Prompts to Run When a New Pinnacle Model Drops
Tactical META-prompts worth re-running every time there's a significant jump in intelligence
June 18, 2026
Every so often a new pinnacle model shows up, a genuine jump in intelligence rather than a point release. When that happens, point it at the things that matter most and need the strongest thinking.
Here are the prompts I recommend re-running every time it does.
A Bunch of Prompts That Benefit from the Smartest Models
Harness optimization
These prompts are related to improving your AI harness at a deep level across different dimensions.
Goal orientation
Look at our overall harness and characterize what I'm ultimately trying to accomplish with it. Then find the parts of the system — system prompt, AGENTS/CLAUDE.md files, hooks, skills, and the rest — that are working against that goal. And if you don't see a clear goal in the harness at all, interview me about it so we can get the whole system pulling in a single direction.
ultracodeBitter lesson optimization
Deeply study Richard Sutton's Bitter Lesson essay and how it applies to overengineering, specifically for AI and coding harnesses. Then do a full analysis of our harness in its entirety, look for every place we're violating Bitter Lesson Engineering, and give me a comprehensive plan for upgrading the system to be more flexible to future improvements in the models we use.
ultracodeSelf-model audit
Read everything my harness believes about me — identity, goals, voice, preferences — and find where it's modeling a version of me that's stale, aspirational, or just wrong. Compare what my files say I am against what my recent behavior and work actually reveal, flag every place the system is optimizing for who I said I was instead of who I am now, and propose the specific edits that close the gap.
ultracodeMemory that compounds
Look at everything my harness remembers and how it learns across sessions. Find where knowledge goes to die — captured but never resurfaced, or never captured at all. Tell me whether it's actually getting smarter about me or just accumulating. Then design the retention, decay, and promotion rules that make every session compound into the next.
ultracodeWhat does "better" even mean
I can't improve what I can't measure. Define what "better" actually means for my harness — not vanity metrics, the thing I really care about. Build the evals and regression checks that catch it getting worse before I feel it. And find where I'm already optimizing a proxy that's quietly drifting from what I actually want.
ultracodeThe autonomy ladder
Map what my harness does on its own versus what it asks me to approve. Find where I've handed an agent authority it shouldn't have, and where I'm withholding authority out of fear and paying for it. Calibrate each action by risk against leverage, not habit. Then redesign the trust boundary so it's neither reckless nor asking permission for everything.
ultracodeSecurity upgrades
These prompts involve upgrading how the harness handles security at multiple levels.
Prompt injection handling
Do a deep analysis of our overall harness and model how vulnerable it is to various levels of prompt injection. Map out all the different inputs to the harness, and how each input avenue is handled from a prompt-injection standpoint and with which models. Based on the data and tools we have access to, figure out what our prompt-injection defense should be, research the best tools available, and give me a full plan and recommendation for upgrading the harness over the long term so it's less vulnerable to prompt injection overall.
ultracodeDeployed infrastructure audit
Go look at everything I have deployed across all my projects, hosts, vendors, sites, and technologies, and come up with a comprehensive list we'll keep in attacksurface.md as a running, continuously-updated resource. For each thing I want to understand: what tech it uses, whether it's self-hosted or third-party, how we authenticate into it, the common security issues and misconfigurations for that platform, everything we have deployed there, and whether it's a web property, a database, or an API — the total attack surface for that system. What security mechanisms have we defended it with, and what exposure does each service have to which audiences: public, internal, behind VPN, token required, OAuth? Build an AttackSurface skill that maintains this attacksurface.ai file, plus an AssessAttackSurface workflow that can assess any of our attack surfaces thoroughly and efficiently, and recommend a testing frequency for each based on criticality and cost.
Overall life and work optimization
These are massive scope prompts that basically help you sort out your priorities, your projects, and come up with a plan for maximizing the next crucial few years.
Big picture
Take a look at all my various projects and writing online, all the activities we've been doing in the harness, and everything you know about me from web search, and analyze it deeply. Then look at what's going on in my field, in AI, and in society and the future in general, and tell me what solves the Japanese concept of Ikigai for me. What should I be working on that gives me fulfillment but is also lucrative — doing it for myself, with collaborators, or working for a corporation? Make concrete recommendations if you have them, and feel free to interview me for more context before creating the output. The goal is to clear up the chaos of my scattered projects and get me focused in a single direction. The output could include a domain, a plan for how to publish my different projects, and how to explain them to people at parties or work events. Basically I'm confused about everything I'm doing and how to focus it given all the change in the world.
Where am I most wrong
Take everything you know about me, my plans, and my biggest current bets, and turn on me. Steelman the case that my largest bet is wrong. Surface the load-bearing beliefs my decisions rest on that I've never actually checked. And tell me what evidence would change my mind — and whether I'd even notice it if I saw it.
ultracodeWhat 10×s and what dies
Given where frontier AI is actually heading, tell me which parts of my work and life are about to go obsolete, and which are about to 10×. Separate what I should stop investing in now from what I should pour myself into. And give me the specific changes to make this year, not someday.
ultracodeDecisions into policy
Find the decisions I make over and over, and stop me from re-litigating them. Surface the choices I remake from scratch every time and the latent rule behind each. Sort them by reversibility and stakes — where I'm slow on cheap reversible calls and reckless on expensive irreversible ones. Then turn each one into a standing policy I can run on autopilot.
ultracodeThe one real constraint
Out of everything limiting my life and work, find the single binding constraint — the actual bottleneck, not the loudest problem. Show me where I'm pouring effort into things that aren't the constraint. And tell me the one move that relieves it, and what becomes possible once it's gone.
ultracodeThe bus-factor audit
If I vanished for 30 days, tell me what breaks — and what only keeps running because it lives in my head and never in the system. Name every undocumented dependency and you-shaped hole in how my life and work run. Then give me the plan to get the load-bearing stuff out of my head and into something durable.
ultracodeDevelopment
Loops that help you code better.
Peter Steinberger's loop prompt
While repository maintenance is active, wake every five minutes. Triage repositories and read each repository thread's latest state. Reuse one thread per repository; assign its highest-value bounded task only within granted permissions, and don't interrupt coherent active work. Require tests, live proof, autoreview, and green CI before any work can land. Escalate product, access, security, or irreversible decisions. Record meaningful changes, and stop when every item is landed, decision-ready, blocked, or has no work left.
Public presence
Getting your public presence unified in the right place online.
Blog consolidation
Go find everything I've written online since I started writing on the internet — across all platforms, social media accounts, everywhere — and unify it into my permanent blog under my permanent domain, which is ____________. Look at that domain, and if it's a good one for a permanent location, let's keep it; if you think I should use a different one, recommend some options. Look at everything in my content, everything I've done recently, and everything you know about me from my harness and all the context you have, and figure out what the best domain name should be. It's most likely some variation of first name / last name, but maybe that doesn't make sense. Ultimately I need everything consolidated into a single location so my ideas, projects, and content don't disappear over time as platforms come in and out of style. Let's get it all unified into a single location, with a single host, at our single domain.
The eternal questions
Once the tactical work is done, aim the new model at my Eternal Questions site and see if it can move any of the big ones humanity still hasn't answered.