Pull Prompt

Pull prompt from Supabase

When this applies

• User says "pull prompt for {client}", "get prompt for {client}", or similar.
• User wants to review, compare, or export a client's current AI assistant prompt and configuration.
• Called by the sync workflow or auto-sync rule to check if prompts need refreshing (see "Freshness check mode" below).

MCP server

• Supabase (user-Supabase): execute_sql with project_id for the Avoca production database.
• Discover project_id via list_projects if unknown. The production project is named "avoca".

Inputs

Workflow

Step 1 — Resolve the client

Match the client to their clients/{Name}-{ID}/ folder. Extract the team ID (numeric suffix). If the user gave a name, glob clients/ for a match. If they gave an ID directly, use it.

Read avoca-client.json from the client folder if it exists — it may have avoca_client_id.

Step 2 — Query system prompts for the team

SELECT sp.id, sp.name, sp.status, sp.template_id, sp.current_version,
       sp.updated_at, sp.variable_overrides,
       spc.company_name
FROM system_prompts sp
JOIN system_prompt_configs spc ON sp.config_id = spc.id
WHERE sp.is_deleted = false
  AND sp.team_id = {team_id}
ORDER BY sp.updated_at DESC;

If multiple prompts exist, list them and ask the user which one to pull (or pull all if they say so). If only one, proceed automatically.

Step 3 — Pull the full data

Run these three queries in parallel for the chosen prompt (by sp.id):

A) Sections (modules):

SELECT
  s->>'name' AS module_name,
  s->>'order' AS module_order,
  s->>'enabled' AS enabled,
  s->>'isCustom' AS is_custom,
  s->>'isLocked' AS is_locked,
  s->>'content' AS content,
  s->>'originalContent' AS original_content
FROM system_prompts, jsonb_array_elements(sections) AS s
WHERE id = {prompt_id}
ORDER BY (s->>'order')::int;

Note: this may return a large result. If it times out or truncates, batch by pulling sections in groups (e.g., WHERE (s->>'order')::int BETWEEN 0 AND 19, then 20–39, then 40+).

B) Config (business variables):

SELECT spc.*
FROM system_prompt_configs spc
JOIN system_prompts sp ON sp.config_id = spc.id
WHERE sp.id = {prompt_id};

C) Template name:

SELECT name, version FROM system_prompt_templates
WHERE id = (SELECT template_id FROM system_prompts WHERE id = {prompt_id});

Step 4 — Build the markdown file

Save to: clients/{Name}-{ID}/prompts/{prompt-name-slug}.md

Slugify the prompt name (lowercase, hyphens, no special chars). If a file already exists, overwrite it (this is a fresh pull).

File structure

# {Prompt Name} — {Company Name}

**Team ID:** {team_id}
**Prompt ID:** {prompt_id}
**Template:** {template_name} (v{version})
**Status:** {status}
**Version:** {current_version}
**Last updated:** {updated_at}
**Pulled:** {today's date and time}

---

## Business config

| Field | Value |
|-------|-------|
| Company name | {company_name} |
| Assistant name | {assistant_name} |
| Location | {company_location} |
| Timezone | {timezone} |
| Trades | {trades_provided.text} |
| Service hours | {service_hours} |
| Live rep hours | {live_representative_hours} |
| Company phone | {company_phone_number} |
| Website | {website} |
| Fees name | {fees_name} |
| Intro line | {ai_intro_line} |
| Closing line | {offer_closing_line} |
| Country | {country} |
| Tech arrival notice | {technician_arrival_duration} |
| Emergency re-fetch | {emergency_re_fetch_mode} |

---

## Policy snapshot

Decode `agent_config_variables` into a human-readable table:

| Policy | Setting |
|--------|---------|
| Booking | {enabled/disabled} — {offer_availabilities.type} |
| Transferring | {enabled/disabled}{, time-based if applicable} |
| Commercial | {action} |
| Cancellation | {action} |
| Rescheduling | {action} |
| ETAs | {action} |
| New customers | {action} |
| Renters | {action} |
| Membership confirmation | {enabled/disabled} |
| Equipment age question | {enabled/disabled} |
| AI disclosure | {enabled/disabled} — {timing} |

---

## Module index

| # | Module | Enabled | Custom | Size |
|---|--------|---------|--------|------|
| 0 | MODULE: VARIABLES | ✅ | — | 4.6 KB |
| 1 | SYSTEM CAPABILITIES | ✅ | — | 1.4 KB |
| ... | ... | ... | ... | ... |

Use ✅ for enabled, ❌ for disabled. Use ✏️ for custom modules.

