Career-Ops: Finding the Right Tool, Making It Mine

Career-Ops automates analysis, not decisions. Every evaluation report is reviewed before I decide whether to apply. Every generated PDF is inspected to verify that keywords are reformulated from real experience, not fabricated. The apply mode fills form fields but stops before clicking submit — I make the final call on every application. This is the same HITL philosophy behind any CRM or applicant tracking system: the tool organizes, filters, and prepares, but a human provides the judgment. The system explicitly discourages applying to roles scoring below 4.0 out of 5, recommending against wasting both my time and the recruiter's time on poor-fit applications. Quality over quantity is a design principle, not a slogan.

A script pipeline follows fixed rules — it cannot reason about context, weigh ambiguous signals, or generate narrative analysis. Career-Ops uses Claude Code as an LLM agent that reads the full job description, cross-references it against my CV and project proof points, and produces a 10-dimension evaluation with contextual reasoning for each score. It detects when a role title says Senior but the responsibilities suggest Staff level. It recognizes when a startup listing compensation as competitive likely means below market. It reformulates keywords from my real experience to match the JD vocabulary without fabricating skills I do not have. The output is a narrative evaluation report with specific rationale per dimension, not a template with blanks filled in. That reasoning capability is why an LLM agent outperforms any deterministic script for this task.

Zero marginal cost per evaluation. Career-Ops runs on a Claude Max subscription, which covers all Claude Code usage across every project: portfolio development, OpenClaw development, and Career-Ops itself. Each evaluation uses approximately 4,000-6,000 input tokens for the JD plus CV context and generates around 2,000 tokens of evaluation output. Under the Max plan, this volume is well within the included capacity. There is no per-API-call billing, no token metering, and no usage caps that would throttle batch processing. The economics make it viable to evaluate every offer the scanner finds rather than pre-filtering manually.

The apply mode uses Playwright to navigate to the application form, take a snapshot of all visible fields, and read the page structure including labels, placeholders, and required field indicators. It then retrieves the cached evaluation report for that company and role, which contains the 10-dimension scoring, matched proof points, and relevant keywords. From this context, it generates coherent responses for each form field: cover letter text draws from the evaluation personalization section, the motivation question pulls from the company research, and skills questions reference the specific proof points that scored highest in the CV match dimension. Every generated response is displayed for my review before any field is filled. I approve, edit, or reject each answer individually. The submit button is never clicked by the system — that is always my action.

The dedup system works in two layers. First, scan-history.tsv stores every URL the scanner has ever seen with a timestamp and source portal. Each new URL is checked against this file using exact string matching after normalization (stripping tracking parameters, trailing slashes, and query strings). Second, even if the URL is new, Career-Ops performs a fuzzy match against applications.md using a normalized company name plus role title comparison. This catches the common case where the same job is posted on multiple portals with different URLs. The company name is lowercased and stripped of suffixes like Inc or GmbH, and the role title is compared after removing seniority prefixes. If both layers pass, the offer enters the pipeline. If either catches a match, it is silently skipped. The result: zero duplicate evaluations.

Yes. Career-Ops is open source on GitHub, created by santifer. You need Claude Code with Playwright browser access enabled for portal scanning and form filling, plus a structured working directory following the data contract: cv.md as the canonical resume, a profile configuration file with your target roles and salary range, and a portals configuration listing the job boards and company career pages to scan. Each of the 12 operational modes is defined as a Claude Code skill file with its own context window, rules, and tool permissions. Fork the repository and customize the scoring dimensions, archetypes, and evaluation criteria for your own career profile. The onboarding flow walks you through setup step by step. That is exactly what I did — forked it, customized the profile for my AI Developer and Autonomous Systems targets, and started using it.