How Prompt Templates Work

Prospector generates proposal drafts by sending a prompt to an AI model. That prompt is assembled from a template you write plus context data from the lead (job description, client history, your profile, score breakdown). Your template defines the structure, tone, and strategy of the proposal. The AI fills in the specifics.

Templates are stored as files in ~/.ghoststack/prospector/templates/. Each template is a Markdown file with YAML frontmatter for metadata and a body that contains the prompt text with variable placeholders.

Template Structure

---
name: 'default'
match_skills: []
match_job_type: ['fixed', 'hourly']
priority: 0
---

Write a proposal for the following freelance job posting.

**Job title**: {{job.title}}
**Job description**: {{job.description}}
**Required skills**: {{job.skills | join(", ")}}
**Budget**: {{job.budget.amount}} {{job.budget.currency}} ({{job.budget.type}})
**Client history**: {{client.total_spend}} spent, {{client.rating}} avg rating, {{client.hires}} past hires

**My relevant skills**: {{profile.skills | intersect(job.skills) | join(", ")}}
**My experience summary**: {{profile.experience}}

Write a concise, professional proposal (150-250 words) that:

1. Opens with a specific observation about the project, not a generic greeting
2. Connects my skills directly to their requirements
3. Mentions a relevant past project if one exists in my portfolio
4. Proposes a clear next step
5. Ends naturally without a formulaic closing

Tone: direct and confident, not salesy. Do not use phrases like "I am excited to apply" or "I believe I am the perfect fit."

Variable Reference

Templates use double-brace syntax for variable interpolation. Available variables:

Variable	Type	Description
job.title	string	Job posting title
job.description	string	Full job description text
job.skills	string[]	Required skills listed on the job
job.budget.amount	number	Budget amount
job.budget.type	string	”fixed” or “hourly”
job.budget.currency	string	Currency code (e.g., “USD”)
job.duration	string	Expected project duration
client.total_spend	number	Client’s total platform spend
client.rating	number	Client’s average rating
client.hires	number	Number of past hires
client.previous_contracts	number	Your past contracts with this client
profile.skills	string[]	Your configured skills
profile.experience	string	Your experience summary
profile.portfolio	object[]	Your portfolio items
score.composite	number	Lead’s overall score
score.signals.*	number	Individual signal scores

Filters

Variables support filters using the pipe (|) syntax:

• join(separator) - joins arrays into a string
• intersect(other_array) - returns common elements between two arrays
• truncate(length) - truncates text to a maximum character count
• lowercase / uppercase - case conversion
• default(value) - fallback if the variable is null or empty

Multiple Templates

You can create specialized templates for different job types. Prospector selects the best-matching template based on match_skills and match_job_type in the frontmatter. Templates with higher priority values are preferred when multiple templates match.

~/.ghoststack/prospector/templates/
  default.md          # priority: 0, matches everything
  react-project.md    # priority: 10, match_skills: ["react", "next.js"]
  api-backend.md      # priority: 10, match_skills: ["node.js", "postgresql"]
  quick-fix.md        # priority: 5, match_job_type: ["fixed"], budget < $500

When a lead’s required skills overlap with a template’s match_skills, that template wins over the default. If multiple specialized templates match, the one with the highest priority is selected.

Writing Better Prompts

The quality of your proposals depends heavily on how well you write your prompt templates. A few guidelines based on what produces good results:

Be specific about structure. Tell the AI exactly how many words you want, what sections to include, and what order. Vague instructions like “write a good proposal” produce generic output.

Specify what to avoid. AI models default to enthusiastic, salesy language unless you tell them not to. Explicitly list phrases and patterns you do not want.

Use score data strategically. Your template has access to the lead’s score breakdown. You can instruct the AI to emphasize your strongest match areas: “If skill_match is above 80, lead with technical expertise. If client_history shows repeat client, reference our previous work.”

Test with real leads. After writing a template, generate drafts for 5-10 recent leads and review them. Look for patterns: if every draft opens the same way, your template is too rigid. If drafts miss important project details, your context assembly might need adjustment.

# Generate a draft without submitting it
ghoststack prospector proposals draft <lead-id> --template react-project --dry-run

Template Versioning

Prospector tracks which template version generated each proposal. When you edit a template, the previous version is preserved so you can compare proposal quality before and after changes. Use the proposals history command to review how template changes affected your win rate over time.