diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json index 4a8d411..06ed02d 100644 --- a/.claude-plugin/plugin.json +++ b/.claude-plugin/plugin.json @@ -3,7 +3,7 @@ "name": "synter", "displayName": "Synter", "description": "Free read-only Synter advertising skills and MCP connector. Review performance, audiences, and plans across 43 ad platforms over OAuth.", - "version": "1.3.1", + "version": "1.3.2", "license": "MIT", "author": { "name": "Synter", diff --git a/.cursor-plugin/marketplace.json b/.cursor-plugin/marketplace.json index 2dbbbbd..0b020a8 100644 --- a/.cursor-plugin/marketplace.json +++ b/.cursor-plugin/marketplace.json @@ -6,7 +6,7 @@ }, "metadata": { "description": "Synter's official plugin marketplace. The AI Agent Operator for Ads.", - "version": "1.3.1" + "version": "1.3.2" }, "plugins": [ { diff --git a/.cursor-plugin/plugin.json b/.cursor-plugin/plugin.json index 2371264..0c82c10 100644 --- a/.cursor-plugin/plugin.json +++ b/.cursor-plugin/plugin.json @@ -2,7 +2,7 @@ "$schema": "https://cursor.com/schemas/cursor-plugin/plugin.json", "name": "synter", "displayName": "Synter", - "version": "1.3.1", + "version": "1.3.2", "description": "Free, read-only Synter MCP plugin for Cursor. Review ad performance and account health across 43 ad platforms over OAuth. No create, update, launch, or spend tools in this listing.", "author": { "name": "Synter", diff --git a/.gitignore b/.gitignore index 39a366e..f108e23 100644 --- a/.gitignore +++ b/.gitignore @@ -13,3 +13,4 @@ node_modules/ # Build artifact — regenerated by scripts/build-zip.sh dist/ +__pycache__/ diff --git a/CHANGELOG.md b/CHANGELOG.md index 23533cd..5649a6b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,11 @@ # Changelog +## [1.3.2] - 2026-11-13 + +### Added +- Added consolidated `outbound`, `outbound-research`, and `outbound-replies` skills under Synter. +- Added prospect discovery, contact enrichment, and CRM directory tools. + ## [1.3.1] — 2026-11-09 ### Changed diff --git a/config/claude-directory/skills-allowlist.json b/config/claude-directory/skills-allowlist.json index b1d76de..20aed82 100644 --- a/config/claude-directory/skills-allowlist.json +++ b/config/claude-directory/skills-allowlist.json @@ -18,6 +18,9 @@ "llm-visibility-audit", "media-plan", "negative-keywords", + "outbound", + "outbound-replies", + "outbound-research", "pixel-capi-auditor", "platform-benchmarks", "quickstart", diff --git a/packages/claude-directory/.claude-plugin/plugin.json b/packages/claude-directory/.claude-plugin/plugin.json index d10229e..e4f8784 100644 --- a/packages/claude-directory/.claude-plugin/plugin.json +++ b/packages/claude-directory/.claude-plugin/plugin.json @@ -3,7 +3,7 @@ "name": "synter", "displayName": "Synter", "description": "Free read-only Synter advertising skills and MCP connector. Review performance, audiences, and plans across 43 ad platforms over OAuth.", - "version": "1.3.1", + "version": "1.3.2", "license": "MIT", "author": { "name": "Synter", diff --git a/packages/claude-directory/skills/outbound-replies/SKILL.md b/packages/claude-directory/skills/outbound-replies/SKILL.md new file mode 100644 index 0000000..e52da1d --- /dev/null +++ b/packages/claude-directory/skills/outbound-replies/SKILL.md @@ -0,0 +1,68 @@ +--- +name: outbound-replies +description: Triage and answer replies to cold email campaigns in Synter Outbound — separate real humans from auto-responders, classify intent, draft a reply in the sender's voice, and route opt-outs to suppression. Use when a user asks about outbound replies, their unibox, who responded to a campaign, or wants to answer a prospect. +--- + +# Synter Outbound Replies + +Most of what looks like a reply is not one. Sort that out before you tell anyone how a campaign is doing. + +## 1. Auto-responders are not replies + +**Roughly 90% of raw replies are out-of-office.** On live campaigns, campaign 216 showed 87 inbox replies: 80 automated, 7 human, 2 actually interested. + +Three things must come out of the count before you report it: + +- **Auto-responders** — out-of-office, vacation, "no longer with the company" +- **The Sent folder** — our own outbound comes back through the same feed; counting it means counting our own words as prospect interest +- **Bounces** — a delivery failure is not a reply + +Only what remains is a reply. Never report the raw number, and never compute a rate from it. + +## 2. Classify what is left + +| Kind | Looks like | Do | +|---|---|---| +| Interested | "happy to discuss", asks about price or timing | draft a reply, propose a time | +| Question | wants detail before committing | answer the question, nothing else | +| Wrong person | "that's not me", "talk to X" | thank, ask for the handoff, never re-pitch | +| Not now | "we're in a good place", "revisit in Q1" | acknowledge, log the date, stop | +| **Opt-out** | "remove me", "unsubscribe", "where did you get my information" | **suppress immediately**, then a one-line apology if any reply at all | + +An angry "where did you get my information?" is an opt-out and a targeting failure. Suppress immediately, and treat it as evidence the list was not qualified — see **outbound-research**. + +## 3. Inspect queue, draft, then wait for approval + +Use `render_card_preview` to inspect pending replies and drafts, and `action_card` only after