lightricks/ltx-leadership-weekly-report

LTX Leadership Weekly Report

Overview

Weekly snapshot of LTX across all three product surfaces, posted every Monday morning to a leadership Slack channel:

• Open Source — HuggingFace download/likes momentum on LTX 2 / 2.3 / 2.3-fp8, plus community feedback themes
• API (LTXV API) — revenue split (Resellers / Contracts / Self-serve), new endpoints, top gen types, self-serve funnel, PQL/PEL counts, top inbound industries
• Studio Enterprise — active vs total contract orgs, new contracts/pilots, total + LTX-model token consumption with WoW%, share-of-video-tokens, share-of-gens

The post leads with an LLM-generated 2-3 sentence TL;DR that follows strict materiality and "no filler" rules — see n8n/tldr_system_prompt.md.

Architecture

Schedule Trigger (Mon 09:00 UTC)
    ↓
Run BigQuery       ← queries/weekly_report_combined.sql  (10 rows: section, data_json)
    ↓ (10 items)
Parse Sections     ← n8n/parse_sections.js               (1 item: { sections: {...} })
    ↓
Render Body        ← n8n/render_body.js                  (1 item: { header, body, ... })
    ↓
Generate TLDR      ← n8n/tldr_system_prompt.md           (Anthropic / OpenAI node)
    ↓
Combine            ← n8n/combine.js                      (1 item: { final_message, ... })
    ↓
Slack Message      (HTTP Request → incoming webhook)

Critical detail: the BigQuery script returns a single result set of 10 rows, each carrying one report block as a JSON-encoded string in data_json. n8n's native BigQuery node (jobs.query) only surfaces the last statement's output, so the original multi-statement script in dwh-data-model-transforms was rewritten into one SELECT ... UNION ALL ... here. KPI logic is unchanged; only the wrapping is different.

Repository layout

ltx-leadership-weekly-report/
├── SKILL.md                          ← this file
├── queries/
│   └── weekly_report_combined.sql    ← single-output SQL for n8n's BQ node
├── n8n/
│   ├── parse_sections.js             ← Code node #1 (parse 10 JSON rows → object)
│   ├── render_body.js                ← Code node #2 (render Slack body)
│   ├── tldr_system_prompt.md         ← System prompt for the LLM node
│   └── combine.js                    ← Code node #3 (TL;DR + body → Slack post)
└── references/
    └── report-template.md            ← visual reference for the rendered post

Instructions — first-time setup

1. Create credentials in n8n

• BigQuery service account — JSON key for a service account with bigquery.jobs.create and read on ltx-dwh-prod-processed.*. Set the credential's default project to ltx-dwh-explore (where the query bills against, no required dataset).
• Anthropic API key — for the TL;DR node.
• Slack incoming webhook — one URL per destination channel. Create the webhook at api.slack.com/apps → your bot → Incoming Webhooks. The webhook is bound to a specific channel; switching channels means a different webhook URL.

2. Build the workflow

Create the 7 nodes in this exact order, matching the names below (the names matter — combine.js references $('Render Body').first().json by name):

| # | Node name | Type | Configuration | |---|------------------|---------------------------------------------------|---------------| | 1 | Schedule Trigger | n8n Schedule trigger | Mondays, 09:00 UTC | | 2 | Run BigQuery | Google BigQuery → "Execute a SQL query" | Paste queries/weekly_report_combined.sql. Use Standard SQL. | | 3 | Parse Sections | Code (JavaScript) | Paste n8n/parse_sections.js. | | 4 | Render Body | Code (JavaScript) | Paste n8n/render_body.js. | | 5 | Generate TLDR | Anthropic → "Message a model" | Model claude-sonnet-4-5, temperature 0.2, max tokens 300. System message: paste content between the BEGIN/END markers in n8n/tldr_system_prompt.md. User message: ={{ $json.body }}. | | 6 | Combine | Code (JavaScript) | Paste n8n/combine.js. | | 7 | Slack Message | HTTP Request | POST to your Slack webhook URL. Body type JSON: { "text": "{{ $json.final_message }}" }. |

3. Test

Execute manually (Monday morning is fine, or any time — the SQL is parameterized off current_date). Inspect each node:

• Run BigQuery: 10 items, each with { section, data_json }.
• Parse Sections: 1 item with sections keyed by A_oss_summary … J_studio_summary.
• Render Body: header, body, week_start, week_end. Body should look like references/report-template.md.
• Generate TLDR: 2-3 sentences, numbers in *bold*, no em-dashes, no banned filler.
• Combine: final_message ready to post.
• Slack Message: posts to the channel.

If you switch LLM provider, the TL;DR field path may change — combine.js already tries the common Anthropic / OpenAI / LangChain shapes, but verify in the run log.

Knobs

| Knob | Where | Current value | Why | |---|---|---|---| | Schedule | Schedule Trigger node | Mon 09:00 UTC | Reflects ISO-week ending the previous Sunday | | Materiality threshold (general) | tldr_system_prompt.md | \|WoW%\| ≥ 25% | Below this is exec-noise | | Materiality threshold (totals) | tldr_system_prompt.md | \|WoW%\| ≥ 15% | Total revenue / total tokens move slowly | | PEL volatility | tldr_system_prompt.md | Always skip | Marketing PEL swings are noisy | | Section coverage | tldr_system_prompt.md | TL;DR may omit quiet sections | Don't pad with "OSS was steady" | | Bold rule (body) | render_body.js | Bold headline numbers only | Not WoW%, not "was X%" context | | Direction arrow | render_body.js | ↑ ↓ 0% | Visual scanning, no +/- chars | | Em-dash policy | both | Banned everywhere | Avoids "AI slop" appearance |

Constraints

DO NOT

• Run the original multi-statement weekly_report_combined.sql from dwh-data-model-transforms through n8n's BQ node — only the last block (J_studio_summary) will come out. Use queries/weekly_report_combined.sql here.
• Rename the n8n nodes "Render Body" or "Generate TLDR" without updating the references in combine.js.
• Add bolding to WoW percentages, "was X%" context, or non-numeric words. Body must follow the locked layout in references/report-template.md.
• Introduce em-dashes (—) or en-dashes (–) anywhere in the body or TL;DR. Use commas / parentheses / colons.
• Trust a single-week PEL move > ±50% — confirm with the body's other numbers before letting the LLM headline it.
• Hardcode the Slack webhook URL inside combine.js. Pass it via the Slack Message node only.

DO

• Update references/report-template.md whenever render_body.js layout changes — the template is the visual contract.
• Bump the Reference TL;DR block in tldr_system_prompt.md after every prompt change so we have a regression anchor.
• Keep queries/weekly_report_combined.sql in lockstep with dwh-data-model-transforms/queries/weekly_report_combined.sql. KPI changes there must be ported here.

References

• Source SQL (multi-statement, dev/EDA version): dwh-data-model-transforms/queries/weekly_report_combined.sql
• Spec doc with KPI definitions: dwh-data-model-transforms/queries/weekly_report_combined.md
• BigQuery cost: ~537 MB scanned per run (dry-run upper bound).