LEPRO Lights Revisited, DIY Storm Guard and Safer AI Agents – HGG683

Posted on Sat July 11, 2026 by Jim

I revisit the LEPRO permanent outdoor lights after nearly six months of real-world use, including the removable wood-shim mounting method, seasonal lighting scenes and the outdoor smart-plug schedule. I also walk through several OpenClaw projects, including the Home Assistant Storm Guard I built for my EcoFlow and BLUETTI batteries and an Amazon Affiliate Manager designed to help me work through years of aging links on TheAverageGuy.tv.

These projects have moved me beyond experimenting with AI tools and into building systems that need to be dependable, observable and safe. I explain why I now break projects into smaller steps, verify sensors and conditions before adding automation, run dry tests, limit authority and save successful implementations as gold standards.

I also cover how ChatGPT, OpenClaw, Hermes and Home Assistant fit into my current workflow, what I have learned about agent costs and memory, how my home-energy dashboards have improved, and why I am adapting parts of the OpenClaw article series for Medium.

Audio Only

LEPRO Lights Revisited, DIY Storm Guard and Safer AI Agents – HGG683

Download Other Options

Chapters

[0:24] — Summer Heat, Publishing and YouTube Updates
[5:05] — Six Months with LEPRO Permanent Outdoor Lights
[13:53] — OpenClaw Lessons About Autonomy and Authority
[17:32] — Building a DIY Home Assistant Storm Guard
[25:45] — Testing Storm Guard During Real Severe Weather
[33:22] — Amazon Affiliate Manager and AI Guardrails
[40:20] — Gold Standards, Documentation and AI Memory
[42:50] — Local AI Costs, Hardware and Community
[50:01] — Better Home Assistant Dashboards and Wrap-Up

This Week on Home Gadget Geeks

The Gutenberg incident and the operating model that followed
DIY Home Assistant Storm Guard for EcoFlow and BLUETTI batteries
Real-storm testing with National Weather Service alerts
Why manual control still matters in weather-aware automation
Using AI to build Home Assistant automations safely
Amazon Affiliate Manager and guarded WordPress repair work
Gold-standard workflows, AI costs, local hardware and Medium experiments
Home Assistant dashboard improvements around energy and storm readiness

Full Show Notes

LEPRO Permanent Outdoor Lights After Nearly Six Months

I begin with an update on the LEPRO permanent outdoor lights first featured on HGG667. I have now had nearly six months to use them through several seasons.

LEPRO supplied the lights for the original HGG667 review.

You can view the LEPRO permanent outdoor lights on Amazon here.

The system includes individually addressable LED lights that can display solid colors, animated patterns and seasonal themes. I have used them for Valentine’s Day, Easter and other neighborhood displays. Because each light can be controlled individually, they can also function as targeted downlighting or supplemental security lighting.

Rather than fastening the lights directly to the vinyl siding or soffits, I mounted them to sections of wooden shims. The shims were then inserted into gaps along the siding and soffit. That avoided drilling permanent holes into the house and made it possible to remove or reposition the lights later.

The improvised installation has held up well. Only one light has come loose, possibly after being disturbed by an animal. A purpose-built, 3D-printed mounting shim could be a useful accessory for installations like mine.

The lights are controlled through the LEPRO app, which provides themes, colors and programmable effects. I also placed the system behind an outdoor smart plug. The lights turn on around sunset and shut off at 10 p.m., preventing them from remaining on late into the night.

After nearly six months, the verdict is positive. The lights have remained reliable, the flexible mounting method has worked, and the ability to change the display throughout the year makes permanent lighting more useful than traditional holiday lights.

I also discuss possible locations for this type of lighting, including decks, patios, sheds, workshops, security-lighting zones and outdoor entertainment spaces.

The Gutenberg Incident and the Operating Model That Followed

I then turn to the ongoing OpenClaw project and the series of articles documenting what I have learned.

The articles are not intended as traditional reviews or step-by-step tutorials. Instead, they are thinking guides based on real projects, mistakes and recoveries. The central lesson is that AI agents should not receive broad autonomy before the operator has established authority, visibility and testing.

That lesson came from experience. In one early project, an agent changed a large collection of WordPress posts from Gutenberg blocks to classic HTML. That disrupted PowerPress shortcodes and other formatting across roughly 100 posts.

