jobtrackingapp/docs/redesign/AI_PRINCIPLES.md

AI Principles

Non-negotiable rule

AI MUST NOT INVENT USER FACTS.

It must never fabricate or silently imply:

• skills,
• technologies,
• qualifications,
• education,
• certifications,
• work history,
• job titles,
• responsibilities,
• achievements,
• metrics,
• projects,
• languages,
• dates,
• salary history,
• personal details,
• experience the user has not provided.

If information is missing, AI should say that it is missing or ask the user to provide it.

Source of truth

The Career Profile is the canonical source of user career information.

AI may also use information explicitly supplied in the current workflow, such as:

• a job description,
• user-entered notes,
• an uploaded document,
• a CV-specific override.

AI must preserve provenance wherever practical.

Example

Bad:

Added Kubernetes experience because the job requires it.

Good:

Kubernetes appears in the job requirements, but it is not present in your Career Profile. I have not added it to your CV.

AI interaction model

Use a hybrid model:

1. Contextual actions beside content.
2. A persistent or easily opened contextual assistant.
3. Command-palette AI actions where useful.

Avoid creating a disconnected AI area that forces users to leave their workflow.

Contextual awareness

The AI assistant should understand what the user is currently working on.

Examples:

On a Job page

• How good a match am I?
• What requirements am I missing?
• What should I emphasise on my CV?
• Prepare interview questions.
• Help me research this role/company.
• Draft a follow-up message.

In a CV

• Make this shorter.
• Make this more natural.
• Tailor this to the selected job.
• Check for unsupported claims.
• Improve readability.

In Career Profile

• Suggest missing details to add.
• Improve wording without changing facts.
• Identify duplicate information.

Review before apply

AI-generated modifications to user-owned content must not be applied silently.

Show a preview or diff before committing changes.

Minimum actions:

• Accept
• Reject

Where practical:

• Accept individual change
• Accept all
• Edit before applying

Evidence

For match analysis, AI should make it clear which Career Profile item supports a claim.

Example:

Strong match: C# / .NET

Evidence:

Work history → Analyst Programmer → C#/.NET backend development

Match scoring

If a numeric match score is shown:

• explain major contributors,
• distinguish required vs preferred criteria where possible,
• do not imply mathematical precision beyond the underlying method,
• do not penalise unknown information as though it were definitely absent unless the product logic explicitly chooses to do so.

User control

Users should be able to:

• disable or ignore suggestions,
• review AI output,
• edit output,
• understand whether content was AI-assisted.

Tone

AI should be:

• professional,
• supportive,
• concise,
• practical,
• non-judgmental.

Avoid exaggerated motivational language.

Privacy and logging

AI-related data handling should be explicitly documented during implementation.

Do not assume that all user content may be retained indefinitely.

Any future analytics about AI usage should prefer aggregate product metrics and avoid exposing private career content to administrators unless required and authorised.