OVERALL KOL SUMMARY ARCHIVE
KOL opinions only · stock news is separateKOL opinions are not confirmed company facts
DAILY UPDATE
Market news
& KOL opinions.
Selected stock news and social media opinions, filtered, compared and summarised.
Historical archive · not a live feed
Daily digest
METHOD & LIMITATIONS
Automation for
daily market updates.
A personal workflow to collect stock news and KOL opinions, reduce repetitive work and keep useful information organised.
01 / PAIN POINTS
Too much information. Too little time.
- Information overload: useful news and opinions are spread across websites and social platforms. Checking them manually is slow and repetitive.
- Provider dependence: as a Hong Kong user, I want to avoid depending on one cloud agent or model account. Regional availability or account access could change. Keeping project knowledge outside the provider is a continuity measure, not a guarantee of uninterrupted access.
- Token cost: passing everything to a large model is easy; keeping each task useful and affordable needs limits on retrieval, reading and retries.
- Security: an agent with broad computer access could expose personal data or execute malicious instructions from retrieved content. Open-source software also needs review before installation or updates.
02 / TASKS & SETUP
Separate routine work from model reasoning.
- Codex / frontier model: planning, comparison and analysis. Hermes runs the configured model for its own research tasks.
- Hermes: local collectors, reusable skills, scheduled jobs and delivery. Deterministic scripts handle routine processing where possible.
- Obsidian: knowledge drafts, project requirements and handoff checkpoints. Executable skills stay in the runtime, not in Obsidian.
- Knowledge writing: create new drafts in Agent_Drafts; do not automatically overwrite formal notes. Keep source claims separate from model analysis.
- Permissions: separate AgentUser account, file access controls and narrowly approved commands. Quote/news access only; no trading access.
- Startup: a Windows login task was installed. Reliable unattended startup after reboot still needs verification.
- Local models: a possible next step for simple tasks; not yet an accepted part of this workflow.
03 / CHALLENGES & SOLUTIONS
Problems encountered during setup.
- Social content access: the current Threads collector uses a dedicated Chromium browser. The earlier Meta API issue needs its original records checked before adding a specific failure claim here.
- Browser disconnected: collection failed even though Chromium was open. Reloading and checking the browser connection restored collection; partial failures stayed visible in the report.
- Too many tabs: research now uses one identified task-owned tab, reuses it and closes only that tab. The original tabs are preserved.
- Missing Unicode output: the browser tool returned empty output for Unicode. Explicit UTF-8 handling was added and tested with both text and tab operations.
- Search rate limit: querying all stocks too quickly failed. Requests were paced to respect the reported limit.
- Scheduled approvals: cron could not request interactive approval. Exact collector commands and a restricted queue gate were approved; broad approval and automatic acknowledgement were not enabled.
- OpenD unavailable: a long collector attempt timed out. A bounded TCP check was added; an open port alone still does not prove a valid login.
- Interrupted processing: persistent pending/prepared batches prevent items being marked complete before the report is received.
04 / LIMITATIONS TO RESOLVE
What still needs work.
- Local dependencies: sleep, reboot, network failures, OpenD and browser connections can interrupt scheduled work.
- Model replacement: Markdown checkpoints support handoff, but switching providers is not yet proven to be seamless. A replacement needs tool and safety checks.
- Token controls: research budgets are documented; live enforcement and measured cost savings are not yet demonstrated.
- Source access: paywalls, login requirements and blocked pages can leave gaps. A public preview is not the full article.
- Security: limited permissions reduce risk but do not eliminate malware, software dependency or data-leak risks. Local execution does not mean model processing is local.
- Source selection: headline filters can misclassify relevance. Important candidates need evidence review, not just a score.
- Website updates: the site now contains selected historical reports for local review. Live daily export and automatic publishing are not connected.
ABOUT ARTHUR
Arthur Kwok
Senior business development and account management professional with 15+ years of experience in revenue growth, strategic account retention and regional market development.
Combines new-business acquisition with long-term account growth, coordinating closely with operations and cross-functional teams to support delivery.
Tech-savvy and adaptable, with a strong interest in AI-enabled workflows and a proven track record of working with internal stakeholders, partners and clients.
My role in this project
I define what the workflow should do, set permissions and cost limits, test the results and track problems that need fixing. Coding and research are AI-assisted.
This project is a practical way to learn how to turn a personal need into a working automation service.