The agent had not intentionally done anything malicious. It simply had more authority than the workflow could safely support and lacked instructions about what must remain unchanged.

Recovering from that event helped shape my current operating model:

Define the agent’s role.
Determine what it is allowed to observe.
Inventory the environment.
Classify what it finds.
Test proposed changes without applying them.
Apply a change to one controlled item.
Verify the result.
Expand only after repeated success.
Reduce the batch size again when an error appears.
Preserve the successful process as a reusable standard.

This incremental approach is slower than issuing one enormous prompt, but it produces more predictable results and makes failures easier to diagnose.

Building a Home Assistant Storm Guard

The most practical example in the episode is my Home Assistant-based Storm Guard system.

I use both EcoFlow and BLUETTI portable power systems. The EcoFlow unit supports the computers and receives solar input, while the BLUETTI supports the monitors. Under normal conditions, I allow the batteries to cycle rather than keeping them at 100 percent.

EcoFlow offers its own Storm Guard feature, which can charge a battery automatically when severe weather is approaching. BLUETTI does not offer the same capability, and the built-in EcoFlow behavior was more aggressive than I wanted.

That created an opportunity to build a custom version in Home Assistant.

The system brings together battery state-of-charge sensors, current power draw, solar input, slow and fast charging controls, National Weather Service alerts, watch and warning classifications, estimated remaining runtime, automatic reset behavior and manual override controls.

The first step was not writing the final automation. It was confirming that Home Assistant and OpenClaw could see the correct sensors and switches.

Next came the weather inputs. National Weather Service alerts were connected to Home Assistant and exposed as sensors. The system needed to distinguish between alerts that affect outage readiness and those that do not. A heat advisory, for example, does not necessarily justify the same battery response as a severe thunderstorm warning, damaging-wind event or winter storm.

My current logic uses different charging responses for different alert levels. A watch can begin a slower charging posture. A warning can move the batteries to faster charging. When the alert is cleared, the system returns the controls to their normal state.

The charging switches also had to be made mutually exclusive so that two incompatible charging modes could not remain active simultaneously.

Testing the Weather Automation During a Real Storm

After simulations and dry runs, the automation was tested during an actual severe-weather event in Bellevue.

I used ChatGPT and OpenClaw to monitor the system while the alert moved through its stages. The process gave me a chance to compare what the automation was expected to do with what Home Assistant and the batteries were actually doing.

The first real alerts also exposed an incorrect assumption in the original configuration. Because the actual National Weather Service data did not arrive exactly as expected, part of the logic had to be corrected.

That was one of the most valuable parts of the project — it is why automations need observable state and a record of what actually happened, not just a code run that did not throw errors.

The successful test was documented so the working behavior could be reviewed and reused later.

Manual Control Still Matters

Although I trust the automation, the dashboard preserves manual controls.

The system calculates how long the batteries may last at the current power draw. That estimate gives me enough context to decide whether to begin charging early, shut down equipment or change the charging rate manually.

The automation assists with preparation, but it does not remove the operator’s ability to make a different decision based on conditions that may not be represented in the weather feed.

The dashboard places battery runtime, charging state and manual slow and fast charging controls together. That makes the system useful both during automated operation and when I want to intervene directly.

Using AI to Build Home Assistant Automations

The same incremental principle that drove the overall operating model applied specifically to the Home Assistant build.

Before this project, I had struggled to create reliable Home Assistant automations. ChatGPT and OpenClaw made the work more accessible by translating natural-language requirements into code, tests and validation steps.

The biggest improvement did not come from asking AI to build the entire system at once. It came from using AI to work through the problem incrementally:

Find the relevant sensor.
Confirm its current state.
Identify the control switch.
Test the switch manually.
Define the alert condition.
Simulate the condition.
Observe the response.
Record what happened.
Correct assumptions.
Repeat the test.

I recommend using a capable frontier model for this kind of work. Smaller models may be adequate for summaries or simple tasks, but complex Home Assistant, API and automation work benefits from stronger reasoning and coding performance.

Creating an Amazon Affiliate Manager

The second major OpenClaw example focuses on years of Amazon affiliate links across TheAverageGuy.tv.