---

## Modules

For each module, output:

### {order}. {Module Name}

{If disabled: **⚠️ DISABLED**}
{If custom: **✏️ CUSTOM OVERRIDE** — content differs from template}

{content — the rendered module text, as-is}

---

## Services provided

{services_provided — full text}

## Services NOT provided

| Service | Description |
|---------|-------------|
| {service} | {description} |

## FAQ

{faq}

## Membership info

{membership_information}

Step 5 — Custom diff detection

For each module where isCustom is true OR where content differs from originalContent:

• Note it in the module index with ✏️
• In the module section, add a collapsed diff block:

<details>
<summary>Diff from template</summary>

**Original:**
{originalContent}

**Current:**
{content}

</details>

Step 6 — Variable overrides

If system_prompts.variable_overrides is non-null, add a section:

## Variable overrides

| Variable | Override value |
|----------|---------------|
| ... | ... |

Step 7 — Confirm

Tell the user:

• File path saved
• How many modules (total / enabled / disabled / custom)
• Any notable config (e.g., "transfers disabled", "booking type: specific_time")

Pulling multiple prompts

If the user says "pull all prompts" for a client with multiple prompts, loop through each and create separate files. Summarize all in one message.

Comparing prompts across clients

If the user says "compare prompts for {client A} and {client B}":

1. Pull both using the workflow above (if not already saved locally).
2. Produce a diff summary focusing on:
   • Modules that differ in content
   • Modules enabled in one but disabled in the other
   • Config differences (services, hours, policies)
3. Save as clients/{Name-A}-{ID}/prompts/compare-vs-{Name-B}.md or output in chat per user preference.

Step 8 — Update avoca-client.json

After writing the markdown file, update avoca-client.json in the client folder with prompt metadata so future syncs can check freshness.

Read the existing file, merge in the prompts key, and write it back:

{
  "avoca_client_id": "2339",
  "prompts": {
    "457": {
      "name": "John Goudie",
      "version": 7,
      "updated_at": "2026-04-02T02:52:32Z",
      "pulled_at": "2026-04-05T22:00:00Z",
      "file": "prompts/john-goudie.md"
    }
  }
}

Fields per prompt entry:

• name — prompt name from system_prompts.name
• version — current_version from system_prompts
• updated_at — updated_at from system_prompts (ISO 8601)
• pulled_at — current timestamp when the pull happened (ISO 8601)
• file — relative path to the markdown file within the client folder

If the client has multiple prompts, each gets its own key under prompts (keyed by prompt ID).

Freshness check mode

When called from the sync workflow or auto-sync rule (not a direct user request), use this lightweight mode instead of a full pull.

When to use

• The sync workflow reads avoca-client.json and finds a prompts section.
• It calls this skill in "check" mode for that client.

Procedure

1. Read avoca-client.json from the client folder. If no prompts key exists, skip entirely — the user hasn't opted in to prompt tracking for this client.

2. Extract the team ID and the stored updated_at for each prompt.

3. Run a single cheap query:

SELECT id, name, current_version, updated_at
FROM system_prompts
WHERE team_id = {team_id} AND is_deleted = false;

4. Compare each prompt's updated_at from Supabase against the stored updated_at in avoca-client.json.

5. If unchanged: Do nothing. Optionally log "prompt is current" if reporting to the user.

6. If changed: Run the full pull workflow (Steps 2–8 above) for the changed prompt(s). This overwrites the markdown file and updates avoca-client.json with the new version/timestamp.

7. If a new prompt appears (ID not in avoca-client.json): Ignore it — don't auto-pull prompts the user hasn't explicitly pulled before.

Output in check mode

Keep it minimal. Examples:

• (nothing — prompt is current, no output needed during silent sync)
• Prompt "John Goudie" updated (v7 → v8) — re-pulled to prompts/john-goudie.md

Safety

• Read-only: never write to Supabase.
• Large prompts may exceed query result limits — batch section pulls if needed.
• Don't dump the full final_prompt text unless the user specifically asks for it; the module breakdown is the point.

Output

Keep chat output concise. Example:

Pulled "John Goudie" prompt to clients/John Goudie-2339/prompts/john-goudie.md 54 modules (53 enabled, 1 disabled) · Template: Modular v3.0.87 · Last updated 4/2/2026