Scripts & Skills — Build Case Study

I turned my job search into a five-agent pipeline — and ran it like a lifecycle program.

Every teardown on this site was produced with help from a system I built: a Python pipeline on the Claude API that discovers matching roles, tailors my resume, drafts the teardown scaffold, red-teams it from the hiring manager's chair, and finds the humans to send it to. I built it the way I'd build a marketing automation stack — routed to the cheapest model that does each job well, cached aggressively, and instrumented so I can see cost per step the same way I'd track cost per send.

Claude API · Python 3-tier model routing Prompt caching · ~90% cheaper cached reads Per-step token & cost telemetry Gmail + Google Docs + Hunter.io integrations

01 — The Pipeline

Five steps, each its own agent.

One orchestrator (python run.py all --company Noom --role "Head of Lifecycle") runs the chain end-to-end, or any step runs alone.

Discover

Scans job boards for lifecycle/retention roles that match my profile, scores fit against my resume, and writes a ranked shortlist to Google Sheets. Runs on a schedule, reads confirmations via the Gmail API.

Sonnet

Tailor

Takes a job URL, maps the posting's language to my actual experience, and drafts the tailored resume bullets and application answers — flagging any claim it can't ground in my history so nothing gets invented.

Opus

Teardown

Assembles my raw field notes and screenshots from testing a company's funnel into a structured teardown scaffold — HTML page plus Google Doc — which I then edit, fact-check, and finish by hand. The analysis is mine; the assembly is automated.

Opus + Sonnet

Critique

Re-reads the finished teardown three times — as the company's CEO, CMO, and the hiring manager — and returns the objections each persona would raise. The shared context is cached once and reused across all three passes.

Opus

Outreach

Finds the right contacts via Hunter.io, then drafts short cold emails that lead with the teardown's most useful finding — value first, ask second. Drafts land in Gmail for my review; nothing sends itself.

Sonnet

02 — The Engineering Choices

Built like a retention stack: route, cache, measure.

The same discipline I'd apply to an ESP bill applies to an API bill. Three choices keep the pipeline fast and cheap without giving up quality where it matters.

Model Tiering

Every agent runs on the cheapest model that does its job well

Content that a human will read (resume bullets, critiques, outreach) runs on Opus. Research, scoring, and scaffolding run on Sonnet at 60% of the input cost. The routing lives in one config line per agent, so when a smaller model gets good enough, the swap is a one-word change — the code comments literally note which tiers are next in line for a downgrade test.

Prompt Caching

Shared context is written once, cached, and reused — cached reads cost ~10% of fresh input

My resume, profile, and each step's system prompt are marked as cacheable prefixes. The critique step is the showcase: the full teardown context is cached on the first persona pass, and the CEO, CMO, and hiring-manager reads all hit the same cache. Stable content up front, dynamic content last — the same rule that makes email templates cacheable in an ESP.

Cost Telemetry

Every call logs tokens, cache hit rate, and dollars — and low hit rates get flagged automatically

Each API call appends a labeled row to a usage log; a report script aggregates cost per step and flags any label with a cache hit rate under 20% — the usual culprit being a timestamp near the top of a prompt breaking the prefix match. It's the same instinct as monitoring deliverability: you don't find out something broke from the invoice.

Why this is on a marketing portfolio

Lifecycle marketing is becoming an orchestration job — this is what my orchestration looks like.

Segmentation, routing, suppression, cost-per-touch, instrumentation: the concepts that make a lifecycle program work are the same ones that make an AI pipeline work. I'd rather show working code than claim "AI-curious" on a resume. The system produced the applications, the teardowns on this site, and the outreach that got them read.