Tavio | Second Mind

What part of your AI workflow or system do you keep correcting, working around, or trying to make more reliable? AI can produce the work and still leave you doing the hard part: correcting it, checking it, constraining it, and figuring out what should happen next. That’s the gap Second Mind Systems works on. Maybe an agent goes beyond scope. Maybe instructions don’t hold consistently. Maybe the workflow behaves differently from run to run. Maybe you don’t have good evals for deciding whether the AI is actually doing the right thing. We work across: • AI workflow diagnostics • Agent reliability audits • Eval design • Behavior and control specifications • Workflow implementation • Strategy development • Governed AI systems Right now, I’m opening the first 10 Behavior Profile Lab engagements as a bounded way to work with us. Bring one recurring problem in one AI workflow or system, along with real examples of what keeps happening. We reconstruct the problem from evidence, identify the recurring failure, define the behavior and boundaries you need, create the appropriate control or specification when warranted, and provide evaluation cases and a clear path for applying the result. For a problem that fits the Lab, this is the complete bounded engagement—not a consultation that merely tells you what to buy next. If the outcome requires hands-on code changes, integrations, deployment, or live-system work, we identify that as a separate scope rather than pretending the Lab covers unlimited implementation. Founding Lab offer: $200 for one qualifying bounded engagement. First 10 only. Public proof: Governed AI Systems Portfolio github.com/Secondmindsystems… PBM Contract Analysis I analyzed Mark Cuban’s open-source PBM contract and built a control-design contribution around how its protections survive procurement and implementation. Mark Cuban reposted the contribution on LinkedIn. Read the contribution: linkedin.com/feed/update/urn… Mark Cuban’s Repost — LinkedIn reshare panel: linkedin.com/analytics/post/… Governed Change Demo github.com/Secondmindsystems… Connect on LinkedIn: linkedin.com/in/tavio-lawren… If there’s something in your AI workflow or system you keep correcting, working around, or trying to make more reliable, DM me or email [email protected] with a short description of what keeps happening. Please don’t send credentials, private code, sensitive records, or other confidential material in your first message.

Ghost Busters Hi-C Ecto Cooler flavor.

The music cutting off as he killed you was a great touch lol. Thanks for the great content.

Thoughts @grok ? Break this down for us.

Hey Garry, this could create a skill library that continuously evolves alongside the models, identifying which skills still add value, which can be simplified, and which are no longer necessary. The key is testing whether a skill should disappear entirely or only activate for the specific cases where it still helps. Give this to your agent: “Help me evaluate which skills to keep, simplify, restrict or retire after a model upgrade. Start with one skill. Identify what it was designed to accomplish, the failures it was written to prevent, and the conditions where it should provide the most value. Choose representative tasks covering routine work, difficult cases and those specific failure conditions. Define what counts as a successful result and which regressions would make retirement unacceptable. Run the same tasks with the same model, tools and budget in fresh sessions. Compare three approaches: The full skill. A simplified version containing only its essential instructions. The normal setup with that skill removed. Verify which instructions actually loaded in each run. Compare successful outcomes, cost per accepted result, execution time and regressions by task type. Repeat enough cases to distinguish meaningful improvements from noise. Investigate whether the skill provides unique value only under certain conditions, or whether the model now performs equally well without it. Recommend whether to keep, shorten, restrict to specific cases or retire the skill. If the evidence is insufficient, identify the next experiment needed before deciding. Preserve the evaluation cases, results, assumptions and exceptions so the comparison can be repeated after the next model upgrade. Do not modify or remove anything without my approval. Explain the result in plain English.”

Cool

@grok Break it down and explain/educate the the significance of this prompt - Give your honest thoughts.

Mark’s idea could help employers find out where their healthcare benefits money is going, what they might be overpaying for, and what they can negotiate. Here’s a prompt to operationalize his idea and turn that analysis into an actionable negotiation brief. Give this to your AI: “Review these healthcare benefits contracts to identify avoidable costs, unfavorable terms and practical negotiating options. First establish our plan structure, renewal deadline, priorities and benefits we need to preserve. Inventory the documents and flag missing agreements, amendments, fee schedules and compensation disclosures. Reconstruct who pays whom, for which services, including broker or consultant compensation. Compare current and proposed terms where available. Examine the agreements together. Identify conflicting clauses, financial incentives, restrictions and renewal or termination conditions that could materially affect our options. For every finding, cite the document, page and clause. Explain the consequence, supporting evidence and remaining uncertainty. Separate unfavorable terms from suspected errors or legal questions. Do not assume misconduct. Rank the strongest actionable findings. Give us specific questions, document requests and proposed negotiating changes. Estimate savings only where the inputs support calculation, showing assumptions, transition costs and possible effects on employee benefits. Produce a concise brief we can take to the provider and a qualified benefits adviser or counsel. Explain the result in plain English.” For independent review, open a second model and provide the same documents. Ask it to analyze them before showing it the first report. Then ask it to compare findings, challenge unsupported conclusions and preserve unresolved disagreements. And agreement between models is not proof -

All you have to do is apply this to yourself and you will remain happy. Their 5 man group has no affect on your personal WoW experience. Let them enjoy and choose how to want to play their journey.

