ApplyOnce

04 / FIELD NOTES

The repetition was the actual problem

The initial frustration was not that job application forms were technically difficult. It was that the same administrative work kept returning.

Name, contact information, employment history, profile links, education, language proficiency, and notice period already existed somewhere. Yet every application system treated them as new information. During a period of applying to many roles, even simple copying and pasting became a noticeable amount of work.

Starting with deterministic matching

I did not want every field to require an AI request. Many form labels can be matched with simpler rules.

ApplyOnce normalizes labels and handles several common synonyms. A remembered answer can have multiple aliases, and the extension calculates a similarity score before treating a field as a match. The matching threshold can be adjusted when the default is either too strict or too permissive.

This makes the common path local, immediate, and free of API usage.

Building memory one answer at a time

When ApplyOnce finds an unfamiliar field, it does not silently guess. The popup shows the question and lets the user enter or select an answer.

Saving the answer also records useful signals from the field, including its label, name, placeholder, and autocomplete value. The next website may use different wording, but those aliases give the matcher more context.

Users can inspect, edit, delete, export, import, or completely clear this local memory.

Using AI as an optional interpreter

AI is most useful during initial setup and when a field cannot be understood through deterministic matching.

The profile importer accepts a user-selected CV or LinkedIn profile. It asks the model to extract factual information while explicitly avoiding unsupported assumptions about salary, identity, demographics, work authorization, and similar sensitive topics.

For ambiguous fields, AI receives the structured local profile, saved answers, field label, and available options. Its output is presented as a suggestion rather than an automatic decision.

This feature requires an OpenAI API key, not a ChatGPT login or session token. It consumes API tokens and does not have access to the user’s existing ChatGPT conversations.

Keeping the public build clean

I used my own CV and LinkedIn profile while developing the seeded personal version. That data could not be included in a public repository because extension source files are visible to anyone who downloads them.

The public build therefore starts with an empty profile and empty memory. Each user has to import or enter their own information.

The API key is kept in Chrome session storage instead of being bundled with the source or permanently saved by the extension. For a larger public release, I would still prefer a backend authentication and proxy layer over asking general users to manage their own API keys.

The browser is not a consistent environment

Native HTML inputs, text areas, select elements, checkboxes, and radio groups are supported by the current version.

The difficult part is that many recruitment platforms build their own form components. A control may visually look like a normal dropdown while behaving like a collection of nested elements managed by a JavaScript framework.

Workday, SuccessFactors, and similar platforms will need dedicated adapters and a larger compatibility test matrix. The current version should be treated as a working foundation rather than universal job-form automation.