explicit user approval. A draft is not a send. + +Every reply to a real person needs explicit approval before it goes. Approval for one reply does not carry to the next message in the same thread. + +**Error recovery & feedback loops:** +- **Rejected or edited draft:** If the user rejects a draft or asks for changes, revise the draft text (`action_card` with `domain: "outbound"`, `action: "edit"`), show the revised text, and wait for explicit approval before calling `action_card`. If the user decides not to reply at all, decline/skip the draft (`action_card` with `domain: "outbound"`, `action: "reject"`). +- **Ambiguous classification:** If a reply is ambiguous (e.g., unclear whether it is a soft "not now", a referral, or an opt-out), do not guess or auto-draft a pitch. Show the verbatim message to the user and ask whether to answer, defer, or suppress. + +Voice rules, enforced by the copy gate: + +- **No em-dashes.** Short sentences. Specific asks. +- No "I hope this finds you well", no "circling back", no "per my last email". +- Never invent a number, a client name, a case study, or a stat. If you do not have the figure, do not use one. +- Say "one prompt", never "click" — Synter is prompt-driven. +- Answer what they asked. A question about pricing gets a price, not a pitch. + +## 4. Booking + +When someone wants time, send the real scheduling link rather than proposing slots blind. Confirm the meeting actually lands in the calendar before telling the user it is booked — "created" is not "booked". + +## 5. What to report back + +When summarizing a campaign's replies, give the user: + +``` +N human replies (M auto-responders excluded) + → X interested + → Y questions + → Z not now + → W opt-outs ← flag these, they are the targeting signal +``` + +Two interested replies from 750 leads is not a reply rate to celebrate, and it is not a volume problem to solve by adding leads. Say what it is: the offer or the list is wrong. diff --git a/packages/claude-directory/skills/outbound-research/SKILL.md b/packages/claude-directory/skills/outbound-research/SKILL.md new file mode 100644 index 0000000..aaed21d --- /dev/null +++ b/packages/claude-directory/skills/outbound-research/SKILL.md @@ -0,0 +1,130 @@ +--- +name: outbound-research +description: Discover and qualify prospects in Synter before writing cold outreach — find people and verified contact info (emails, phone numbers, LinkedIn URLs), verify a checkable fact from first-party signals, rank competing signals, and land it in the email without sounding AI-generated. Use when finding prospects or leads, building or topping up an outbound list or lead list, writing a first-touch cold email to a researched prospect, or deciding whether someone should be contacted at all. +--- + +# Synter Outbound Research & Prospect Qualification + +Find a checkable fact about the person, or do not mail them. + +## 0. Finding & Enriching Prospects in Synter + +Use Synter's native discovery and signal tools to find the right person and verified contact channels: + +1. **Find Decision-Makers (`company_intel_search_people`)**: + - Search by company domain, job function, seniority level (`c_suite`, `vp`, `director`), technology stack, or keyword query. + - Returns matching people with their LinkedIn URLs, titles, companies, and enriched contact data. +2. **Enrich Emails, Phones & LinkedIn (`enrich_contacts`)**: + - Pass any combination of LinkedIn URLs, work emails, or phone numbers to `enrich_contacts` to resolve verified `work_email`, personal `emails`, direct `phone` / `phones`, `linkedin_url`, `job_title`, and `company`. +3. **Check Company Signals (`company_intel_get_signals` / `company_intel_get_profile`)**: + - Pull hiring velocity, open roles, tech stack changes, and firmographics to ground outreach in a verifiable fact. + +## The test + +For one real lead on the list, ask: **if they replied "where did you get my information?", what would I say?** + +If the honest answer is "you match an industry and a job title," the research is not done. Cut the cohort rather than send it. That question is not hypothetical — it is a verbatim reply from a real campaign that mailed 750 people a character-identical email. + +## 1. Disqualifiers first — they veto, they do not score + +Check these before any positive research (`get_crm_contacts`). One hit stops the send, no matter how strong the positives. + +| Check | Blocks because | +|---|---| +| In the CRM already | known contact or open deal | +| We have already met | cold copy insults someone whose call you attended | +| Current paying customer | they already buy from us | +| Contacted before, any channel | double-touch reads as spam | +| Suppressed, bounced, unsubscribed | legal and deliverability | + +A **lapsed** customer is the deliberate exception: a win-back target, and a strong one. Say so plainly rather than pretending it is a cold approach. + +## 2. Positive signals — use the strongest you can actually fetch + +One fact per email. Never stack them. + +``` +they clicked our ad ............... strongest 30 day window +they used the product ............ very strong 45 day window +they engaged our social post ..... strong 60 day window (comment > reaction) +someone at their company visited . moderate 30 day window (company-level only) +new role, promotion, funding ..... strong 90 day window +role + industry only ............. NOT A SIGNAL — do not send +``` + +Ranking rules: + +1. **Recency beats strength within a tier.