Process Analyst Agent

Analyzes a business process, finds gaps, clarifies with the human, generates a complete specification for building an agent.

When to use

• Before building a new agent
• "analyze process X"
• "what is needed to automate Y"

Dependencies

• Skills: dispatcher, memory
• Data: CRM schema, PM data, existing skills, existing tools

Input

Process name or number from the Process Map:

How to execute

Step 1: Gather context

For the specified process, read:

1. Existing skill (if any) — from $SKILLS_PATH/skills/
2. Existing tool (if any) — scripts, API clients
3. Data — which CSV/files the process reads or writes
4. Schema — $CRM_PATH/schema.yaml
5. Adjacent processes — what runs before/after this process
6. Email Pipeline as reference — $GOOGLE_TOOLS_PATH/ (the only fully automated agent)

Step 2: Analysis by checklist

For each process, fill in:

## Process Analysis: [Name]

### 1. TRIGGER (what starts the process)
- [ ] Trigger defined (schedule / event / manual)
- [ ] Frequency defined
- [ ] Launch conditions are clear

### 2. INPUT (input data)
- [ ] Data sources defined
- [ ] Data format is clear
- [ ] Data access is available (API keys, credentials)
- [ ] Data volume is estimated

### 3. PROCESSING (processing logic)
- [ ] Business rules described
- [ ] Edge cases defined
- [ ] Dependencies on other processes defined
- [ ] AI component needed? Which model?

### 4. OUTPUT (result)
- [ ] What is created / modified
- [ ] Where it is written (CSV, file, API)
- [ ] Who is the consumer of the result
- [ ] Output format is defined

### 5. ERROR HANDLING
- [ ] What to do on API error
- [ ] What to do with invalid data
- [ ] Retry logic
- [ ] Alerting (where to report an error)

### 6. HUMAN-IN-THE-LOOP
- [ ] Which decisions require human approval
- [ ] Approval format (Telegram notification? CLI prompt?)
- [ ] What to do if human did not respond

### 7. INTEGRATION
- [ ] Which other agents depend on this one
- [ ] Which agents does this one depend on
- [ ] Shared state (which files are shared)
- [ ] Are race conditions possible?

### 8. GAPS (what is missing)
- [ ] List of questions for the owner
- [ ] Missing tools
- [ ] Missing data
- [ ] Missing credentials

Step 3: Dialogue with the human

For each unfilled item -- formulate a clear question:

Question format:

[SECTION] [QUESTION]
Context: what is already known
Options: if there are obvious choices
Default: if there is a recommendation

Rules:

• No more than 5 questions at a time
• From most important to least important
• Suggest a default where possible
• If something is obvious from context -- don't ask, just record it

Step 4: Generate Spec

After all clarifications -- create a file:

$AGENTS_PATH/specs/[process-name].spec.md

Spec structure:

# Agent Spec: [Name]

## Meta
- Process ID: #
- Priority: high/medium/low
- Complexity: simple/medium/complex
- Estimated components: N files

## Overview
One paragraph on what the agent does.

## Trigger
- Type: schedule / event / manual
- Schedule: cron expression (if schedule)
- Event: what triggers it (if event)

## Pipeline

[Input] → [Step 1] → [Step 2] → ... → [Output]

## Steps

### Step 1: [Name]
- Input: what it receives
- Logic: what it does
- Output: what it produces
- Error: what on error
- Model: if AI (sonnet/haiku/opus)

### Step 2: [Name]
...

## Data Flow

| Reads | Writes |
|-------|--------|
| file.csv | file.csv |

## Human Approval Points
- Where approval is needed
- How to notify (Telegram/CLI)
- Timeout: what if no response

## Dependencies
- Other agents: [list]
- Tools: [list]
- Credentials: [list]

## Testing Plan
- Unit: what to test in isolation
- Integration: which agents to test with
- Dry-run: how to run without side effects

## Rollback
- How to rollback if something went wrong

Step 5: Validation

• Re-read the spec
• Check that all gaps are closed
• Check consistency with other specs (if any)
• Save and notify the human

Output

Spec file in $AGENTS_PATH/specs/[name].spec.md

Examples

Example 1: Daily Briefing Analysis

User: analyze process #14 Daily Briefing

Claude: [reads daily-briefing.md, pm_tasks_master.csv, email agent, activities.csv]
Claude: Here is the analysis of the Daily Briefing process:

TRIGGER: schedule, every morning at 08:00
INPUT: pm_tasks_master.csv, activities.csv, email summaries
PROCESSING: AI summary (haiku)

GAPS -- need to clarify:
1. [TRIGGER] What time for the morning briefing? Default: 08:00
2. [OUTPUT] Where to send the result? Options: Telegram / CLI / both
3. [HUMAN] Is approval needed before sending, or automatic?

Claude: creates spec after responses

Related skills

• agent-builder — takes the spec and builds the agent
• agent-tester — tests the built agent
• dispatcher — task routing
• memory — context from previous sessions