The site contains links created under several generations of Amazon affiliate tools and formats. Some remain valid, some point to products that no longer exist, some use older link structures, and some are difficult to interpret without reviewing the surrounding text.

Instead of telling OpenClaw to replace every questionable link, I created a dedicated Amazon Affiliate Manager with a specific job and limited authority.

The manager’s work begins with observation:

Inventory the existing links.
Identify the different link formats.
Determine which links still work.
Classify broken or ambiguous links.
Compare the linked product with the surrounding article.
Identify candidates for safe replacement.
Separate clear cases from those requiring manual review.

Access to Amazon’s Creators API would make some of this work easier because the system could search for current products and approved links. Listener use of the site’s Amazon affiliate link helped me reach the transaction threshold needed to pursue that access.

Even with better product data, the system still needs guardrails. A valid Amazon link is not necessarily the correct replacement for an old reference. The visible text, original product and context of the article must all agree.

Authority, Guardrails and Batch Size

I return several times to the importance of controlling agent authority.

An agent needs to know both what it should do and what it must not do. For the Amazon Affiliate Manager, that includes rules such as preserving Gutenberg, avoiding changes to unrelated content, not inventing product matches and stopping when the evidence is ambiguous.

Batch size is also part of the safety model.

During a separate YouTube project with Hermes, I initially applied changes one item at a time. After several successful updates, the batch size was increased to five. An error appeared in that first larger batch, so the process returned to single-item changes until the problem was understood.

That experience reinforced a practical rule: increasing scale is not proof that the underlying process has become safe. Every larger batch creates another opportunity to discover an assumption that did not appear in earlier examples.

Creating a Gold Standard

When a workflow finally produces the correct result, I save that implementation as a gold standard.

A gold standard can include the final structure, required fields, approved formatting, validation checks, examples of correct output, restrictions on what may change, recovery instructions and known edge cases.

Saving this locally in OpenClaw is more reliable than assuming a cloud-based AI service will remember every detail from an earlier conversation.

Memory systems can be useful, but they are limited. I recently had to condense my ChatGPT memory by giving the stored material to another AI, asking it to compress the information and then replacing the longer version with the condensed result.

That experience further emphasized the value of keeping operational knowledge in documented standards rather than relying exclusively on conversational memory.

AI Costs and Local Hardware

These projects have not required expensive new servers.

OpenClaw and Hermes are running on an older Linux laptop. My primary recurring AI expense is the $20 ChatGPT subscription.

The larger concern is token consumption. Agentic systems can generate enormous amounts of usage when they inspect files, write code, rerun checks and repeat tasks. Metered API access can become expensive quickly.

I have been using an OpenAI OAuth connection associated with my ChatGPT subscription. It still has limits, but those limits have been more practical for my workload than paying separately for every token consumed by the agents.

The broader lesson is that running the software may be inexpensive, but poorly controlled agent workflows can consume model capacity rapidly. Guardrails, smaller jobs and deliberate testing help manage both risk and cost.

Expanding the OpenClaw Series to Medium

I am beginning to adapt parts of the OpenClaw article series for Medium.

The goal is not simply to copy the WordPress posts onto another platform. Medium provides an opportunity to experiment with shorter versions, alternate structures and more practical material, including some of the prompts used during the projects.

The Medium effort is still an experiment. I plan to see how readers respond before deciding how far to expand it.

The original OpenClaw series on TheAverageGuy.tv remains the fuller account of the lessons, mistakes and operating principles behind the work.

Home Assistant Dashboard Improvements

The episode closes with a broader update on Home Assistant.

My dashboards now provide a clearer view of EcoFlow and BLUETTI battery levels, current power consumption, charging rates, solar input, projected runtime, weather-alert state and manual charging controls.

The solar dashboard shows the daily production curve as the panels begin generating power, rise toward peak production and taper later in the day.

Building these dashboards has changed how I use Home Assistant. Instead of displaying every available sensor, the dashboards are increasingly organized around specific decisions: how much power remains, whether severe weather is approaching, whether charging should begin and what controls need to be available.

I am considering a future episode with Gavin Campbell focused on walking through the dashboard design and Home Assistant improvements.