** A reaction from yesterday beats an ad click from four weeks ago. +2. **Honour the decay windows.** Opening with "I saw you engaged with my post" about something from March is worse than no personalization. +3. **Person-level beats company-level.** "You clicked" is a fact. "Someone at your company visited" is an inference and must be written as one. + +## 3. Land it without the AI tell + +**Never name the surveillance.** The fact goes in, the mechanism stays out. + +``` +BAD "I noticed you visited our pricing page on Tuesday." +BAD "Our system flagged that you clicked one of our ads." +GOOD "Your take on the PMax thread stuck with me." +GOOD "You were looking at programmatic pricing recently, so this may be timely." +``` + +The test: would the sentence be creepy said aloud at a conference? If yes, reference the topic, not the event. + +**Banned openers** — the tells that read as generated: + +- "I hope this email finds you well" +- "I noticed that..." / "I saw that you..." / "I came across your profile" +- "As a {TITLE} at {COMPANY}, you probably..." +- Em-dashes anywhere +- The word "click" — Synter is prompt-driven; say "one prompt" + +**Shape of the opening:** + +``` +line 1 timing or season anchor, short and human +line 2 the fact, stated as theirs not ours +line 3 why it matters to them specifically +``` + +Keep the fact out of line 1. Leading with it is what makes it feel like a merge field. A human beat first, then specificity, is what makes it feel written. Body under 90 words. + +## 4. Timing anchors + +| Send day | Use | Avoid | +|---|---|---| +| Mon | "kicking off the week" | **"Happy Monday!"** — the most-used cold-email opener there is; it signals bulk mail instantly | +| Tue-Wed | nothing, these are the best days | forcing an anchor | +| Thu | "before this week gets away" | | +| Fri | "Before the weekend," | asking for a call today | +| Sat-Sun | do not send cold | | + +Rotate anchors across a cohort — the same one on every email is its own tell. And only use a day anchor if the schedule actually matches it; "before the weekend" landing on Tuesday is worse than nothing. + +## 5. Seasonal and fiscal + +Read from **their** calendar, not ours. For a DTC brand in October the anchor is peak retail season, not our quarter. + +- **Jan** new budget, planning reopens · **Feb-Mar** Q1 under pressure +- **Apr-Jun** mid-year reforecast · **Jul-Aug** thin teams, keep asks small +- **Sep** everyone back, strongest window of the year +- **Oct-Nov** Q4 planning, next-year budget locking, retail peak +- **Dec** do not start sequences after the second week + +## 6. Life events — the strongest human hook + +| Event | Window | Angle | +|---|---|---| +| New role | **first 90 days**, the best window there is | choosing a stack, wants early wins | +| Promotion | 60 days | scope grew, budget grew | +| Funding | 90 days | pressure to deploy into growth | +| Team growth | 60 days | hiring paid roles means scaling spend | + +Congratulate in **one clause**, then get to the substance. A paragraph about their promotion is transparently a device. + +``` +GOOD "Three weeks into the growth seat - you're probably still + picking what the paid stack looks like." +BAD "Congratulations on your well-deserved promotion! It's always + exciting to see talented leaders like yourself step into..." +``` diff --git a/packages/claude-directory/skills/outbound/SKILL.md b/packages/claude-directory/skills/outbound/SKILL.md new file mode 100644 index 0000000..7e86936 --- /dev/null +++ b/packages/claude-directory/skills/outbound/SKILL.md @@ -0,0 +1,68 @@ +--- +name: outbound +description: Find prospects (verified emails, phone numbers, LinkedIn URLs) and run cold email outbound through Synter Outbound — search contacts, enrich identifiers, review the pipeline, inspect campaigns, check sending health, and pause or resume. Use when a user asks to find prospects, emails, phone numbers, LinkedIn profiles, or asks about outbound, cold email, sequences, their lead pipeline, reply rates, or wants to launch, pause, or top up an email campaign. +--- + +# Synter Outbound & Prospect Discovery + +Synter Outbound is Synter's native prospect discovery, contact enrichment, and cold-email engine. You read and discover freely; you never send email without an explicit yes. + +## 1. Finding Prospects (Emails, Phone Numbers, LinkedIn URLs) + +When a user asks to find prospects, decision-makers, emails, phone numbers, or LinkedIn profiles: + +- `company_intel_search_people` — search people at target companies by `domains`, `job_functions`, `job_levels`, `countries`, `technologies`, `query`, `search`, or `linkedin_urls`. Automatically enriches returned profiles via Synter Signals with verified work emails, personal emails, phone numbers, and LinkedIn URLs. +- `enrich_contacts` — pass a list of emails, phone numbers, or LinkedIn profile URLs (`identifiers`) to resolve and enrich verified work email, personal emails, direct phone numbers, LinkedIn URL, job title, and company firmographics. +- `company_intel_search_people` — natural-language prospect search across companies and contacts with automatic email/phone/LinkedIn