--- name: phase-2-packet description: Build a Phase 2 outreach packet for a single greenlit role. Triggers on "build packet for X", "tailor CV for X", "Phase 2", "packet X", "/packet", or any request to prepare an application for a specific company / role. Pulls the full JD via the ATS redirect, identifies the right contacts, generates a JD-tailored CV (docx + PDF) per the discipline rules, drafts channel-appropriate outreach (Connect+Note + longer DM), and saves the packet under outreach/{slug}/. Stops before Phase 3 - waits for the user's explicit "send" approval. --- # phase-2-packet Build the Phase 2 outreach packet for a single role. Pull the JD, identify contacts, tailor the CV, draft outreach, save the packet, present in chat. ## When to invoke The user says: "build packet for X", "tailor CV for X", "Phase 2 for X", "packet X", "/packet X", or "let's go" after a Phase 1 greenlight summary. ## Inputs The user names a role by company or by slug. Resolve to: - `company` (string) - `role_title` (string) - `linkedin_url` (the `/jobs/view/{id}/` URL) - `slug` (kebab-case, e.g. `company-role-slug`) If the role is in the dashboard but no slug yet, generate one. If the role isn't in the dashboard, scaffold a minimal entry first. ## Workflow ### Step 1: Pull the full JD Delegate to the `jd-extractor` agent (or run inline). The LinkedIn jobs page renders a skeleton in unauthenticated views; the real body lives behind the "Apply on company website" redirect. Recover it via: 1. Navigate to the LinkedIn jobs URL in Chrome 2. Find the Apply link's `href` 3. Decode the `/safety/go/?url=...` query param to recover the destination ATS URL 4. Navigate to the destination (Teamtailor, Workable, Greenhouse, Lever, Ashby, company-careers page) 5. Extract the full JD body text including: - About the role / mission - Responsibilities / day-in-the-life - Required experience / qualifications - Comp band if stated - Anti-bot screener words / phrases (look for `mention the word "..."` or `include "..." in your application`) - Reporting line and sit-on-SLT signals - Customer / partner names called out by name - In-office requirement (number of days) - Pre-employment check requirements (FCA, security clearance, criminal/credit checks) ### Step 2: Sanity-check tier classification If the JD body reveals signals that contradict the Phase-1 card-level triage (e.g., title says "Head of Product" but body says "individual contributor / chief of staff"), STOP. Surface the mismatch to the user with two options: - A) **Skip** - reclassify as Skip, log reason - B) **Reframe and proceed** - tailor for the actual scope (e.g., "I have been the CPO three times; force-multiplier IC at this scale-up is by choice") - C) **Apply low-effort** - careers application only, no Connect+Note, no CV refresh Do NOT proceed to Step 3 without the user's call. ### Step 3: Identify contacts Apply the contact-discovery cascade from `${CLAUDE_PLUGIN_ROOT}/config/channels.yaml`: 1. LinkedIn job poster (the "Meet the hiring team" panel) 2. Hiring manager named in JD body (often "Reports to: ..." or specific name) 3. 1st-degree connection at the company in product/eng leadership 4. 2nd-degree with strong mutuals 5. Founder / CEO for founder-led companies 6. Recruiter-first (de-anon path) for recruiter-masked listings For each candidate contact, check: - Degree of connection (1st / 2nd / 3rd) - Open Profile status (do they show a `Message` button without InMail paywall?) - Public LinkedIn URL When the role's reporting-to-CTO is in another timezone or appears to be transitioning out (e.g., now CEO at a different company), consider parallel reach to both the CTO and the CEO of the hiring company. ### Step 4: Tailor the CV Delegate to the `cv-tailor` agent. Pass: - `--role-slug` - the kebab slug - `--baseline` - path to the user's master CV (default: `${CLAUDE_PLUGIN_ROOT}/templates/cv-baseline.docx`, but the user may have replaced this with their own) - Replacements derived from the JD per the discipline: - Headline-positioning fragment (one phrase that captures the role's mandate) - Adjacency sentence (last sentence of executive summary, lands the role-specific analogue) - One or two career-highlight cells (rename + body) - Italic context line per relevant past role - Geo-conditional header per `geo-rules.cv_header_templates` - "Tailored for {company}, {role}" tag The agent applies the discipline rules from `outreach-style.yaml`: - Identity stable - Max 2 of 6 highlight cells changed - One phrase per past role's italic context line - No em/en dashes anywhere - Geo header conditional Outputs `{cv_filename_prefix}-{slug}.docx` and `.pdf` in `outreach/{slug}/`. ### Step 5: Draft outreach Delegate to the `outreach-drafter` agent. Generate: 1. **Connect+Note (Premium 300 char or standard 280 char)** for the primary contact, leading with one concrete JD hook + one pattern-match anchor 2. **Connect+Note** for the backup contact (parallel reach) 3. **Longer DM (120-180 words)** for use after a connection accepts, or if Open Profile path lands Apply `outreach-style.yaml` rules: peer-to-peer, no banned phrases, no em/en dashes, signed off with first name. For role-specific anti-bot screener words detected in the JD body, draft a Summary / cover-letter blurb that includes the word naturally (per `outreach-style.yaml.screener_test_words.action`). ### Step 6: Save the packet Create `outreach/{slug}/` and write: - `{cv_filename_prefix}-{slug}.docx` - `{cv_filename_prefix}-{slug}.pdf` - `jd.md` - JD notes from `${CLAUDE_PLUGIN_ROOT}/templates/jd.md.template`, populated - `outreach.md` - outreach drafts from `${CLAUDE_PLUGIN_ROOT}/templates/outreach.md.template`, populated ### Step 7: Append to dashboard Append a new role entry to the dashboard's `ROLES = [...]` array with: - `id` - one greater than the current max - `slug` - the kebab slug - `role`, `company`, `location`, `comp`, `score` - `jdUrl` - the canonical `/jobs/view/{id}/` URL (NOT the search URL) - `contact` - primary contact name + role - `contactUrl` - primary contact's LinkedIn URL - `channels` - planned channels - `sendDate: null` - until Phase 3 - `status: "drafted"` - `note` - long-form note capturing the score rationale, friction flags, and pattern match anchors ### Step 8: Present in chat Output: - Score and tier - JD URL + primary contact LinkedIn URL (per `feedback_packet_links.md` rule: ALWAYS show JD URL + contact URL in chat at greenlight, not just in saved files) - Connect+Note draft (verbatim, with char count) - Friction flags worth confirming with the user before send - Computer:// links to the saved CV PDF, DOCX, jd.md, outreach.md ### Step 9: STOP Do NOT proceed to Phase 3. Wait for the user's explicit "send" per role. ## Approval gate Before saving the packet, surface for the user: - Tier classification post-JD-extraction (may have changed from Phase 1) - Comp band alignment - Geo friction - Title-scope alignment (line-product vs IC vs specialist) If any of these have shifted from the Phase-1 read, surface the shift before generating the CV. The user gets the call. ## References See `${CLAUDE_PLUGIN_ROOT}/skills/phase-2-packet/references/cv-tailoring-discipline.md` for the canonical replacement-key pattern. See `${CLAUDE_PLUGIN_ROOT}/skills/phase-2-packet/references/contact-discovery-recipes.md` for LinkedIn search patterns to identify contacts.