Chinh (lelouvincx) lelouvincx

Similar to the idea ((69abe62c-5ba2-468f-ba5c-c0e9274cc6d0)), we have github repo claude-mem they it automatically captures everything Claude does

Tooling investment is paying off

• This week saw 4 agent skills set up (Notion, Holistics, BigQuery, weekly report automation).
• The weekly report automation alone will save 1–2 hours per week.
• The risk is spending more time building tools than using them.
• => Consider a cooling-off period before starting new skill builds.

Logseq Automation behavior draft up

Tasks

Done

• ((69a2a1cc-fc1f-4802-ba56-48183fa9e9a5))

• ((69a81acd-1ec9-457d-ab25-abb87c57d581))

• ((69a7f8db-8d65-4c0b-9148-a11223fc395a))

• DONE Prepare a high level draft of our deliverables that we completed

• Sent to Vincent in slack.

Tasks

Done

• DONE Data tracking - AI Enablement

DONE Init a document for me and Phuong Nguyen to discuss.

• Asked chi Phuong, she wants to:
  • Track how many prospects/customers enabled AI?
  • Know how long for a tenant since created to enabling AI?
• My follow-up questions to make things clearer:
  • Suppose we know the it’s long enough for a tenant to enable AI, then what’s next?
    • Data policy?
      • Action 1: Make our data policy more explicit.
      • Action 2: Review policy.
    • Data modeling is not well-prepared?
    • Pricing
      • Free or not? Make it more explicit.
  • These survey-like questions are currently not being tracked under any form.
  • Although hard to convince it’s good enough, but it’s quick to build from data side, there’s no reason to reject this. => Set at #P3
• Asked Giap (engineer):
  • Toggle AI: public.tenants.settings->enablement.
  • Banner: public.users.settings->personal->ai->hide_banner.
  • These data are already stored in bigquery via Data Import -> dbt snapshot.
• Expected output: an analytical-ready dataset for answering written questions.
• I found a dashboard already tracking these information: AI Settings.
• Ask: If I add the hide_banner tracking, does this dashboard meet your requirements?
• Phuong answered we can suspend this if the ROI is low. Will get back if later trigger.

DONE Support Pass The Keys sharing object dependency data

Engineering team’s AI adaptation approach:

• Each squad chooses a task, do it.
• Revert, use AI to do it.
• Compare the results.
• Note learnings.
• Adapt current process.

DEFINE Correct and Accurate?

• Correct means free from error, conforming to a standard or following rules.
• Accurate means high degree of precision, fidelity to fact, or truth.

Tasks

Done

• ((69a2a1cc-fc1f-4802-ba56-48183fa9e9a5))
• DONE Continue commenting the project DE-201 from ((6984ba77-ed0a-4312-a25c-c83021e0606a))

Only focus on things that need to do during the PR preview to facilitate the reviewing process.

[[Down The Rabbit Hole]]

• Impressions from Vincent’s demo call
  • I’m writing impressions after watching a call to practice using Benjamin Franklin’s self-taught writing method.
  • Call Link
  • Slack
  • He keep talking almost the call.
    • I don’t know whether it’s good for now.
    • But I remembered someone said if you want to become a streamer, you have to be able to talk all the time, even no one is listening.
    • So that may be a thing I should practice to improve myself.
  • If I’m an audience, hitting the spot when answering my questions matter.
    • I will most remember and be impressed by that.
  • He’s not too technical about coding but he knows the product features very well, even new, small features.
    • For example, generate AI field descriptions.
    • Leo (prospect) said: “Nothing above what stood out to me as we were going through, but yeah, there’s a lot here. I’m just kind of annoyed that I haven’t played with anything in much detail beyond Looker because it’s so 2019.”
      • Meant, when try the product there will be hidden features that they don’t know.
      • As soon as attending the calls (demo/onboarding).
    • Impressive.
  • Stand in customers’ shoe and be honest.
    • Don’t overestimate our features. They are great to us but not always great to them.
    • For example, AI pricing. Vincent mentioned Holistics allows customers to use their own key instead of locking them into the BI.
    • As a BI tool they do not train LLM models. So a BI tool and any LLM should be free to customers to choose.

Tasks

Done

• ((69a2a1cc-fc1f-4802-ba56-48183fa9e9a5))
• DONE Continue commenting the project DE-201 from ((6984ba77-ed0a-4312-a25c-c83021e0606a))

{{embed ((69a2a1cc-94ff-4e02-972e-3b4d720b63ce))}}

https://agents.craft.do/mermaid

• Beautified mermaid, as a JS package (I guess) => then can be used in website & documentation.

https://ampcode.com/notes/200k-tokens-is-plenty

• He said it right, “200k tokens? Plenty, if you make use of threads”.
• Long threads:
  • Worse inference output.
  • Expensive.
• Simply explained: agents get drunk if you feed them too many tokens.
• He used Amp CLI and utilize the read_thread tool. Each small task is a new thread, then he will have some way to library/manage them all.
• https://media.secondbrain.lelouvincx.com/2026/03/1f4004bc50bad9179b277331f04cce69.png
• “Some people use multiple markdown files with specs and histories to store and transfer context between threads (it’s me tbh) - I’m not one of those people. I tell Amp to run git diff or to inspect previous commits.”

