Admin prompt manager - Alliance IGAD Innovation Hub

The Admin Prompt Manager provides a centralized system for managing AI prompts that power the Alliance IGAD Innovation Hub’s proposal generation workflow.

Overview

Prompts are the instructions given to AWS Bedrock (Claude) for each step of the proposal workflow. The prompt manager allows administrators to:

Create and edit system and user prompts
Organize prompts by section, sub-section, and category
Publish prompts to make them active
Track version history and changes
Preview and test prompts before deployment

Prompt structure

Each prompt consists of two components:

System prompt

The system prompt sets the AI’s role, expertise, and behavior:

You are an expert proposal writer specializing in international development projects.
You have deep knowledge of RFP requirements, evaluation criteria, and donor expectations.

Your task is to analyze the provided RFP document and extract key information
that will guide proposal development.

User prompt template

The user prompt template contains the specific instructions and context, with placeholders for dynamic data:

Analyze the following RFP document:

{{RFP_TEXT}}

Extract the following information:
1. **Requirements**: Technical, organizational, and eligibility requirements
2. **Evaluation Criteria**: Scoring framework and weighting
3. **Deliverables**: Expected outputs and milestones
4. **Submission Guidelines**: Format, page limits, and deadlines

Provide your analysis in JSON format.

Prompt metadata

Each prompt includes organizational metadata:

Field	Description	Example
`section`	Main feature area	`proposal_writer`, `newsletter_generator`
`sub_section`	Workflow step	`step-1`, `step-2`, `rfp-analysis`
`category`	Specific task	`RFP Analysis`, `Concept Generation`
`categories`	List of applicable categories	`["RFP Analysis", "Requirements Extraction"]`
`is_active`	Whether this prompt is currently used	`true` / `false`
`version`	Version number	`1`, `2`, `3`

Prompt organization

Sections and sub-sections

Prompts are organized hierarchically:

proposal_writer/
├── step-1/
│   ├── RFP Analysis
│   ├── Concept Analysis
│   └── Requirements Extraction
├── step-2/
│   ├── Concept Generation
│   └── Section Elaboration
├── step-3/
│   ├── Structure Generation
│   └── Workplan Creation
└── step-4/
    ├── Draft Feedback
    └── Quality Analysis

Active prompt selection

When the system needs a prompt, it queries DynamoDB:

# From source code: backend/app/tools/proposal_writer/rfp_analysis/service.py
filter_expr = (
    Attr("is_active").eq(True)
    & Attr("section").eq("proposal_writer")
    & Attr("sub_section").eq("step-1")
    & Attr("categories").contains("RFP Analysis")
)

response = table.scan(FilterExpression=filter_expr)
prompt = response["Items"][0]  # Returns the active prompt

Only ONE prompt per section/sub-section/category combination can be active at a time. Publishing a new prompt automatically deactivates the previous one.

Managing prompts

Create a new prompt

curl -X POST https://api.igad-hub.com/api/admin/prompts \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "RFP Analysis - Enhanced Requirements Extraction",
    "system_prompt": "You are an expert proposal writer...",
    "user_prompt_template": "Analyze the following RFP...{{RFP_TEXT}}...",
    "section": "proposal_writer",
    "sub_section": "step-1",
    "categories": ["RFP Analysis", "Requirements Extraction"],
    "is_active": false,
    "comments": "Added more detailed requirement extraction"
  }'

Response:

{
  "prompt_id": "prompt#uuid-here",
  "version": 1,
  "created_at": "2024-01-15T10:00:00Z",
  "created_by": "admin@igad-hub.com"
}

Update an existing prompt

curl -X PUT https://api.igad-hub.com/api/admin/prompts/{prompt_id} \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "user_prompt_template": "Updated template with {{NEW_VARIABLE}}...",
    "comments": "Added support for new variable"
  }'

Updates:

Increment version number
Update updated_at timestamp
Preserve change history in comments

Publish a prompt

To make a prompt active:

curl -X PUT https://api.igad-hub.com/api/admin/prompts/{prompt_id} \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"is_active": true}'

The system automatically:

Sets this prompt’s is_active to true
Finds any other prompts with matching section/sub_section/category
Sets their is_active to false

Publishing a prompt immediately affects production! Test prompts thoroughly before activating.

List all prompts

curl -X GET https://api.igad-hub.com/api/admin/prompts \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN"

Optional query parameters:

section: Filter by section (e.g., proposal_writer)
sub_section: Filter by sub-section (e.g., step-1)
is_active: Filter by active status (true / false)

Delete a prompt

curl -X DELETE https://api.igad-hub.com/api/admin/prompts/{prompt_id} \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN"

Cannot delete an active prompt. Deactivate it first by publishing a different prompt.