enrichment. +- `synter_audience` — build and export full B2B contact or US adult signal audiences at scale. +- `company_intel_get_profile` / `company_intel_get_signals` — inspect company firmographics, hiring velocity, and tech stack buying signals before outreach. + +## 2. Where outbound stands + +Start every outbound campaign conversation with the real state: + +- `get_crm_contacts` — leads by stage, active campaigns, replies awaiting review, and pipeline metrics. +- `render_card_preview` — recent human replies classified by intent and draft response status. + +Sending capacity is the fleet's, not one campaign's. Every campaign draws on shared sending domains and mailboxes, so one campaign's bounces damage all of them. + +## 3. Read reply numbers honestly + +**Roughly 90% of raw "replies" are out-of-office auto-responders.** Measured across live outbound campaigns: + +| campaign | inbox replies | automated | human | interested | +|---|---:|---:|---:|---:| +| 216 | 87 | 80 | 7 | 2 | +| 235 | 23 | 20 | 3 | 0 | +| 195 | 28 | 25 | 3 | 2 | + +Campaign 216 read as "85 replies, 3.0%" while producing **two** interested humans from 750 leads. + +Always quote human reply rate (`reply_rate_pct`: human replies ÷ sends) and `total_interested`. **Never quote `total_replied_including_automated`, and never compute a rate from it.** If the human figure is null, the filtered stats have not synced yet — say so rather than substituting the inflated one. + +Bounce rate above ~2% is a fleet emergency, not a campaign problem. Say so plainly. + +## 4. Before anything sends + +Cold email is outward-facing and irreversible. Confirm before every send, resume, or lead push — approval for one campaign never carries to the next. + +Two gates, both mandatory: + +1. **Qualification** — run the **outbound-research** skill. If you cannot name a checkable fact about a recipient, do not mail them. Matching an industry and a job title is not a fact. +2. **Deliverability** — every address verified via `enrich_contacts`, duplicates removed, nobody on the suppression list (`get_crm_contacts` / `get_crm_contacts`), and merge tokens rendering cleanly. + +## 5. Launch, pause, and reply management + +- `action_card` — creates a drafted campaign sequence. Creating is not sending without explicit launch confirmation. +- `action_card` — execute `approve`, `pause`, or `resume` on an outbound campaign card (`domain: "outbound"`, `action: "approve" | "pause" | "resume"`). Requires explicit user confirmation for launch or resume. +- `render_card_preview` — inspect pending prospect replies and draft responses. +- `action_card` — approve and send a drafted reply to a prospect. Ask first, every time. + +Pause immediately via `action_card` (`domain: "outbound"`, `action: "pause"`), without asking, if you see: bounce rate over 5%, a merge token printing literally in a sent email, or a lead who asked to be removed still receiving mail. Report it after. + +## 6. What not to do + +- Do not top up a campaign to fix a low reply rate. A campaign at 0.3% has an offer problem; adding leads spends mailbox reputation to buy nothing. +- Do not quote open rates. Open tracking is off across the platform. +- Do not send to someone already in the CRM, already met, or already a customer — see **outbound-research**. +- Do not write email bodies from scratch with generic AI prose. Templates are human-written; the model fills variable slots from real facts only, and never invents a number, a client name, or a stat. diff --git a/packages/claude-directory/tools/directory-tools.json b/packages/claude-directory/tools/directory-tools.json index d84d815..e3801c5 100644 --- a/packages/claude-directory/tools/directory-tools.json +++ b/packages/claude-directory/tools/directory-tools.json @@ -32,5 +32,13 @@ "enable_campaign", "pause_campaign", "update_campaign_budget", - "get_job_status" + "get_job_status", + "company_intel_search_people", + "company_intel_get_profile", + "company_intel_get_signals", + "enrich_contacts", + "synter_audience", + "get_crm_contacts", + "render_card_preview", + "action_card" ] diff --git a/plugin.json b/plugin.json index 9c5ce5f..169b07e 100644 --- a/plugin.json +++ b/plugin.json @@ -1,7 +1,7 @@ { "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json", "name": "synter", - "version": "1.3.1", + "version": "1.3.2", "description": "Plan and operate advertising campaigns through Synter\u2019s hosted MCP server over OAuth. Review performance, write ad copy, and approve campaign changes before execution.", "author": { "name": "Synter", diff --git a/scripts/check-directory-tools.js b/scripts/check-directory-tools.js index fcea814..7b2ec3c 100644 --- a/scripts/check-directory-tools.js +++ b/scripts/check-directory-tools.js @@ -11,7 +11,7 @@ const DATA_IDENTIFIERS = new Set([ 'conversion_id', 'custom_label', 'event_id', 'google_product_category', 'identifier_exists', 'li_fat_id', 'openai_ads', 'order_id', 'paid_social', 'product_type', 'rdt_cid', 'sale_price', 'transaction_id', 'utm_campaign', - 'utm_content', 'utm_medium', 'utm_source', 'utm_term', + 'utm_content', 'utm_medium', 'utm_source', 'utm_term', 'c_suite', 'job_functions', 'job_levels', 'job_title', 'linkedin_url', 'linkedin_urls', 'reply_rate_pct', 'total_interested', 'total_replied_including_automated', 'work_email', ]); const ROOT = path.resolve(__dirname, '..'); const SURFACES = ['skills', 'commands', 'rules', 'agents']; @@ -51,9 +51,9 @@ function listFiles(dir) { function readAllowlist(root) { const names = JSON.parse(fs.readFileSync(path.join(root, 'tools/directory-tools.json'), 'utf8')); - if (!Array.isArray(names) || names.length !