LinkedIn Profile Optimizer Agent

An AI agent that audits your LinkedIn profile against general best practices and returns an actionable report: an overall score, per-section scores and critiques, ready-to-paste rewritten copy, and a suggested keyword list to improve how easily recruiters find you.

You give it your profile as a LinkedIn "Save to PDF" export — no LinkedIn login, no API keys for LinkedIn, no scraping. The agent parses the PDF, sends the sections to Claude in a single structured call, and renders a clean report in your browser.

Status: in active development. See docs/superpowers/specs/2026-07-14-linkedin-profile-agent-design.md for the full design.

---

Why PDF export instead of the LinkedIn API?

LinkedIn's profile-editing API is not available to most developers, and scraping public profiles violates LinkedIn's Terms of Service and is fragile. Exporting your own profile to PDF is fully within your rights, needs no authentication, and works for everyone. LinkedIn provides this via Profile → Resources → Save to PDF.

---

Features

• Overall profile score (0–100) with a per-section breakdown.
• Section-by-section critique for Headline, About, Experience, Education, and Skills.
• Ready-to-paste rewrites of each section, with a copy-to-clipboard button.
• Keyword suggestions to improve search visibility for recruiters.
• Local-first & private — runs on your machine; your profile is only sent to the Anthropic API for analysis, never stored.

---

How it works

A fixed, easy-to-reason-about pipeline: parse → analyze → render.

Browser (upload page)
    │  POST /analyze  (your LinkedIn PDF)
    ▼
FastAPI backend
    │
    ├─ pdf_parser   PDF bytes ─► { headline, about, experience, education, skills }
    │
    ├─ analyzer     sections  ─► Claude (one structured call) ─► Report
    │
    └─ render report page  ◄── Report (scores, critiques, rewrites, keywords)

Scoring rubric

The agent scores your profile against these best practices:

Section	What "good" looks like
Headline	More than a job title — states value/specialty, keyword-rich
About	Strong hook in the first lines, first person, quantified impact, clear CTA
Experience	Achievement-oriented bullets with metrics, not a list of duties
Skills	Covers relevant industry terms so recruiter searches surface you
Completeness	Every section present and with meaningful depth

---

Tech stack

• Python 3.11+
• FastAPI + Uvicorn — backend and local server
• Jinja2 — server-rendered pages (no SPA, no build step)
• pypdf — PDF text extraction
• Anthropic SDK — Claude for the analysis (default model claude-sonnet-4-6, configurable)

---

Getting started

These steps describe the intended setup. Application code is being implemented against the approved design spec.

1. Clone and install

git clone https://github.com/phpdev-expert/linked-in-AI-agent.git
cd linked-in-AI-agent
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Configure your Anthropic API key

export ANTHROPIC_API_KEY="sk-ant-..."
# optional: override the default model
export ANTHROPIC_MODEL="claude-sonnet-4-6"

Get a key from the Anthropic Console.

3. Run the app

uvicorn app.main:app --reload

Open http://localhost:8000, upload your LinkedIn PDF export, and view the report.

---

Usage

1. In LinkedIn, go to your profile → Resources → Save to PDF.
2. Open the app at http://localhost:8000.
3. Upload the downloaded PDF.
4. Review your scores, critiques, and rewrites — copy the rewritten sections straight into LinkedIn.

---

Project layout

app/
  main.py          FastAPI routes; serves the pages
  pdf_parser.py    PDF bytes → sectioned profile text (pure, no network)
  analyzer.py      sections → Report via Claude (LLM client injected)
  models.py        Pydantic schemas (also the JSON contract for Claude)
  templates/
    upload.html    upload page
    report.html    report page (scores, critiques, rewrites, keywords)
docs/superpowers/specs/
  2026-07-14-linkedin-profile-agent-design.md   full design spec
tests/             unit + integration tests (LLM mocked)

---

Development

Modules are intentionally small and independently testable:

• pdf_parser is pure — unit-tested against a sample PDF fixture.
• analyzer takes the Anthropic client as a parameter, so tests run against a mocked client (no network, no cost).
• main has one integration test for POST /analyze with the LLM mocked.

pytest

---

Privacy

Your profile PDF is parsed locally and its text is sent to the Anthropic API only for the duration of the analysis request. The app does not persist your profile or the generated report.

---

License

MIT (to be added).