How to use Claude in a Job Search

Over the last six months, I’ve learned a lot about how to best use Claude in my job search. This blog post summarizes my latest evolution in partnering with the AI – turning it into a career coach that makes the search process easier and more impactful.

While the below scenario is tailored to using Claude to improve your job search and to create a tailored, hallucination-free resume, the approach is also applicable to a wider range of scenarios – customer emails, marketing deliverables, documentation, etc. It’s been a fun learning experience as well as a huge time-saver for me.

This approach delivers 95%+ quality resumes that are 100% hallucination-free. I now spend 20 minutes tailoring a resume instead of 90 minutes recreating from scratch.

Claude as my Wingman

When I evaluate a prospective role (identified by my Job Search AI Agent), Claude helps me evaluate the role and recommends whether I should apply to it – I provide it with information about the company and the job description, and it provides two things:

1. An initial assessment of whether the role is a potential match
2. A starter resume that I can use for submission

My explicit goal here is to not apply to everything – I want Claude to help me filter out everything except excellent fits, and then to help me curate the best data points from my career to tell a story that resonates for the hiring manager. Claude brings an objective perspective – one that challenges me and helps me be more targeted.

Below is a screenshot of a conversation when I bring the AI a mismatch:

As you can see, Claude is brutally honest in its initial assessment – the prompt has been tuned to generate assessments that lean more towards ‘not a fit’ rather than wasting the time of both me and the hiring team(s).

Now, let’s compare that against a solid fit – where Claude not only provides an assessment, but outlines a fully detailed analysis on concern areas for both me and the hiring manager. This analysis helps you think through the full implications of a role and where you should lean into interview preparation. Personally, I am horrible at interviews, so this has helped me think about how I present myself and limit my ramblings on topics that I think are interesting, but that are irrelevant to a busy interviewer. I’m still working on this part of myself as I look for my next role.

Claude’s personality in our discussions is also something that I am quite proud of, but that’s a topic for another blog post.

Setting up the LLM Project Space

To accomplish this analysis, I set up a project space within Claude with a very long set of instructions and a collection of files that are to be used as reference points.

Note: Claude Projects is a workspace feature that allows you to provide custom instructions and reference documents that persist across all conversations in that project.

Claude Project Files

I run this as a project inside of my Claude account that is specific to my job search, which contains the files below. I used Microsoft Word files (.docx) for easy editing, but Markdown (.md) files work just as well and may be easier for version control if you’re technical.

Note on file naming: The numeric prefixes (01-, 02-, etc.) help Claude prioritize sources – it reads files in order when resolving conflicts or uncertainties.

• A master career fact document (01-FACTS-Master-Career-Facts.docx) with my core facts: timelines (career, education, locations), a list of verified metrics and achievements, and a list of what I never worked on (the edge cases that may cause hallucinations)
• A gaps and limitations document (02-FACTS-Gaps-And-Limitations.docx) that lists things that Claude should be wary of
  • Nevers: roles I never had, products I never worked on
  • Nuances: metrics that may look different between similar facts; title variations to be careful of (e.g., ‘VP of’ vs ‘Head of’; ‘PMM’ vs ‘Product Marketing Manager’)
  • Claims that the AI can claim, cannot claim, and explicitly gray areas
  • Areas and topics that the AI should seek clarification on
• An achievement context file for each major role/company (03-FACTS-Achievement-Context-Company.docx) that provides core facts for that role
  • Bulleted achievements/metrics that can be cited
  • Framing advice – the default framing for the facts, and additional variants that the LLM can use depending on the target role (e.g., DevRel role vs Dev Marketing role vs Product Marketing role)
  • Team context – for me, this outlines when someone reported to me or ‘dotted-line’ to me
• A bullet point library file for each major role/company (04-Libaries-Company-Bullet-Options.docx)
  • Resume-ready bullets for use by the LLM
  • I grouped them by emphasis area – this enables the AI to select the right wording and context based on the emphasis areas in the job post description and company type
    Example headers: Product Marketing Emphasis, Developer Relations Emphasis, Business Impact Emphasis, Technical Depth Emphasis, AI Emphasis, Startup/Growth Emphasis
• A sample resume (Resume - sample.docx) that I used for an application – providing the LLM with an example to emulate. The sample resume is for formatting only – Claude never pulls content from it, only layout and structure.

To be honest – the creation and curation of these files took a while (a couple of afternoons). It’s a lot of detail and feels repetitive, but the effort is inversely proportional to the hallucinations in the output.

To accelerate the process, I provided Claude with 10+ unique resumes I previously used and Claude did an initial population of the files above. From there, I waded in by hand to clean them up and ensure accurate positioning.

