Skip to main content
Version: Next

Ghostwriter

What it is​

Ghostwriter is the per-job AI chat assistant in JobOps.

Ghostwriter uses:

  • current job description and metadata
  • reduced profile snapshot
  • global writing style settings
  • the configurable Ghostwriter system prompt template from Settings

The UI behavior is one persistent conversation per job, shown in the right-side drawer from job details.

Why it exists​

Ghostwriter helps you produce job-specific writing quickly while preserving consistency with your profile and style settings.

Typical use cases:

  • role-specific answer drafting
  • cover letter and outreach drafts
  • interview prep tied to the job description
  • rephrasing with tone constraints
  • multilingual drafting when you want replies in a specific language

How to use it​

  1. Open a job in discovered or ready.
  2. Open the Ghostwriter drawer.
  3. Enter your prompt and stream a response.
  4. Use the Copy button on any completed Ghostwriter reply to copy the full output.
  5. Stop or regenerate responses when needed.

Writing style settings impact​

Global settings affecting generations:

  • Tone
  • Formality
  • Constraints
  • Do-not-use terms

Ghostwriter follows the output language you request in your prompt. For example, Ecris en français should produce a French reply.

If you want a persistent default language, set it in Settings → Writing Style & Language.

If you need to change Ghostwriter's base behavior more deeply, edit the Ghostwriter prompt template in Settings → Prompt Templates. That editor is advanced on purpose: removing instructions or placeholders can make responses less reliable, but reset restores the default template quickly.

Do-not-use terms are passed as guidance in the prompt. They are not enforced by a hard post-generation filter, so the model should avoid them but may still use them occasionally.

Defaults:

  • Tone: professional
  • Formality: medium
  • Constraints: empty
  • Do-not-use terms: empty

Context and safety model​

  • Job snapshot is truncated to fit prompt budget.
  • Profile snapshot includes relevant slices only.
  • System prompt enforces read-only assistant behavior.
  • Logging stores metadata, not full prompt/response dumps.

API surface​

  • GET /api/jobs/:id/chat/messages
  • POST /api/jobs/:id/chat/messages (streaming)
  • POST /api/jobs/:id/chat/runs/:runId/cancel
  • POST /api/jobs/:id/chat/messages/:assistantMessageId/regenerate (streaming)

Compatibility thread endpoints remain, but UI behavior is one thread per job.

Common problems​

Responses feel too generic​

  • Verify the job description is complete and current.
  • Confirm style constraints in Settings are specific enough.
  • If you customized the Ghostwriter prompt template, compare it with the default or reset it to confirm the regression comes from the template.

Generation quality is lower than expected​

  • Check model/provider configuration in Settings.
  • Tighten prompts with explicit output intent (for example, "3 bullet points for recruiter outreach").
  • If you need a non-English response every time, set it in Settings → Writing Style & Language.

Missing context in answers​

  • Update profile data and relevant project details used by Ghostwriter context.
  • Regenerate after updating job notes/description.

I need to reuse a reply outside JobOps​

  • Use the Copy button shown on each completed Ghostwriter response.
  • If the button changes to Copied, the full reply is already on your clipboard.