-logged reply queue with Slack alerts and drafted responses.

_Runs `assist` until the Gojiberry key is set and one clean run has happened, then graduate to `auto`._

Classification rules.

- *Machine mail is its own outcome.* Out-of-office: keep the lead in the sequence, move the next touch to one working day after a stated return date (never a guessed one), and record nothing, not a reply, not a stage change. Delivery deferrals: no action. Hard bounces: suppress permanently before the next send.

- *"Not interested" is an objection, not an unsubscribe.* Only an explicit stop request stops contact.

- *A "wrong person" reply is a referral.* Thank them, ask for the right name, and route it the same day.

- *A buying question is a buy signal.* Answer in one line and ask for the meeting in the same message.

- *"Let's talk" means send the booking link plus what you'll bring.* Never "which day works?".

- Every classified reply gets a CRM note and a dated next touch with its reason, so no warm thread is orphaned.

- Auto-replies and bounces never count towards reply or positive-reply metrics.

Purpose. The weekly read on performance + signals, and the hub that dispatches improvements back into the system (closes the loop).

Trigger: weekly, Mon 09:00.

1. Funnel + performance. Pull the Gojiberry funnel for active campaigns, found → contacted → accepted → replied → interested, plus emails sent, open rate, bounce rate, unsubscribes. Record the metric values (Leads sourced, Contacted, Acceptance rate, Reply rate, Open rate, Bounce rate, Unsubscribes, Emails sent, Positive replies).

2. Signals. Per Agent: `get_agent_logs` (last 7 days) + `get_intent_type_counts`, which signals produced leads (`nbResultsLastLaunch`), volume trend, failed runs, intent mix.

3. Dispatch, close the loop. Route each finding to the process that owns the fix:

- Weak / quiet signals (low `nbResultsLastLaunch`, dead keywords or competitors) → ICP & Signal Agent to retune targeting + signals.

- Copy underperformance (low open or reply by step / A-B variant) → Copy Review.

- Funnel drop-off / structural issues (low acceptance, channel mix, timing, caps) → Campaign Structure Review.

4. Report out. Record a report on this system (funnel table, signal table, dispatched actions, week-over-week) and post a short Slack digest.

Output: weekly report + Slack digest + logged metric values, with explicit hand-offs to signal tuning, copy and structure.

Diagnose in layers, stop at the first break.

1. *Delivery*, bounces, complaints, placement. If this is broken, change nothing else this week.
2. *Engagement*, acceptance and opens (opens are diagnostic only; tracking is unreliable).
3. *Replies*, by first answered step. Healthy opens with near-zero replies point to the offer or copy; weak opens and weak replies mean fixing delivery and subject first.
4. *Reply quality*, split positive, referral, not-now and negative before judging anything. Lots of replies but mostly negative is a segment problem.

Replies but no meetings? Check reply speed and the ask before touching copy; interest fades within hours.

Rules. Quote the sample size with every rate. Leave out auto-replies. Read acceptance per sending account (one account dropping is an account problem; all dropping together is targeting or copy). Prescribe one change per week, with the result that would confirm it.