Project Instructions

The LLM instructions are long – I’m not going to lie – but let me break it down. I posted a generalized instructions file on GitHub, if you want to see it in action.

Below are the high-level sections of the instructions:

• Career Coach Context & Approach provides the context on how I want the LLM to behave and the coaching style I want it to follow
• Critical Data Accuracy Rules are my first set of rules that enforce no-hallucinations, telling the LLM where to pull facts from and to always ask when unsure
• Resume Format & Content Guidance inform the LLM how to create the resume file:
  • The Format Reference section tells it to start with the format of the sample resume (format only; no content!)
  • Content Sources (in priority order) informs it on which files to use pull content from and how to use the information
  • Bullet Point Optimization Standards provides instructions on how to create the bullets and reinforcing that it can only use verified content:
    • Target Format details the length and reinforces that ‘readability and impact matter more than rigid line counts’
    • Optimization Priority gives a step-by-step process for selecting and combining bullets from the available files
    • Safe Optimization Techniques gives a list of ALLOWED and NEVER ALLOWED steps, and includes examples of safe and unsafe bullet optimizations
  • Resume Creation Process provides a clear six-step process for creating the resume
  • Content Quality Standards provides guidelines for creating the resume section-by-section (professional summary, core competencies, and bullet points). This section focuses on final presentation and guidelines on length and which facts belong where in the document.
• Standard Job Analysis Process is the core of the instructions – what it is doing and how to do it. Everything above and below supports this set of nine steps.
• Validation Checklist provides the list of what must be checked before the resume can be considered ‘done’ and presented to the user
  • The checklist includes sections on Content Integrity, Optimization Quality, Format & Readability, and Hallucination Red Flags
  • System Boundaries reinforce a list of What I CAN Do and What I CANNOT Do to the LLM.
  • Key Principles reinforce using the optimization rules in the final review
• And Critical Reminders provides one last reinforcement of ‘no hallucinations’

How to implement this

If you want to build a similar system, here’s how to start:

1. Create a Claude Project for your job search
2. Add the LLM instructions similar to the full example I posted on GitHub
3. Upload a sample resume for formatting reference only
4. Start with your Master Facts file – timeline, key metrics, achievements
5. Add your Gaps & Limitations file – what you DIDN’T do (this prevents hallucinations)
6. Build bullet libraries from 3-5 of your best resumes that you have used
7. Pass the LLM a job title and description, and iterate on the content of your Claude Project

Initial setup: 2-3 afternoons. Worth every minute.

And note that you can accelerate steps 4-6 by providing some of your past resume submissions and asking Claude to create the fact and bullet library files based on the instructions in the project. Claude can a fantastic collaborator on helping setup and iterate on the project space.

Why this approach?

This approach is an evolution 6+ months in the making – starting with a simple prompt that asked Claude to be a career coach and run an analysis on my resume using a job description and ATS best practices. And while helpful, the resume tailoring process was time-intensive and everything was ‘a great fit.’

I followed the below journey as I learned about AI:

• I moved the instructions into a project and created a repository of prior work and prior resumes
• I separated guidelines from clear instructions
• I began providing greater context and asking it to select ‘the best resume’ from the project as a starting point
• I began doing debriefs with Claude to understand where hallucinations were coming from and how to improve the prompt

Example: Claude kept claiming I ‘led the [PMM | Xbox | Sales] team’ when I’d only ‘collaborated with’ them. The fix? A Gaps & Limitations file that explicitly listed roles I never held or teams that I led efforts with rather than managerially led.

This approach came from one particular debrief with Claude, as I was frustrated with an increasing number of hallucinations. The quality decrease arose from my resume portfolio – there was enough variation in the text that the LLM was getting confused. After a long ‘discussion’, we settled on the approach above.

This approach has worked out really well – the initial resumes using this format gave 80%+ quality, with issues in the summary (buzzword bingo) and in the bullets, but it was 100% free of hallucinations. After iterating on the instructions, I added the guidelines to tailor bullets, optimize readability, and stay hallucination-free. And I’m super happy with the end-result – I still would never submit the resumes as-is, it’s now 95%+ quality. I now spend 20 minutes to tailor it, rather than 90 minutes.

What’s next

Looking ahead, I will continue to evolve the project as I learn about LLMs.

Although it’s solid, I’m working improving the following areas of using Claude in a job search:

• Streamlining the instructions (there’s definitely some redundancy to remove)
• Integrating this approach with upstream AI agents that pre-filter job postings in n8n and Zapier
• Documenting my “lessons learned” for others implementing similar systems

If you implement this approach, I’d love to hear what works (and what doesn’t) for using Claude in your job search – either below or on LinkedIn.