Placeholder formats

Supported formats

The system supports TWO placeholder formats:

Double curly braces: {{VARIABLE_NAME}}
Curly + square brackets: {[VARIABLE_NAME]}

Both formats are replaced during prompt execution:

# From source code
for key, value in context.items():
    # Format 1: {{KEY}}
    template = template.replace("{{" + key + "}}", str(value))
    # Format 2: {[KEY]}
    template = template.replace("{[" + key + "]}", str(value))

Common placeholders

Placeholder	Description	Example
`{{RFP_TEXT}}`	Full RFP document text	RFP content
`{{CONCEPT_TEXT}}`	User’s concept document	Concept content
`{{RFP_ANALYSIS}}`	Structured RFP analysis result	JSON object
`{{CONCEPT_ANALYSIS}}`	Concept analysis result	JSON object
`{{SELECTED_SECTIONS}}`	User-selected sections for elaboration	Array
`{{USER_COMMENTS}}`	User feedback on sections	Object
`{{PROPOSAL_STRUCTURE}}`	Generated proposal structure	JSON object
`{{DRAFT_CONTENT}}`	User’s draft proposal text	Text content

Version control

Version history

Every prompt update increments the version:

{
  "prompt_id": "prompt#abc123",
  "version": 3,
  "created_at": "2024-01-10T10:00:00Z",
  "updated_at": "2024-01-15T14:30:00Z",
  "change_history": [
    {
      "version": 1,
      "timestamp": "2024-01-10T10:00:00Z",
      "user": "admin@igad-hub.com",
      "comment": "Initial prompt"
    },
    {
      "version": 2,
      "timestamp": "2024-01-12T11:00:00Z",
      "user": "admin@igad-hub.com",
      "comment": "Improved requirement extraction"
    },
    {
      "version": 3,
      "timestamp": "2024-01-15T14:30:00Z",
      "user": "editor@igad-hub.com",
      "comment": "Added budget framework extraction"
    }
  ]
}

Rollback

To rollback to a previous version:

Copy the old prompt’s content
Create a new prompt with the old content
Publish the new prompt to make it active

Testing prompts

Preview endpoint

curl -X POST https://api.igad-hub.com/api/admin/prompts/preview \
  -H "Authorization: Bearer YOUR_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt_id": "prompt#abc123",
    "context": {
      "RFP_TEXT": "Sample RFP content for testing...",
      "CONCEPT_TEXT": "Sample concept..."
    }
  }'

Returns the compiled prompt with placeholders replaced:

{
  "system_prompt": "You are an expert proposal writer...",
  "user_prompt": "Analyze the following RFP:\n\nSample RFP content for testing...\n\nExtract..."
}

Test with Bedrock

For thorough testing, administrators can:

Create a test proposal
Upload test documents
Temporarily publish the new prompt
Trigger the workflow step
Review the AI output
Publish the production prompt if satisfied

Best practices

Prompt design

✅ Be specific: Clearly define the task and expected output format
✅ Use examples: Include sample outputs in the prompt
✅ Set constraints: Specify word limits, formatting requirements
✅ Request JSON: For structured data, always request JSON output
✅ Include context: Provide relevant background and objectives

Version management

✅ Test before publishing: Always test in a non-production environment
✅ Document changes: Use the comments field to explain modifications
✅ Incremental updates: Make small, focused changes to isolate issues
✅ Monitor impact: Track proposal quality metrics after prompt changes

Security

✅ Restrict access: Only grant admin privileges to trusted users
✅ Audit changes: Review change_history regularly
✅ Backup prompts: Export active prompts periodically

Authentication

All admin prompt endpoints require:

JWT Bearer token with admin privileges
is_admin: true claim in the token

Users without admin privileges receive:

{
  "detail": "Admin access required",
  "status_code": 403
}

Next steps

Admin API reference

Complete API documentation for prompt management

Proposal workflow

See how prompts are used in the workflow

Bedrock integration

Learn about AWS Bedrock and Claude integration

User management

Manage admin users and permissions

​Overview

​Prompt structure

​System prompt

​User prompt template

​Prompt metadata

​Prompt organization

​Sections and sub-sections

​Active prompt selection

​Managing prompts

​Create a new prompt

​Update an existing prompt

​Publish a prompt

​List all prompts

​Delete a prompt

​Placeholder formats

​Supported formats

​Common placeholders

​Version control

​Version history

​Rollback

​Testing prompts

​Preview endpoint

​Test with Bedrock

​Best practices

​Prompt design

​Version management

​Security

​Authentication

​Next steps

Admin API reference

Proposal workflow

Bedrock integration

User management

Overview

Prompt structure

System prompt

User prompt template

Prompt metadata

Prompt organization

Sections and sub-sections

Active prompt selection

Managing prompts

Create a new prompt

Update an existing prompt

Publish a prompt

List all prompts

Delete a prompt

Placeholder formats

Supported formats

Common placeholders

Version control

Version history

Rollback

Testing prompts

Preview endpoint

Test with Bedrock

Best practices

Prompt design

Version management

Security

Authentication

Next steps