Leveraging AI in the Job Hunt — While Still in the Trenches
Before we startI'm writing this while actively job hunting — not from the comfortable distance of a signed offer letter. I haven't "made it" yet. I think that actually makes this more useful: everything here is being tested in real time, in a real market. No polished retrospective, just what's working now.
I recently relocated to Toronto — stepping into what I can only describe as "The perfect storm." Unemployment hit 14.6% last July before easing back to 8.6% in February.
For virtually every role I find interesting, LinkedIn cheerfully reports: “Reposted 20 hours ago · Over 100 applicants.”
I had a choice: play the numbers game and mass-apply with generic materials, or treat this like the product problem it actually is. I went with option two.
To be clear: I'm not talking about mass-generating low-effort spam. The exact opposite. Fewer applications, done dramatically better — each one precisely targeted, genuinely tailored, and backed by real research on the company. Here's the full workflow I've built.
What's covered in this post
- The full 7-step process outline — from benchmarking to interview prep
- The prompts that actually work — including Yfaat Gal's benchmarking prompts
- Setting up the AI workflow — tech stack, context files, and the trigger prompt
Phase 1 — The Process
The Full Process Outline
This isn't a one-click solution. It's a structured process where AI amplifies each step — so the output is genuinely good, not just fast. Here's how it flows:
Compare your LinkedIn profile against professionals in similar roles at similar companies. Look for keywords, skills, measurable outcomes, and prerequisites you're missing or under-emphasizing.
Based on benchmarking insights, decide where you have the clearest competitive edge — role types, company sizes, verticals. Startup vs. corporate, product vs. growth. Deliberate focus beats broad volume every time.
Build one strong "master" version informed by benchmarking. This is the foundation Claude tailors per role — you never start from scratch again. Each application generates a new file version.
Configure Claude Cowork with your Notion context files. 20–30 minutes of setup. Hours saved on every application that follows. Full setup guide below ↓
Find a relevant role, run the workflow. Claude performs a gap analysis against the job description, proposes specific changes, and generates a tailored .docx. You review and approve throughout.
After the resume, Claude researches the company and drafts a cover letter with multiple angles to choose from. You select, refine, approve. The .docx lands in your Cowork folder and the application is logged to Notion automatically.
Once you land an interview, use your tailored resume and the job description to prompt Claude for a targeted prep session: likely questions, strong talking points from your experience, and research on the company's current priorities.
Step 5 in detail: How the tailored resume works
- Claude compares your master resume against the JD to discover gaps, then lists the deltas and proposes specific changes — current vs. proposed, section by section.
- You review each suggestion and provide your own input — e.g., experience that's relevant but wasn't on the resume, or corrections if Claude misread something.
- Accept or decline each suggestion. Claude incorporates your feedback and reworks if needed.
- A new .docx resume is generated for that specific role and dropped into your Cowork folder.
Step 6 in detail: How the cover letter works
- Claude drafts a cover letter drawing on three signals: your experience, the highlighted JD requirements, and research into the company's current priorities — announcements, product launches, blog posts, LinkedIn activity.
- Claude presents a list of possible angles to lead with. You pick the ones that best represent you.
- Claude generates the final letter. You can still edit before sending.
- The .docx is saved to your Cowork folder. Claude logs the full application to your Notion tracker — date, company, role, salary expectations, notes.
💡 The cover letter edge most people missThe real differentiator isn't keyword matching — it's company-specific context. Asking Claude to research recent blog posts, press releases, and LinkedIn activity before writing gives you insight into what the company is focused on right now. Referencing that signals you've done real homework in a way generic letters never can.
Phase 1 — Benchmarking Prompts
The Prompts That Actually Work
Big credit to Yfaat Gal, who shared these at a JVS Toronto event a couple of weeks ago. They've become a regular part of my benchmarking routine.
To identify what you're under-signaling on LinkedIn:
What am I under-signaling?""What outcomes do leaders in this role usually highlight?"
To decode what a job description is really screening for:
"What are the top 5 skills this job description is actually screening for?"
"Which of my experiences best match these requirements?"
🎤 Bonus: Interview prep promptsOnce you've landed the interview: "Based on this job description and my resume, what are the 5 most likely interview questions? What strong answers could I give based on my experience?" and "What should I know about this company's current focus before this interview?"
Phase 2 — Setup
Setting Up the AI Workflow
One-time investment. Once it's in place, every application runs from a single trigger prompt.
The tech stack
Core of the per-application workflow. Generates tailored .docx files into a desktop folder. Slightly slower than Projects, but token limits are less of a bottleneck. Works separately from Projects with no shared memory.
Working well
Best for the benchmarking phase — deep research sessions, LinkedIn profile analysis, strategy work. Maintains context across conversations. Faster, but hits daily limits quicker.
Working well
Beta. Theoretically useful for reading job descriptions in-browser. Still unreliable in practice — worth revisiting once it matures.
Beta — unreliable
Hosts all context documents. Claude reads instructions and resume via MCP integration, and writes application entries back to the tracker. Notion becomes the control center.
Solid
🤖 Honest Cowork reviewCowork has what I've come to affectionately call "selective blindness and amnesia" — it can't see other Claude Projects and has no memory between sessions. You give it full context every time via the Notion files. Once that's built into the workflow it stops being a limitation — and it's still my favourite tool for the job.
What you need in Notion first
- Master resume file — your optimized baseline, also stored in your Cowork desktop folder so Claude can access it directly.
- Instruction file — the full workflow instructions for Claude: what to do, in what order, how to format output.
- Scope & requirements file — your targeting parameters: preferred roles, company types, what to prioritize.
- Job application tracker — a Notion database where Claude logs each application automatically. Useful for follow-up timing and pattern-spotting over time.
The trigger prompt — starts every application
Once your Notion context files are in place and your master resume is in the Cowork folder, every application begins with:
Here's my context: [link to your Notion instructions/scope file]
Please run the full job application workflow for this role:
[paste job description URL — or text if the URL is blocked]
From there: gap analysis → accept or refine suggestions → tailored resume .docx generated → cover letter drafted with selectable angles → final letter approved → everything logged to Notion. Five to ten minutes, start to finish.
⚠️ Practical noteSome company career pages block Claude from reading the URL directly. If that happens, paste the job description text into the prompt manually. The workflow runs identically either way.
Where I'm at — and what's next
Still in the search. But applying with more precision, more confidence, and significantly less 11pm panic-editing. That's a real improvement worth sharing.
Next in this series: The full Cowork setup walkthrough — the instruction file structure, the Notion tracker schema, and exactly how I've wired Claude's memory workaround. If you're building something similar and want to compare notes, reach out — I'd genuinely enjoy the conversation.