Plan for week 10 (Ampcode)

• P1
  • Add2Cart
    • ((69a2a14f-91e8-49be-be93-da1c506869aa)) (carry-over from friday, push Anurag).
    • ((69a2a1cc-94ff-4e02-972e-3b4d720b63ce))
  • Internal
    • ((69a42ffd-22c7-4b43-8b34-0f663743abcf))
    • ((69a2a1cc-fc1f-4802-ba56-48183fa9e9a5))
• P2
  • Internal
    • ((697889da-2e31-4c43-9676-81a898db1ea6))
  • Presales
    • ((69841d05-88ea-4209-82ca-30b11b26e91c))
  • Duty Support
    • Document MedEnterprises performance use case on metrics sheet.
• P3
  • Personal
    • Learn logseq’s internal query, and potentially build a skill to query tasks faster.

Tasks

Done

• ((69a2a14f-91e8-49be-be93-da1c506869aa)) (push Anurag)
• {{embed ((69a2a1cc-fc1f-4802-ba56-48183fa9e9a5))}}
• DONE Code Review MRR https://github.com/holistics/dbt/pull/812

Chilling afternoon at Home Coffee https://media.secondbrain.lelouvincx.com/2026/03/c747aba549a69c538f38a9486b16186f.png

[[Google Data Portability API]]

IDEA Archive all amp theads into a repository.

• Motivation: These threads are context / information. In worst cases, someday ampcode does not support well at getting historical threads (like chatgpt did), I need to backup all these information.

Tasks

Done

• DONE Teach Yen

DONE Teach Duc Anh

DONE Schedule weekly logseq report

Doing ((69a2a14e-6983-4250-83b5-2640f836b634))

LEARNING Really impressive with the performance and smartness of ampcode. Much better than claude code for sure. The kill point here is its tool calling skills.

Weekly report 2026 W09 (Feb 10 - 28): https://github.com/lelouvincx/second-brain-logseq/pull/1

IDEA Schedule a cron for weekly pull request, showing all logseq changes this week.

• Context:
  • I use logseq as second brain library, to note all tasks and personal knowledge.
  • Then sync them as-code on github (github.com/lelouvincx/second-brain-logseq).
  • Now I want a personal assistant that can do weekly sync up for myself, on all the tasks / issues / learnings this week / month / quarter / half / year.
  • This can be used for:
    • Personal revision and future planning.
    • Easier when writing company updates such as Half Performance Review (for example, [[Performance Review - 2025H2]]), weekly sync every wednesday, 1on1 with manager.
    • Career / life recommendation.
• But firstly I need to do it manually.
  • Tried it here: ampcode
    • It looks so smooth with the assistance of amp.
  • Manual workflow:
    • Every week is a new branch. Convention: weekly/2026-w10. logseq.order-list-type:: number
    • Look at the changes in their Logseq repository and create a weekly personal report as a GitHub pull request. logseq.order-list-type:: number
      • Example PR: https://github.com/lelouvincx/second-brain-logseq/pull/1 logseq.order-list-type:: number
    • Read the reference links from the notes and tasks to enrich the PR description: logseq.order-list-type:: number
      • Slack links - use the slackcli skill. logseq.order-list-type:: number
      • Notion links - use the notion skill. logseq.order-list-type:: number
      • Ampcode thread links - use read_thread. logseq.order-list-type:: number
      • GitHub links - use gh CLI. logseq.order-list-type:: number
      • Gemini links - use read_gemini_app skill (not built yet - researching at https://gemini.google.com/app/e2f57fa7dd0a35bc). logseq.order-list-type:: number
      • Web links - use read_web_page. logseq.order-list-type:: number
• My plan is to schedule it as a cronjob to run 2 times a week:
  • One on end of tuesday: for data sync meeting on wednesday - create a pull request and leave there til weekend.
  • One on end of friday: for comprehensive weekly report - update the existing pull request.
• Use amp cli (headless) to do it in background.
• Run fully local (macos).
• Highlight this week: win story with Trust Wallet (Slack)
• IDEA Create a github bot account to automate things for me (create pr, code review, deployment) just like duyetbot. I name it lelouvincx-bot

Tasks

Done

• DONE Setup bigquery cli skills

https://benn.substack.com/p/no-really-everything-becomes-bi

• Insight is relative.
• It is dependent on what people already know.
• You can’t tell people something surprising without knowing what is expected.
• Insight isn’t the data, it is what’s not in our audiences’ head.

https://benn.substack.com/p/go-crazy-folks-go-crazy

• He told a common pattern he realized in the software industry: google is afraid of publishing chatgpt, and openai told why not?; copilot is afraid of yolo editting the code and claude code told why not?; same and same thing.
• Before we all had computers and phones and instagram, making art was hard. Because art was expensive and somewhat scarce, we valued the art.
• Then it became easier to make. Everybody can make art. Even without thinking or caring about their art. The value fell.
• But to someone who really make art, it will become extremely valuable. Like craftman things.
  • For example, tranh thêu Tô Châu https://media.secondbrain.lelouvincx.com/2026/02/b19a35a9766eb9fd4e344c0867e43d9c.png

Tasks

Done

• TODO Answer Vincent & Sriram on sales motion (trial-first vs call-first)

Full context moved to [[Lead Funnel by Sales Motion]]

Tasks

Done

• TODO Fix the temp bridge table

• Approach: Treat atc_price_history as alternative to retailer_current_pricing.

• Current state: using a table bridge_product_retailer as a mapping between atc_price_history and retailer_current_pricing. The result is when querying historical prices, Redshift has to join a long path from master_product -> retailer_current_pricing -> bridge_product_retailer -> atc_price_history just to query the historical price, while expect to join master_product to atc_price_history => leads to slow performance.

• The only difference between those 2 is that atc_price_history contains historical price, while retailer_current_pricing contains present price. But similarly they are fact tables telling the price of a product inside a retailer.