renatocaliari/cali-degustia-client-news-watch

cali-degustia-client-news-watch

Monitor AI and tooling news for updates that matter to past clients. Reads existing client report files (generated by cali-degustia-diagnostico or similar), deduplicates recommended tools across all clients, searches for relevant news, and cross-references findings back to the clients who could benefit.

---

Workflow

Step 1 — Collect client report files

Ask the user where their client report files are stored. Default locations to check:

• ~/.agents/skills/cali-degustia-diagnostico/assets/ — HTML templates (the source of truth if filled reports are there)
• A user-specified directory (e.g., ~/clientes/, ~/reports/, or a cloud-synced folder)

Read all files with .html or .md extensions in the given directory. For .html files, extract the visible text content (strip tags, scripts, and styles) — the semantic structure (client name, recommended tools, pain points, context notes) is usually present in headings, lists, and paragraphs.

If no directory is specified, ask the user to provide one or to paste the reports inline.

---

Step 2 — Extract and deduplicate

For each report file, extract:

| Field | Description | |-------|-------------| | Client identifier | Name, initials, or any unique label found in the report | | Company name | The client's company or organization name, if present in the report | | Recommended tools | Every tool, platform, or practice recommended (e.g., ChatGPT, Make, Notion, a specific automation approach) | | Client context | The client's role, industry, pain points, and goals as described in the report | | Problems addressed | The specific tasks or bottlenecks the recommendations target |

Deduplication rules:

• Merge identical tool references across clients into a single entry.
• For each unique tool, keep a list of which clients received it.
• If the same tool was recommended for different use cases (e.g., ChatGPT for copywriting in one client and for data analysis in another), note both use cases.
• If a tool name has multiple variations (e.g., "ChatGPT" vs "GPT-4" vs "OpenAI"), normalize to the most specific correct name.

Output of this step:

## Deduplicated tool map

### ChatGPT
  - Client A: copywriting briefs, 120 min/week saved
  - Client B: data analysis from CSV exports
  - Client C: drafting client emails

### Make (Integromat)
  - Client A: automated report generation
  - Client D: invoice processing pipeline

### ... (continue for all tools)

---

Step 3 — Research relevant news

For each unique tool and for each unique use-case pattern (e.g., "copywriting automation", "report automation for SMBs"), search for meaningful updates.

Sources to use (in order):

1. last30days skill (try first) — Best for detecting community sentiment, Reddit discussions, X posts, Exa Web, YouTube, and Hacker News threads about a tool or category. Run one query per tool and one per use-case pattern. This is the primary source for social/community signals.
2. researcher-web tool — For deeper, authoritative sources: official blog posts, changelogs, pricing updates, deprecation notices, acquisition news, competitor launches. Use this to follow up on strong signals found by last30days.
3. researcher-docs tool — If a tool has updated its API or documentation in a way that affects how it was recommended (e.g., a new endpoint, a sunset feature).

For every finding, capture the source URL. Without a verifiable URL, omit the source_url field in the JSON output. This ensures every news item is traceable back to its origin.

What qualifies as "relevant":

| Signal Type | Priority | Example | |-------------|----------|---------| | Tool deprecation / sunset | 🔴 Critical | "Make is shutting down free tier" | | New feature that replaces a manual step | 🟢 High | "ChatGPT now reads uploaded PDFs natively" | | Alternative tool that solves the same pain better | 🟢 High | "New open-source alternative to Make" | | Pricing change | 🟡 Medium | "Notion AI dropped to $8/mo" | | Community workaround / tip | 🔵 Low | "Reddit thread on how to use Make with Gmail better" | | General AI news not tied to any recommended tool | ⚪ Ignore | "Google launched a new AI model" (unless it directly replaces a recommended tool) |

Filter aggressively: Do not include news that is generic, hype-driven, or has no clear implication for the client's specific use case.

---

Step 4 — Cross-reference and prioritize

Map each news item back to the specific clients who would benefit.