Key Takeaways

Permanent outdoor lighting is more useful when it can change with seasons, events and security needs.
AI agents should earn authority through observation, testing and small verified changes.
Home Assistant automations are safer when sensors, switches and alert conditions are verified before control logic is added.
Weather-aware battery charging needs observable state, manual override and careful alert classification.
Affiliate-link repair requires context, not just a replacement URL.
Gold-standard workflows are more reliable than assuming a chat model will remember every operational detail.
Local hardware can run useful agent workflows, but token consumption and model limits still matter.
Dashboards work best when they support decisions instead of displaying every available sensor.

FAQ

What is Home Gadget Geeks 683 about?

HGG683 is a solo episode where I revisit LEPRO permanent outdoor lights, explain my DIY Home Assistant Storm Guard automation and walk through OpenClaw lessons about safer AI agents, Amazon affiliate repair work and gold-standard workflows.

Who hosts HGG683?

I host HGG683 as a solo episode of Home Gadget Geeks.

What LEPRO product is discussed in this episode?

I discuss the LEPRO permanent outdoor lights first featured on HGG667 and explain how they have performed after nearly six months of use.

Did LEPRO supply the lights?

Yes. LEPRO supplied the lights for the original HGG667 review, and that disclosure is preserved in this post.

What is Storm Guard?

Storm Guard is my Home Assistant-based approach for preparing EcoFlow and BLUETTI battery systems when severe weather alerts indicate higher outage risk.

How does Storm Guard use weather alerts?

The system uses National Weather Service alert context to distinguish watch and warning conditions, then maps those conditions to different charging postures while preserving manual control.

Why did I build my own Storm Guard?

EcoFlow has a built-in Storm Guard feature, but I wanted a more measured Home Assistant approach that could also account for BLUETTI, solar input, battery state and manual override controls.

What did I learn during the real storm test?

The real storm test showed why actual alert data matters. One assumption in the original configuration did not match the live National Weather Service data, so the logic had to be corrected.

What is the Amazon Affiliate Manager?

The Amazon Affiliate Manager is an OpenClaw workflow designed to inventory, classify and carefully repair aging Amazon affiliate links on TheAverageGuy.tv without guessing replacements or damaging Gutenberg content.

What do I mean by agent authority?

Agent authority means what an AI system is allowed to observe, change or automate. I argue that authority should be limited at first and expanded only after repeated verified success.

What is a gold-standard workflow?

A gold-standard workflow is a documented successful process that captures structure, checks, restrictions, examples and recovery notes so future work can repeat the known-good path.

Does HGG683 cover Home Assistant dashboards?

Yes. I close with updates on dashboards for battery levels, runtime, solar input, charging posture, weather-alert state and manual controls.

Products Discussed in This Episode

Affiliate disclosure: Some product links may be affiliate links. If you use those links to make a purchase, The Average Guy Network may earn a small commission at no additional cost to you. Thanks for supporting Home Gadget Geeks.

LEPRO Permanent Outdoor Lights
EcoFlow portable power system
BLUETTI AC180
Home Assistant
OpenClaw
Hermes
ChatGPT

Links and Resources Mentioned

LEPRO Permanent Outdoor Lights

LEPRO Permanent Outdoor Lights on Amazon: https://amzn.to/44rWDZp
Original LEPRO discussion from HGG667: https://theaverageguy.tv/2026/01/17/celebrating-15-years-of-home-gadget-geeks-with-lepro-outdoor-lighting-ai-hardware-shifts-and-crypto-signals-hgg667/?highlight=hgg667

Affiliate disclosure: LEPRO supplied the lights for the original review. As an Amazon Associate, The Average Guy Network may earn from qualifying purchases made through affiliate links. Using the link does not change the price paid by the listener.

Join the Conversation

What permanent outdoor lighting are you using around your home? Have you built weather-aware battery automations, Home Assistant dashboards or AI-assisted workflows of your own?

I would like to hear what is working for listeners, particularly practical automations that solve a real household or content-management problem.

Send feedback to me at TheAverageGuy.tv, leave a comment on YouTube or join the discussion in the Average Guy Network Discord community.

Listen and Subscribe

Listen to Home Gadget Geeks using the player above, subscribe in your favorite podcast app, support the show on Patreon, and join the Average Guy Network community on Discord.