== 34 || new Set(names).size !== 34 || + if (!Array.isArray(names) || names.length !== 42 || new Set(names).size !== 42 || names.some(name => typeof name !== 'string' || !/^[a-z][a-z0-9_]+$/.test(name))) { - throw new Error('tools/directory-tools.json must contain 34 unique tool names.'); + throw new Error('tools/directory-tools.json must contain 42 unique tool names.'); } return new Set(names); } diff --git a/scripts/check-directory-tools.test.js b/scripts/check-directory-tools.test.js index 7ea0295..c77958a 100644 --- a/scripts/check-directory-tools.test.js +++ b/scripts/check-directory-tools.test.js @@ -56,7 +56,7 @@ test('checks nested files in every instruction surface, including optional comma assert.equal(cli.status, 1); assert.match(cli.stderr, /unlisted tool unknown_commands/); fs.writeFileSync(path.join(root, 'tools/directory-tools.json'), '["list_campaigns", "list_campaigns"]'); - assert.throws(() => scan(root), /34 unique tool names/); + assert.throws(() => scan(root), /42 unique tool names/); }); function response(id, result, { sse = false, session = false } = {}) { diff --git a/scripts/test_build_packages.py b/scripts/test_build_packages.py index a2ff75f..e4594c5 100644 --- a/scripts/test_build_packages.py +++ b/scripts/test_build_packages.py @@ -34,7 +34,7 @@ def test_all_packages_follow_rsa_and_bidding_guidance(self): def test_directory_source_is_reproducible_and_has_only_static_runtime(self): files = mod.sync_claude_directory() self.assertEqual(files, mod.collect(mod.ROOT, 'claude-directory')[1]) - self.assertEqual(sum(name.endswith('/SKILL.md') for name in files), 27) + self.assertEqual(sum(name.endswith('/SKILL.md') for name in files), 30) self.assertEqual(sum(name.startswith('agents/') for name in files), 3) self.assertNotIn('hooks', json.loads(files['.claude-plugin/plugin.json'])) self.assertEqual(json.loads(files['.claude-plugin/plugin.json'])['privacyPolicyUrl'], 'https://synterai.com/privacy') @@ -102,10 +102,10 @@ def test_directory_generated_source_detects_drift_and_symlinks(self): def test_archives_match_reviewed_rsa_correction_bytes(self): expected = { - 'claude': '0cb7a407899bd9593f9c14fbda786656fc85e786594afe36d5465e50459ec3ff', - 'claude-directory': '6287957db5cae87857b2f01177f25fd47990fdf41883fddf0fb7d3f98d47de2d', - 'openai': '3b5313edba57d9e24422d9f39d264b9ebfed15b63ba8d0dad37864e17712d04b', - 'openai-new-draft': '3998eddad1a726fabf916dcc0ab5e531ecaaa7d8d4d035b2dbc9fdfa6eb20435', + 'claude': 'e88988640d232b94158fffbfa4fd58b28570deffa5f7e155621db028e088249d', + 'claude-directory': '57a9f244fd3e74f2afbe06f48abaf5cd7c1d5a7fb09cd31da85d047853a7f9a3', + 'openai': 'de72c10a09b57f5128fade50979672dca03f1e87e37862df06c8b2edac92d5ec', + 'openai-new-draft': '4e61d4fc480e234fd5d23a16098b8303ad7ab07871187fdba749940061baff05', } # Snapshot after the approved RSA/bidding correction across all hosts. # Update deliberately when reviewed package contents change. @@ -130,7 +130,7 @@ def test_openai_existing_identity_and_no_hooks(self): self.assertFalse(any(name.startswith(('scripts/', '.claude-plugin/', '.git/', 'hooks/')) for name in files)) self.assertNotIn('.env', files) self.assertEqual(files['assets/logo.png'], (mod.ROOT / 'assets/logo.png').read_bytes()) - self.assertEqual(len([n for n in files if n.endswith('/SKILL.md')]), 27) + self.assertEqual(len([n for n in files if n.endswith('/SKILL.md')]), 30) def test_claude_identity_and_executable_present(self): _, files = mod.collect(mod.ROOT, 'claude') diff --git a/skills/outbound-replies/SKILL.md b/skills/outbound-replies/SKILL.md new file mode 100644 index 0000000..e52da1d --- /dev/null +++ b/skills/outbound-replies/SKILL.md @@ -0,0 +1,68 @@ +--- +name: outbound-replies +description: Triage and answer replies to cold email campaigns in Synter Outbound — separate real humans from auto-responders, classify intent, draft a reply in the sender's voice, and route opt-outs to suppression. Use when a user asks about outbound replies, their unibox, who responded to a campaign, or wants to answer a prospect. +--- + +# Synter Outbound Replies + +Most of what looks like a reply is not one. Sort that out before you tell anyone how a campaign is doing. + +## 1. Auto-responders are not replies + +**Roughly 90% of raw replies are out-of-office.