Track application

You are managing the user's job application pipeline.

STEP 1: Read current memory

Read ${CLAUDE_PLUGIN_ROOT}/references/CUSTOMIZATION_MEMORY.md to understand current patterns and what has been learned. This is the user's real (git-ignored) file. If it does not exist yet, seed it by copying ${CLAUDE_PLUGIN_ROOT}/references/CUSTOMIZATION_MEMORY.example.md to that path, then proceed.

STEP 2: Determine the action

Based on the user's input:

If logging a new application:

• Extract: company name, role title, date applied, source (referral / direct / recruiter)
• Determine company type (AI / Consumer / Enterprise / Fintech / Telco)
• Confirm: was the resume and cover letter generated using this plugin?
• Ask: any customizations made that worked well?

If updating an existing application:

• Extract: company, new status, outcome notes
• If rejected: what was the stated reason? What can we learn?
• If progressing: what round, who is interviewing?

If checking pipeline:

• Summarize all open applications by status
• Flag any applications needing follow-up (no response after 7 days)
• Show conversion rates by company type

STEP 3: Update CUSTOMIZATION_MEMORY.md

Read the current CUSTOMIZATION_MEMORY.md, then propose specific additions to the "Successful Patterns" or "Unsuccessful Patterns" sections based on the outcome.

Write the updated file back to: ${CLAUDE_PLUGIN_ROOT}/references/CUSTOMIZATION_MEMORY.md

Format additions as:

#### [Company Type] - [Company Name] - [Outcome]
Date: [date]
Resume strategy: [what was emphasized]
Cover letter hook: [what opener was used]
Result: [screen / reject / offer / pending]
Learning: [what this tells us for future applications]

STEP 4: Persist the structured tracker (machine-readable — required)

CUSTOMIZATION_MEMORY.md holds the prose learnings. It is NOT queryable, so it cannot drive the patterns (rejection analysis) or followup (cadence) skills. Those skills read a structured tracker instead. Maintain it here.

${CLAUDE_PLUGIN_ROOT}/references/applications.md is the user's REAL tracker. It is git-ignored (so private job-search data never reaches a public repo). Read it; if it does not exist, create it — seed it from ${CLAUDE_PLUGIN_ROOT}/references/applications.example.md if that exists, and put the privacy header at the top (see the example file). It has exactly one table with this schema — keep the columns in this order:

# Application Tracker

| Num | Company | Role | Type | Status | Applied | Responded | Interview | Source | Score | Outcome reason | Contact | Follow-ups | Last follow-up |
|-----|---------|------|------|--------|---------|-----------|-----------|--------|-------|----------------|---------|-----------|----------------|
| 1   | Northstar AI | Director of Product | AI | applied | 2026-06-30 | — | — | referral | 4.2 | — | [email protected] | 0 | — |

Field rules:

• Num — sequential, never reused.
• Type — AI / Consumer / Enterprise / Fintech / Telco / Other.
• Status — one of: evaluated, applied, screening, responded, interview, offer, rejected, discarded, skip. Use these exact tokens (the patterns/followup skills classify on them). discarded = went dead with no formal rejection; skip = self-filtered before applying.
• Applied — YYYY-MM-DD you submitted the application.
• Responded — YYYY-MM-DD the company/recruiter first responded with a real reply (not an auto-ack); — until it happens. Set this the moment a reply lands (even if you also move Status to screening/responded). followup keys its urgency off it, and patterns reads it as proof the row reached the screening stage.
• Interview — YYYY-MM-DD of the first interview (scheduled or held); — until it happens. followup keys momentum off it, and patterns reads it as proof the row reached the interview stage.
• Source — referral / direct / recruiter.
• Score — the evaluation score (X.X/5) if the role was evaluated; else —.
• Outcome reason — on rejected/discarded/skip, the stated or inferred reason (e.g. "geo-restricted", "stack mismatch", "seniority gap"). This is what patterns mines — never leave it blank on a negative outcome.
• Follow-ups — integer count of follow-ups actually SENT (not drafted); starts at 0. followup bumps it on a confirmed send and treats 2+ as COLD.
• Last follow-up — YYYY-MM-DD of the most recent follow-up sent, else —.

On a new application: append a row (Responded, Interview —; Follow-ups 0). On an update: rewrite the existing row in place (build the new row as a fresh line, don't mutate columns piecemeal) — and when a reply or interview happens, fill the Responded / Interview date, don't just change Status. Write the file back. If the file still contains the shipped fictional example rows, delete them when appending the first real row.

STEP 5: Output pipeline summary

Always end with a pipeline summary table (derived from applications.md):

APPLICATION PIPELINE

| Company | Role | Status | Date | Follow-up Needed? |
|---------|------|--------|------|-------------------|
| [name]  | [role] | [status] | [date] | [yes/no] |

Conversion Rates:
- Applications to screens: X%
- Screens to interviews: X%
- Interviews to offers: X%

Patterns Working: [what's getting responses]
Patterns to Change: [what's not working]

Zero-denominator guard (mandatory — same rule as the patterns skill): each conversion rate above is numerator ÷ denominator. If a denominator is 0 (e.g. no screens yet, so "screens to interviews" has nothing to divide), print "insufficient data for this rate" for that line — never divide by zero, and never print 0% when what you mean is "no data yet". A 0% reported on a zero denominator reads as a real failure signal when it is really just absence of data.

If there are fewer than 5 applications, note: "Pipeline is thin — consider expanding outreach or applying to more roles."

When the tracker has 5+ outcomes, suggest: "Run the patterns skill to analyze what's converting." When any application is aging past its cadence, suggest: "Run the followup skill to check follow-up cadence."