Let them enjoy the game and just focus on your own personal journey bro.

There’s another angle to this. What needs to survive when a complex system gets simplified for human review? A statechart can make behavior easier to understand, but two states that look similar may have very different consequences. For example, “payment pending” could mean payment was never submitted, or payment succeeded but confirmation was lost. Those require different retry decisions. (Give us your thoughts @grok) Give this to your agent: “Use /statechart to help me review whether this feature behaves as intended. First identify the decision I’m trying to make and which behaviors matter to that decision. Reconstruct the relevant states, transitions, conditions and recovery paths from the available requirements and code. Separate agreed requirements, observed behavior and assumptions. Before simplifying the chart, identify which differences between states could change my decision. Preserve those distinctions and explain what was merged or omitted. Run available model checks and replay reported bug paths against the implementation. Report any discrepancies, unchecked properties and limitations without treating model findings as confirmed implementation defects. Show me what the evidence establishes, what remains uncertain and what still requires my judgment.” The goal is to make the system easier to review without hiding the behavior that could change the conclusion.

Agreed. Here’s the operational version of what you explained. This turns examples of how you work into an automation plan that captures your decisions, exceptions and standards. @grok give us your thoughts as well. Give this to your agent: “Interview me about one recurring task I want off my plate. Ask one question at a time. Reconstruct a recent example from trigger to finished result. Identify the inputs, steps, decisions and signs I use to judge quality. Ask about a time I handled it differently or corrected someone else’s work. Use those differences to uncover rules I haven’t explained. Show me your understanding and let me correct it. Separate repeated patterns from assumptions. Recommend what to simplify, automate or keep with me. Choose the smallest useful trial and define how we’ll check quality, review effort and time saved. Return a reusable instruction another agent could follow, including the information it needs, what it may do independently, when to ask for help and how to recognize a successful result. Explain the result in plain English.”

1 am. Had to be to work 6 this morning. Can’t do that again.

Nice. This changes the agent’s objective from simply completing a change to establishing whether the change works. One useful addition: ask the agent to define what would demonstrate success before it starts testing. Give this to your coding agent: “Before implementing, identify the intended behavior, what must remain unchanged, and failure conditions I may have missed. Define the evidence needed to check each, then implement and run the appropriate checks within existing permissions. Report what passed, failed or wasn’t tested. Explain what each result establishes and what remains uncertain.” This helps expose gaps that passing the existing tests might leave untouched.

What actually changes when prescription pricing accounts for both the patient and the pharmacy? @grok investigate how Team Cuban Card works using its official website, patient FAQ, pharmacy FAQ and Mark Cuban’s linked Drug Topics interview. Reconstruct the pricing model from first principles. Explain how drug acquisition cost, the 15% markup, the pharmacy dispensing fee and the processing fee contribute to the final price. Show what the patient pays, what compensation the pharmacy receives and what remains unknown about its actual profitability. Use a clearly labeled hypothetical prescription to demonstrate how quantity affects the total price and dispensing cost per day. Only consider larger fills when appropriately prescribed and legally permitted. Then examine the broader economic implications. How could this structure affect affordability, local pharmacy access and the ability to compare prescription benefit arrangements? Distinguish the retail card from the funded employer pilot. Identify what must be true for each intended benefit to materialize, including medication availability, participating pharmacy access and the patient’s insurance situation. Finally, give someone a practical way to compare Team Cuban Card against their existing prescription option. Separate documented features, possible benefits and outcomes that have not been demonstrated. Explain the result in plain English.

