Personalising an application

The mail-merge version — same letter, different company name — reads exactly like what it is. Recruiters see dozens a week and it signals that the applicant did nothing, which is worse than a short honest note.

Real personalisation says something that could only be said about this application: this requirement against this experience, this product against this project. That requires actual information about the role and the candidate, not a template with a slot.

A blunt test settles it. Remove the company name and the requirement list; if what remains would work for any employer in the sector, nothing specific ever reached the generator, whatever the prompt asked for.

To say something specific, the system needs material: the posting in full, what the company does, the team if it is identifiable, and the candidate’s relevant history. Most of this is available without scraping anything sensitive.

Keep it bounded. Trawling for personal details about the hiring manager is an easy way to produce something that reads as intrusive rather than interested, and candidates will not notice the line was crossed until it is in front of the employer.

Cache what you gather per company rather than per application. A candidate applying to three roles at one employer should not trigger three research passes, and the company facts are stable in a way the posting-specific material is not.

• The posting, in full, including the parts below the fold
• What the company does, in its own words
• The candidate achievements relevant to this specific role
• Nothing personal about individuals beyond a publicly stated role

Not every fact is worth mentioning. A recent funding round says nothing about why this candidate fits, and reciting a company’s own marketing back to them is filler that a reader skips.

The references that land connect the company to the candidate: a technical problem in the posting they have solved, a market they know, a product they use. If the connection cannot be stated in one sentence, leave it out.

The best material usually comes from the candidate rather than from research. One question — why this one — produces something no amount of gathering can, and a scrappy two-sentence answer in their own words beats a well-composed observation the system found.

Generative systems always produce something, and something is often generic. Without a gate, the engine reliably emits competent, forgettable text for every application and the candidate never knows.

Check each output for specificity before it is used: does it name a concrete requirement, does it cite a real achievement from the record, would it read as wrong if pasted into a different application. If it passes that last test, it is not personalised and should be rejected.

Implement the gate in code rather than as a self-assessment. A model asked whether its own output is specific enough will say yes, which is the same reason every other guardrail belongs outside the thing it is constraining.

Verify the claims at the same point. Personalised text invents more freely than a CV does, because it is prose with no structure pulling it back, so anything unsupported by the profile should be removed rather than hedged.

Genuine personalisation needs genuine connection, and a candidate does not have a genuine connection to fifty companies a week. At that volume the system starts producing plausible sentences about companies the applicant does not care about.

Cap it. Personalise deeply where the candidate actually wants the role, and use a short honest application elsewhere. A modest number of specific applications outperforms fifty generated ones by a margin everyone in hiring recognises.

Make the tier an explicit choice at the moment a role is surfaced rather than at the moment of applying. Deciding later means deciding when someone is tired, and tired reliably chooses the automated path for a role that deserved the other one.

Producing nothing is a legitimate output. Where the candidate has no specific reason and the company offers nothing distinctive, a two-line note with a strong CV reads as efficient, and a manufactured page reads as exactly what it is.