conversational-ai-assistant

What this skill does

# Conversational AI Assistant

This skill gives the OpenClaw Greek Accounting system a conversational interface. An accounting assistant can ask questions in plain English, request information about any client, understand what needs doing today, and trigger actions — without needing to know CLI commands, file paths, or Greek regulatory terminology. The assistant handles the translation in both directions: English questions into skill commands, and skill outputs back into clear English answers.


## Setup

```bash
export OPENCLAW_DATA_DIR="/data"
which jq || sudo apt install jq
```

No external credentials required. Provides a natural language query interface that routes questions to other installed skills. Requires the OpenClaw LLM runtime.


## Core Philosophy

- **English In, English Out**: Every interaction is in English. Greek data — names, addresses, regulatory terms, AFM numbers — is presented in English context without requiring the assistant to understand Greek
- **Read First, Act Second**: The vast majority of interactions are queries. The assistant surfaces information freely. Actions that change data require the same human confirmation gates as the rest of the system
- **Honest About Uncertainty**: When data is incomplete, when a calculation has low confidence, or when a question requires professional judgement, the assistant says so clearly rather than guessing
- **Skill Orchestration, Not Duplication**: The assistant does not reimplement any skill logic. It calls the appropriate skills, collects their outputs, and presents them coherently. It is a translation layer, not a processing layer
- **Context Awareness**: Within a conversation session, the assistant remembers what has been discussed. If an assistant asks about a client and then asks a follow-up question, the assistant resolves the reference without requiring the AFM to be repeated
- **Professional Tone**: Responses are clear, concise, and professional — appropriate for an accounting firm environment. No unnecessary hedging, no excessive caveats, no waffle
- **Scope Awareness**: The assistant knows what it can and cannot do. It does not attempt to answer questions outside its data — it directs the user to the right place instead

---

## OpenClaw Commands

### Starting a Conversation
```bash
# Start an interactive session
openclaw chat start --user "maria.g"
openclaw chat start --user "maria.g" --client-context EL123456789   # pre-load a client
openclaw chat start --user "maria.g" --mode voice                    # voice-optimised brevity

# Single-turn query (no session)
openclaw chat ask --user "maria.g" --query "What VAT returns are due this week?"
openclaw chat ask --user "maria.g" --query "Show me Alpha Trading's compliance status"

# Session management
openclaw chat history --user "maria.g" --last 10-sessions
openclaw chat resume --user "maria.g" --session-id S20260218-001
openclaw chat end --session-id S20260218-001 --save-summary
```

### Query Mode — Information Retrieval
```bash
# Client queries
openclaw chat ask --query "What documents are we still waiting on from Beta Services?"
openclaw chat ask --query "When did we last file VAT for EL123456789?"
openclaw chat ask --query "Which clients have compliance gaps right now?"
openclaw chat ask --query "How many active clients do we have in the retail sector?"

# Deadline queries
openclaw chat ask --query "What's due this week across all clients?"
openclaw chat ask --query "Is there anything overdue?"
openclaw chat ask --query "When is the next EFKA deadline for Alpha Trading?"
openclaw chat ask --query "What are the VAT deadlines for February?"

# Compliance queries
openclaw chat ask --query "Has Alpha Trading's VAT return been submitted for January?"
openclaw chat ask --query "Are there any clients with overdue filings?"
openclaw chat ask --query "What AADE notifications have come in this week?"

# Financial queries
openclaw chat ask --query "What was Alpha Trading's VAT liability last month?"
openclaw chat ask --query "How many employees does Beta Services have registered with EFKA?"
openclaw chat ask --query "What bank statements are still unprocessed?"
```

### Action Mode — Triggering Workflows
```bash
# Initiate actions (assistant prepares, human confirms)
openclaw chat do --user "maria.g" --action "process this month's documents for Alpha Trading"
openclaw chat do --user "maria.g" --action "prepare VAT return for EL123456789 for January 2026"
openclaw chat do --user "maria.g" --action "run the morning check"
openclaw chat do --user "maria.g" --action "send me a summary of what needs doing today"

# Draft requests (produces output for review, never submits automatically)
openclaw chat draft --user "maria.g" --type "document-request" --client EL123456789 --missing "bank statements October 2025"
openclaw chat draft --user "maria.g" --type "deadline-reminder" --client EL123456789 --deadline "VAT 2026-02-25"
```

### Explanation Mode — Understanding Greek Accounting
```bash
# Explain regulatory concepts in plain English
openclaw chat explain --topic "what is myDATA"
openclaw chat explain --topic "VAT return deadlines Greece"
openclaw chat explain --topic "EFKA contribution categories"
openclaw chat explain --topic "what does E1 form mean"
openclaw chat explain --client EL123456789 --topic "why is their VAT liability high this month"

# Explain system status
openclaw chat explain --topic "what happened during last night's processing"
openclaw chat explain --topic "why is this document flagged for review"
```

---

## Conversation Design

### Intent Recognition

The assistant maps natural language to one of five intent types, then routes accordingly:

```yaml
Intent_Types:

  query_client:
    triggers: ["tell me about", "what's the status of", "show me", "how is", "what does X owe"]
    routes_to: "openclaw clients view + openclaw dashboard client-overview"
    requires_auth: "check-access --action view_client"

  query_deadline:
    triggers: ["what's due", "when is", "any deadlines", "overdue", "upcoming"]
    routes_to: "openclaw deadline check + openclaw dashboard deadlines"
    requires_auth: "read-only, no check needed"

  query_compliance:
    triggers: ["has X been filed", "what's been submitted", "compliance status", "any gaps"]
    routes_to: "openclaw clients compliance-history + openclaw dashboard compliance-check"
    requires_auth: "check-access --action view_compliance"

  trigger_action:
    triggers: ["process", "prepare", "run", "do", "start", "generate"]
    routes_to: "openclaw greek [relevant workflow command]"
    requires_auth: "check-access --action [relevant action]"
    note: "Always shows what will happen before doing it. Government submissions always require explicit confirmation."

  explain_concept:
    triggers: ["what is", "explain", "what does X mean", "why", "how does"]
    routes_to: "internal knowledge base + openclaw aade status-check if regulatory"
    requires_auth: "none"
```

### Context Resolution

Within a session, the assistant tracks references so assistants do not need to repeat themselves:

```yaml
Context_Tracking:
  active_client:
    example_exchange:
      user: "Tell me about Alpha Trading"
      assistant: "[loads EL123456789 context]"
      user: "When did they last file VAT?"
      assistant: "[resolves 'they' to Alpha Trading — no need to repeat AFM]"

  active_period:
    example_exchange:
      user: "What's due in February?"
      assistant: "[loads February 2026 deadline context]"
      user: "Are there any that are already overdue?"
      assistant: "[resolves 'any' to February deadlines already loaded]"

  active_task:
    example_exchange:
      user: "Prepare the VAT return for Alpha Trading for January"
      assistant: "[prepares return, shows summary]"
      user: "Looks good, submit it"
      assistant: "[resolves 'it' to the prepared VAT return — shows confirmation gate]"

  context_reset:
    trigger: "New client or topic introduced explicitly"
    exa