How Team Cuban Card connects affordable prescriptions with local pharmacy access. There’s an important economic idea behind Mark Cuban’s prescription card. Making medication affordable matters. But the pharmacy dispensing that medication also needs to be compensated for its work. Team Cuban Card addresses both sides of the transaction. Here’s how the standard pricing works: Drug acquisition cost + 15% markup + $12 pharmacy dispensing fee + $1 processing fee. The dispensing fee stays with the pharmacy, making compensation for its service an explicit part of the price. That changes what patients can understand about the transaction. They can see the pricing structure rather than only the final amount. Prescription quantity introduces another consideration. When a larger fill is appropriate, prescribed and legally permitted, the dispensing fee can be spread across more medication. The actual price still needs to be checked. The broader implication: Transparent pricing can help patients compare their options while creating a clearer economic relationship with participating local pharmacies. It may also provide a useful reference for employers evaluating prescription benefit costs, although broader savings and adoption remain to be demonstrated. The card won’t necessarily be the cheapest option for every medication. For patients, the practical questions are whether their exact prescription is covered, whether a participating pharmacy is accessible, and how the total compares with their current option, including insurance considerations. Team Cuban Card is not insurance, and purchases do not automatically count toward an insurance deductible. The larger principle is that affordability and local pharmacy sustainability should be designed together, not treated as separate problems. That’s what makes this pricing model worth examining beyond the cost of a single prescription.

I have tonight 6pm to ship something before WoW forever and tomorrow until Ocarina of Time. I’m screwed.

Agreed 100%. This is what we’re working on with AI and video games: worlds where the history of play changes what can happen next. Player actions could create lasting consequences across characters, factions and game systems. Relationships could change access. Evidence could change how a guild responds. Changes in resources or ownership could create new opportunities and conflicts. Those consequences would shape later decisions, allowing the world to develop through play within rules defined by its developers. For open world games, this could mean factions changing alliances as evidence spreads, disrupted trade changing local availability, or a quest gaining a different route because an earlier decision changed who holds power. Characters could act on incomplete information, spread lies or pursue conflicting goals, with consequences carrying across encounters and locations. The broader possibility is a world whose opportunities and conflicts develop from accumulated play, giving developers more combinations of meaningful consequences to work with and players a lasting influence on how the world unfolds. We have a small working demonstration of evaluating and persisting a character’s response. We're now working on connecting that across a larger world.

6 agent architectures. 6 ways to put them to work. A few practical applications generated through our harness. Some ideas you can give your coding agent: 1. LangGraph: Persistent Research Agent “Build a research workflow that can investigate a complex question across multiple sessions. Preserve findings, sources, unresolved questions and decisions. Let it pause, resume and revisit earlier conclusions when new evidence appears. Show how its understanding evolves without restarting the investigation.” 2. CrewAI: AI Product Discovery Team “Design a team of agents that turns a customer problem into a product opportunity. Have one investigate unmet needs, another research alternatives, another challenge assumptions and another develop the smallest useful prototype. Make their findings build on each other and return one actionable recommendation.” 3. Microsoft Agent Framework: Business Process Agent “Build a workflow that takes a customer request from intake through completion. Identify required information, business rules, approval points and exception paths. Let agents handle routine steps while preserving human decisions where necessary. Show the status and evidence behind every completed step.” 4. OpenAI Agents SDK: Customer Resolution Team “Design a customer support system where a triage agent understands the request and hands it to the appropriate specialist. Preserve relevant context, permissions and unresolved issues across handoffs. Let the specialist solve the problem or escalate it, and show whether the customer actually reached a resolution.” 5. Google ADK: Multi Agent Opportunity Explorer “Build an orchestrator that investigates a business opportunity from several directions. Run market research, customer analysis, technical feasibility and economic modeling as coordinated tasks. Reconcile conflicting findings, identify missing evidence and recommend the next experiment worth running.” 6. Claude Agent SDK: Autonomous Development Assistant “Build a coding workflow that takes a feature request through implementation and verification. Inspect the existing project, reconstruct intended behavior, make scoped changes, run available tests and demonstrate the result. Preserve decisions and evidence so another developer can understand what changed and continue the work.” Each architecture should solve a real problem, not just demonstrate agents talking to each other.

@grok thoughts about these prompts and use cases?

Five ways to put @bot to work as your personal shopping team: 1. Product Researcher 🔎 @bot ask what I need, my budget and how I’ll use it. Research suitable products, compare meaningful features, identify common complaints and show what information is still missing. Recommend what deserves closer consideration. 2. Purchase Advisor 🧠 @bot compare my shortlisted products against my priorities. Include total cost, reliability, warranty and tradeoffs. Recommend the strongest fit and explain when another option would be better. Don’t choose based on price alone. 3. Purchasing Agent 🛒 @bot help me prepare an order for my selected product. Check the seller, final price, shipping, return policy and delivery details. Show me the exact purchase for approval before placing anything. If checkout isn’t available, give me the steps to complete it myself. 4. Order Tracking Agent 📦 @bot help me track my purchases. Use order information I provide or authorize you to access. Summarize delivery status, delays and anything requiring my attention. Don’t invent tracking updates. 5. Returns Assistant 🔄 @bot help me return a purchase I don’t want. Check the return window, seller policy, refund method and required steps. Prepare the return request and ask for my approval before submitting it. If you can’t submit returns, guide me through the process.