** On live campaigns, campaign 216 showed 87 inbox replies: 80 automated, 7 human, 2 actually interested. + +Three things must come out of the count before you report it: + +- **Auto-responders** — out-of-office, vacation, "no longer with the company" +- **The Sent folder** — our own outbound comes back through the same feed; counting it means counting our own words as prospect interest +- **Bounces** — a delivery failure is not a reply + +Only what remains is a reply. Never report the raw number, and never compute a rate from it. + +## 2. Classify what is left + +| Kind | Looks like | Do | +|---|---|---| +| Interested | "happy to discuss", asks about price or timing | draft a reply, propose a time | +| Question | wants detail before committing | answer the question, nothing else | +| Wrong person | "that's not me", "talk to X" | thank, ask for the handoff, never re-pitch | +| Not now | "we're in a good place", "revisit in Q1" | acknowledge, log the date, stop | +| **Opt-out** | "remove me", "unsubscribe", "where did you get my information" | **suppress immediately**, then a one-line apology if any reply at all | + +An angry "where did you get my information?" is an opt-out and a targeting failure. Suppress immediately, and treat it as evidence the list was not qualified — see **outbound-research**. + +## 3. Inspect queue, draft, then wait for approval + +Use `render_card_preview` to inspect pending replies and drafts, and `action_card` only after explicit user approval. A draft is not a send. + +Every reply to a real person needs explicit approval before it goes. Approval for one reply does not carry to the next message in the same thread. + +**Error recovery & feedback loops:** +- **Rejected or edited draft:** If the user rejects a draft or asks for changes, revise the draft text (`action_card` with `domain: "outbound"`, `action: "edit"`), show the revised text, and wait for explicit approval before calling `action_card`. If the user decides not to reply at all, decline/skip the draft (`action_card` with `domain: "outbound"`, `action: "reject"`). +- **Ambiguous classification:** If a reply is ambiguous (e.g., unclear whether it is a soft "not now", a referral, or an opt-out), do not guess or auto-draft a pitch. Show the verbatim message to the user and ask whether to answer, defer, or suppress. + +Voice rules, enforced by the copy gate: + +- **No em-dashes.** Short sentences. Specific asks. +- No "I hope this finds you well", no "circling back", no "per my last email". +- Never invent a number, a client name, a case study, or a stat. If you do not have the figure, do not use one. +- Say "one prompt", never "click" — Synter is prompt-driven. +- Answer what they asked. A question about pricing gets a price, not a pitch. + +## 4. Booking + +When someone wants time, send the real scheduling link rather than proposing slots blind. Confirm the meeting actually lands in the calendar before telling the user it is booked — "created" is not "booked". + +## 5. What to report back + +When summarizing a campaign's replies, give the user: + +``` +N human replies (M auto-responders excluded) + → X interested + → Y questions + → Z not now + → W opt-outs ← flag these, they are the targeting signal +``` + +Two interested replies from 750 leads is not a reply rate to celebrate, and it is not a volume problem to solve by adding leads. Say what it is: the offer or the list is wrong. diff --git a/skills/outbound-research/SKILL.md b/skills/outbound-research/SKILL.md new file mode 100644 index 0000000..aaed21d --- /dev/null +++ b/skills/outbound-research/SKILL.md @@ -0,0 +1,130 @@ +--- +name: outbound-research +description: Discover and qualify prospects in Synter before writing cold outreach — find people and verified contact info (emails, phone numbers, LinkedIn URLs), verify a checkable fact from first-party signals, rank competing signals, and land it in the email without sounding AI-generated. Use when finding prospects or leads, building or topping up an outbound list or lead list, writing a first-touch cold email to a researched prospect, or deciding whether someone should be contacted at all. +--- + +# Synter Outbound Research & Prospect Qualification + +Find a checkable fact about the person, or do not mail them. + +## 0. Finding & Enriching Prospects in Synter + +Use Synter's native discovery and signal tools to find the right person and verified contact channels: + +1. **Find Decision-Makers (`company_intel_search_people`)**: + - Search by company domain, job function, seniority level (`c_suite`, `vp`, `director`), technology stack, or keyword query. + - Returns matching people with their LinkedIn URLs, titles, companies, and enriched contact data. +2. **Enrich Emails, Phones & LinkedIn (`enrich_contacts`)**: + - Pass any combination of LinkedIn URLs, work emails, or phone numbers to `enrich_contacts` to resolve verified `work_email`, personal `emails`, direct `phone` / `phones`, `linkedin_url`, `job_title`, and `company`. +3. **Check Company Signals (`company_intel_get_signals` / `company_intel_get_profile`)**: + - Pull hiring velocity, open roles, tech stack changes, and firmographics to ground outreach in a verifiable fact. + +## The test + +For one real lead on the list, ask: **if they replied "where did you get my information?", what would I say?** + +If the honest answer is "you match an industry and a job title," the research is not done. Cut the cohort rather than send it. That question is not hypothetical — it is a verbatim reply from a real campaign that mailed 750 people a character-identical email. + +## 1. Disqualifiers first — they veto, they do not score + +Check these before any positive research (`get_crm_contacts`). One hit stops the send, no matter how strong the positives. + +| Check | Blocks because | +|---|---| +| In the CRM already | known contact or open deal | +| We have already met | cold copy insults someone whose call you attended | +| Current paying customer | they already buy from us | +| Contacted before, any channel | double-touch reads as spam | +| Suppressed, bounced, unsubscribed | legal and deliverability | + +A **lapsed** customer is the deliberate exception: a win-back target, and a strong one. Say so plainly rather than pretending it is a cold approach. + +## 2. Positive signals — use the strongest you can actually fetch + +One fact per email. Never stack them. + +``` +they clicked our ad ............... strongest 30 day window +they used the product ............ very strong 45 day window +they engaged our social post ..... strong 60 day window (comment > reaction) +someone at their company visited . moderate 30 day window (company-level only) +new role, promotion, funding ..... strong 90 day window +role + industry only ............. NOT A SIGNAL — do not send +``` + +Ranking rules: + +1. **Recency beats strength within a tier.** A reaction from yesterday beats an ad click from four weeks ago. +2. **Honour the decay windows.** Opening with "I saw you engaged with my post" about something from March is worse than no personalization. +3. **Person-level beats company-level.** "You clicked" is a fact. "Someone at your company visited" is an inference and must be written as one. + +## 3. Land it without the AI tell + +**Never name the surveillance.** The fact goes in, the mechanism stays out. + +``` +BAD "I noticed you visited our pricing page on Tuesday." +BAD "Our system flagged that you clicked one of our ads." +GOOD "Your take on the PMax thread stuck with me." +GOOD "You were looking at programmatic pricing recently, so this may be timely." +``` + +The test: would the sentence be creepy said aloud at a conference? If yes, reference the topic, not the event. + +**Banned openers** — the tells that read as generated: + +- "I hope this email finds you well" +- "I noticed that..." / "I saw that you..." / "I came across your profile" +- "As a {TITLE} at {COMPANY}, you probably..." +- Em-dashes anywhere +- The word "click" — Synter is prompt-driven; say "one prompt" + +**Shape of the opening:** + +``` +line 1 timing or season anchor, short and human +line 2 the fact, stated as theirs not ours +line 3 why it matters to them specifically +``` + +Keep the fact out of line 1. Leading with it is what makes it feel like a merge field. A human beat first, then specificity, is what makes it feel written. Body under 90 words. + +## 4. Timing anchors + +| Send day | Use | Avoid | +|---|---|---| +| Mon | "kicking off the week" | **"Happy Monday!"** — the most-used cold-email opener there is; it signals bulk mail instantly | +| Tue-Wed | nothing, these are the best days | forcing an anchor | +| Thu | "before this week gets away" | | +| Fri | "Before the weekend," | asking for a call today | +| Sat-Sun | do not send cold | | + +Rotate anchors across a cohort — the same one on every email is its own tell. And only use a day anchor if the schedule actually matches it; "before the weekend" landing on Tuesday is worse than nothing. + +## 5. Seasonal and fiscal + +Read from **their** calendar, not ours. For a DTC brand in October the anchor is peak retail season, not our quarter. + +- **Jan** new budget, planning reopens · **Feb-Mar** Q1 under pressure +- **Apr-Jun** mid-year reforecast · **Jul-Aug** thin teams, keep asks small +- **Sep** everyone back, strongest window of the year +- **Oct-Nov** Q4 planning, next-year budget locking, retail peak +- **Dec** do not start sequences after the second week + +## 6. Life events — the strongest human hook + +| Event | Window | Angle | +|---|---|---| +| New role | **first 90 days**, the best window there is | choosing a stack, wants early wins | +| Promotion | 60 days | scope grew, budget grew | +| Funding | 90 days | pressure to deploy into growth | +| Team growth | 60 days | hiring paid roles means scaling spend | + +Congratulate in **one clause**, then get to the substance. A paragraph about their promotion is transparently a device. + +``` +GOOD "Three weeks into the growth seat - you're probably still + picking what the paid stack looks like." +BAD "Congratulations on your well-deserved promotion! It's always + exciting to see talented leaders like yourself step into..." +``` diff --git a/skills/outbound/SKILL.md b/skills/outbound/SKILL.md new file mode 100644 index 0000000..7e86936 --- /dev/null +++ b/skills/outbound/SKILL.md @@ -0,0 +1,68 @@ +--- +name: outbound +description: Find prospects (verified emails, phone numbers, LinkedIn URLs) and run cold email outbound through Synter Outbound — search contacts, enrich identifiers, review the pipeline, inspect campaigns, check sending health, and pause or resume. Use when a user asks to find prospects, emails, phone numbers, LinkedIn profiles, or asks about outbound, cold email, sequences, their lead pipeline, reply rates, or wants to launch, pause, or top up an email campaign. +--- + +# Synter Outbound & Prospect Discovery + +Synter Outbound is Synter's native prospect discovery, contact enrichment, and cold-email engine. You read and discover freely; you never send email without an explicit yes. + +## 1. Finding Prospects (Emails, Phone Numbers, LinkedIn URLs) + +When a user asks to find prospects, decision-makers, emails, phone numbers, or LinkedIn profiles: + +- `company_intel_search_people` — search people at target companies by `domains`, `job_functions`, `job_levels`, `countries`, `technologies`, `query`, `search`, or `linkedin_urls`. Automatically enriches returned profiles via Synter Signals with verified work emails, personal emails, phone numbers, and LinkedIn URLs. +- `enrich_contacts` — pass a list of emails, phone numbers, or LinkedIn profile URLs (`identifiers`) to resolve and enrich verified work email, personal emails, direct phone numbers, LinkedIn URL, job title, and company firmographics. +- `company_intel_search_people` — natural-language prospect search across companies and contacts with automatic email/phone/LinkedIn enrichment. +- `synter_audience` — build and export full B2B contact or US adult signal audiences at scale. +- `company_intel_get_profile` / `company_intel_get_signals` — inspect company firmographics, hiring velocity, and tech stack buying signals before outreach. + +## 2. Where outbound stands + +Start every outbound campaign conversation with the real state: + +- `get_crm_contacts` — leads by stage, active campaigns, replies awaiting review, and pipeline metrics. +- `render_card_preview` — recent human replies classified by intent and draft response status. + +Sending capacity is the fleet's, not one campaign's. Every campaign draws on shared sending domains and mailboxes, so one campaign's bounces damage all of them. + +## 3. Read reply numbers honestly + +**Roughly 90% of raw "replies" are out-of-office auto-responders.** Measured across live outbound campaigns: + +| campaign | inbox replies | automated | human | interested | +|---|---:|---:|---:|---:| +| 216 | 87 | 80 | 7 | 2 | +| 235 | 23 | 20 | 3 | 0 | +| 195 | 28 | 25 | 3 | 2 | + +Campaign 216 read as "85 replies, 3.0%" while producing **two** interested humans from 750 leads. + +Always quote human reply rate (`reply_rate_pct`: human replies ÷ sends) and `total_interested`. **Never quote `total_replied_including_automated`, and never compute a rate from it.** If the human figure is null, the filtered stats have not synced yet — say so rather than substituting the inflated one. + +Bounce rate above ~2% is a fleet emergency, not a campaign problem. Say so plainly. + +## 4. Before anything sends + +Cold email is outward-facing and irreversible. Confirm before every send, resume, or lead push — approval for one campaign never carries to the next. + +Two gates, both mandatory: + +1. **Qualification** — run the **outbound-research** skill. If you cannot name a checkable fact about a recipient, do not mail them. Matching an industry and a job title is not a fact. +2. **Deliverability** — every address verified via `enrich_contacts`, duplicates removed, nobody on the suppression list (`get_crm_contacts` / `get_crm_contacts`), and merge tokens rendering cleanly. + +## 5. Launch, pause, and reply management + +- `action_card` — creates a drafted campaign sequence. Creating is not sending without explicit launch confirmation. +- `action_card` — execute `approve`, `pause`, or `resume` on an outbound campaign card (`domain: "outbound"`, `action: "approve" | "pause" | "resume"`). Requires explicit user confirmation for launch or resume. +- `render_card_preview` — inspect pending prospect replies and draft responses. +- `action_card` — approve and send a drafted reply to a prospect. Ask first, every time. + +Pause immediately via `action_card` (`domain: "outbound"`, `action: "pause"`), without asking, if you see: bounce rate over 5%, a merge token printing literally in a sent email, or a lead who asked to be removed still receiving mail. Report it after. + +## 6. What not to do + +- Do not top up a campaign to fix a low reply rate. A campaign at 0.3% has an offer problem; adding leads spends mailbox reputation to buy nothing. +- Do not quote open rates. Open tracking is off across the platform. +- Do not send to someone already in the CRM, already met, or already a customer — see **outbound-research**. +- Do not write email bodies from scratch with generic AI prose. Templates are human-written; the model fills variable slots from real facts only, and never invents a number, a client name, or a stat. diff --git a/tools/directory-tools.json b/tools/directory-tools.json index d84d815..e3801c5 100644 --- a/tools/directory-tools.json +++ b/tools/directory-tools.json @@ -32,5 +32,13 @@ "enable_campaign", "pause_campaign", "update_campaign_budget", - "get_job_status" + "get_job_status", + "company_intel_search_people", + "company_intel_get_profile", + "company_intel_get_signals", + "enrich_contacts", + "synter_audience", + "get_crm_contacts", + "render_card_preview", + "action_card" ] diff --git a/tools/readonly-skills.json b/tools/readonly-skills.json index b1d76de..20aed82 100644 --- a/tools/readonly-skills.json +++ b/tools/readonly-skills.json @@ -18,6 +18,9 @@ "llm-visibility-audit", "media-plan", "negative-keywords", + "outbound", + "outbound-replies", + "outbound-research", "pixel-capi-auditor", "platform-benchmarks", "quickstart", diff --git a/tools/readonly-tools.json b/tools/readonly-tools.json index 336d060..df23446 100644 --- a/tools/readonly-tools.json +++ b/tools/readonly-tools.json @@ -24,5 +24,11 @@ "get_campaign_plan", "run_launch_preflight", "get_plan_execution", - "get_job_status" + "get_job_status", + "company_intel_search_people", + "company_intel_get_profile", + "company_intel_get_signals", + "enrich_contacts", + "get_crm_contacts", + "render_card_preview" ]