⚠️ Privacy rule: The deduplicated tool map links tools to client identifiers. This map is for internal use only. Never expose which client uses which tool outside this brief. When presenting findings, anonymize client identifiers (e.g., "Client A", "Client B") or omit the map entirely if sharing externally.

## News: Make launched pre-built templates for invoice processing

Clients affected: A, D
Why it matters: Client A uses Make for report generation — templates could cut setup time.
                 Client D uses Make for invoice processing — this directly addresses their workflow.
Suggested action: Let both clients know with a short sentence and a link.

Priority: High

Prioritization matrix:

| Priority | Criteria | Action | |----------|----------|--------| | High | Directly affects a recommended tool OR opens a clearly better alternative | Prepare a message to the client | | Medium | Relevant but not urgent — could be mentioned in next check-in | Add to a digest list | | Low | Interesting but not actionable for any client | Skip or file for reference |

---

Step 5 — Produce the brief

Output two deliverables:

A. Human-readable brief (same format as before)

# Client News Watch — [Date]

## 🔴 Critical Alerts
(Any sunset/deprecation that requires immediate action)

## 🟢 Opportunities for Specific Clients
(New features or alternatives mapped to clients)

## 🟡 Digest (Medium-signal updates)
(News worth knowing but not urgent)

## 📊 Coverage Summary
- Total clients scanned: N
- Unique tools tracked: N
- Total news items found: N
- Items matched to clients: N
- Clients with at least one relevant update: N

## Raw Tool Map (for reference)
(The deduplicated tool list from Step 2)

B. Per-client structured output (for integration tooling)

After the brief, output a JSON array with one entry per affected client. This format is designed to be consumed by CRMs (like NocoBase), automation tools (Make, n8n), or any script that processes notifications.

{
  "brief_date": "2026-07-21",
  "generated_by": "cali-degustia-client-news-watch",
  "clients": [
    {
      "client_id": "cliente-a",
      "name": "Nome do Cliente",
      "company": "Nome da Empresa",
      "email": "[email protected]",
      "priority": "high",
      "relevant_news": [
        {
          "tool": "Make (Integromat)",
          "news": "Make launched pre-built templates for invoice processing",
          "relevance": "This directly addresses Client A's invoice processing workflow. Templates could reduce setup time from hours to minutes.",
          "source_url": "https://example.com/news",
          "suggested_message": "Oi [nome], tudo bem? Sobre a automação de notas fiscais que a gente montou: a Make lançou uns templates prontos pra isso. Deve cortar seu tempo de setup. Quer que eu dê uma olhada?"
        }
      ]
    },
    {
      "client_id": "cliente-b",
      "name": "Nome do Outro Cliente",
      "company": "Outra Empresa",
      "email": "[email protected]",
      "priority": "medium",
      "relevant_news": []
    }
  ]
}

Rules for the JSON output:

• Include ALL clients from Step 2, not only those with news. Clients with no relevant news get an empty relevant_news array and priority: "none".
• The client_id should be a stable slug (derived from the client identifier in the report) so the same ID is reused across runs.
• The email field is optional — include it only if available in a structured source (e.g., a clientes.csv found alongside the reports). If unavailable, omit the field entirely rather than leaving it blank.
• Each suggested_message should be a short (2-3 sentences), personalized message in the client's language (Portuguese for Brazilian clients) that the user could review and send directly.
• If a client has multiple news items, sort by priority (critical > high > medium > low > none).

---

Config

If the user prefers a default report directory, they can set it in references/config.yaml:

client_reports_dir: ~/clientes/reports/  # default path to scan

Otherwise, the skill prompts for the directory each run.

---

Known limitations

• Relies on the quality and completeness of existing client reports. If a tool was recommended verbally or in a channel not captured in a report file, it won't be tracked.
• The last30days skill requires API keys for some sources (X, TikTok, Instagram). The brief will note what sources were actually queried.
• Deduplication is text-based; if two tools are functionally identical but named differently (e.g., "Zapier" vs "Zapier AI"), they will be separate entries. The agent should use judgment to flag possible duplicates.