How We Designed Prompts That Give Genuinely Useful CV Feedback

Early versions of Cvaluate sounded like every other AI career tool: "Use stronger action verbs." "Quantify your achievements." "Tailor your CV to the role." Technically true, practically useless. Candidates do not need a model to repeat blog posts — they need feedback that could only apply to their CV. Prompt engineering for feedback that is specific, honest, and actionable took more iterations than we expected, and taught us that LLM prompt design for production is closer to curriculum design than to magic spells.

The anti-generic philosophy

Our content team jokes that we are allergic to the phrase "leverage synergies". The product team shares that allergy. Prompts explicitly ban hollow coaching language and require each weakness to cite a location in the document. If the model cannot point to a bullet, it cannot criticise it.

We also rank feedback by leverage. Fixing a parsing-breaking layout matters more than swapping "helped" for "facilitated". Prompts instruct the model to prioritise issues that affect screening — missing keywords, invisible experience, unquantified outcomes in relevant roles — before stylistic nitpicks.

What version one got wrong

Our first evaluation prompt optimistically asked for "constructive feedback". Typical output:

• Paragraphs of encouragement with no actionable edits
• Rewrites that changed job titles or invented metrics
• Identical advice regardless of seniority or industry
• Strengths lists that read like horoscopes — specific wording, vague substance

Users told us they wanted what a good recruiter friend would say: "This bullet is weak because X; try Y instead." That clarity drove the next generation of prompts and tied directly into hallucination guardrails.

Structuring the evaluation prompt

Modern prompts are modular:

1. Role and rubric — you are an expert reviewer; score against explicit dimensions (parsing survival, keyword fit, evidence of impact, clarity).
2. Inputs — structured CV JSON, optional job description, parsing confidence metadata.
3. Output contract — JSON schema for scores, strengths, weaknesses, rewrites, requirements matrix.
4. Negative constraints — do not invent facts; do not recommend white-text tricks; do not guarantee interviews.
5. Examples — few-shot pairs of bad generic advice vs good specific advice (not user data).

Separating rubric from examples lets us tune tone without rewriting the entire instruction set. It also makes regression testing feasible when we swap model versions.

Resume rewrite prompts that actually help

Resume rewrite prompts are the highest-risk part. Models love to upgrade candidates into fiction. We require rewrites to preserve factual claims, flag when a metric is missing rather than inventing one, and show before/after pairs side by side.

Before: Responsible for social media for the brand.

After: Managed organic and paid social for the brand across three channels, growing follower engagement through A/B-tested content calendars.

The "after" adds structure and plausible mechanism language without fabricating a percentage. If the extraction layer already includes a metric, the prompt pushes the model to foreground it. That is calibrated AI feedback — ambitious about clarity, conservative about facts.

When a job description is present

JD-aware prompts add a requirements matrix: each must-have and nice-to-have mapped to evidence, partial evidence, or gap. The model must quote or paraphrase from the CV when claiming a match. This is where generic "mirror the job description" advice becomes a checklist you can act on in fifteen minutes — the same workflow we describe in our 15-minute tailoring guide, but automated.

Calibrating tone

Job seekers arrive stressed. Prompts instruct direct language without mockery. "This bullet is vague" is fine; "this is terrible" is not. We A/B tested user reactions to blunt versus soft phrasing and found that specificity matters more than politeness — people accept hard truths when the rewrite shows the fix.

Senior CVs get different emphasis than graduate CVs. Prompt variants reduce unhelpful "add more experience" notes for juniors and push principals on strategic outcomes instead of task lists.

How we test prompt changes

Every prompt revision runs against a frozen set of anonymised CV fixtures — multi-column disasters, career changers, sparse graduate CVs, executive summaries. We diff JSON outputs for schema compliance, grounding errors, and score stability (see scoring reproducibility). Human reviewers spot-check rewrites for plausibility.

Prompt design never finishes. New model releases shift behaviour; we retune constraints rather than assuming yesterday's prompt still holds.

Why this matters for you

The promise of AI CV tools is speed with substance. Prompts are where that promise is kept or broken. For context on how LLMs fit into hiring, read our guide to AI in hiring. To see the current prompt generation on your document, run your CV through Cvaluate's free analysis — compare the rewrites to what a generic chatbot suggests, and notice which ones could only apply to you.