Quick answer: Three practical AI upgrades you can use now (these AI upgrades take minutes to set up): (1) save useful corrections in a lessons file so the AI learns your rules, (2) run agents in a plan, do, check, correct, learn loop, and (3) structure your website so AI search can quote it.
These are not future ideas. They can reduce repeat mistakes, speed up complex work, and make our websites easier for AI search tools to understand. Updated July 2026 for Opus 5, which moved part of the original skill into the tool itself.
- Save useful corrections.
- Let agents check and fix their work.
- Make our websites easy to quote.
These are not future ideas. They can reduce repeat mistakes, speed up complex work, and make our websites easier for AI search tools to understand. Updated July 2026 for Opus 5, which moved part of the original skill into the tool itself.
Upgrade 01
Teach AI once.
AI often repeats the same mistake because the correction disappears inside an old chat. A lessons file turns each useful correction into a rule the system can reuse.
Spot the mistake
A person corrects the output, or the AI catches its own error.
Save one clear rule
The lesson explains what went wrong and what to do next time.
Reuse the lesson
The rule is added to future prompts, so the same error is less likely to return.
Example
Weak lesson
“Make the design better.”
Better Lesson
Specific and reusable
“For Be Known vertical graphics, keep the bottom third clear for subtitles and use the uploaded logo without redrawing it.”
Copy This Instruction
Paste it at the top of a new project instruction, custom instruction, or long-running project chat. Keep the same chat or project for related work so the lessons remain available.
CLAUDE.mdAGENTS.mdLESSONS.mdOne rule we learned the hard way: not every correction deserves to become permanent. A lesson file that logs everything gets long, contradicts itself, and quietly slows every future run. Keep a one-off as a note for that job only. Promote it to a standing rule when the cause is understood, the fix is specific and reusable, nothing already covers it, and the mistake either repeated or cost real time.
Upgrade 02
Use agent loops.
An agent loop does more than complete a task. It checks the result, corrects problems, and carries the lesson into the next attempt.
Plan
Define the goal and what a good result must include.
Do
The agent completes one clear, bounded task.
Check
It compares the result with the rules and success criteria.
Correct
It fixes gaps instead of passing weak work forward.
Learn
It saves a useful lesson, then starts the next cycle smarter.
Plan → Do → Check → Correct → Learn → Repeat
Content
Build and review a campaign
One agent researches buyer questions. Another drafts hooks. A reviewer checks each idea against the ICP, brand voice, and offer before anything is published.
Research
Compare sources in parallel
Several agents collect evidence. A lead agent removes weak sources, resolves conflicts, and turns the findings into one recommendation.
Website Work
Make changes without guessing
An agent edits one section. It then checks the live page, tests links and forms, fixes defects, and records what caused them.
Lead Generation
Research and qualify prospects
Agents find and score prospects against clear rules. A human reviews the final list and message before outreach begins.
Use the smallest loop that works. Several agents working in parallel is the most expensive version of this, and the newest tools only do it when you actually ask. Most jobs need one worker, clear success criteria, and one honest check. Reach for a team of agents when the work genuinely splits into separate pieces that later have to be joined back together.
Version 4.0 · Rewritten For Opus 5
The Intelligence Layer
This is a practical operating model for complex work: research, content, websites, automation, coding, migrations, campaign planning, and long multi-step jobs. The strongest model does the hard thinking, routing, and final quality check. Right-sized models do the clearly specified work. Nothing ships on confidence alone. The version we shared first was written before Opus 5. That release moved a large part of the old skill into the tool itself and disproved one of its main assumptions, so the skill below is a rewrite rather than a patch.
Strongest model: route, judge, verify Right-sized model: execute clear tasks Every route: check, correct, learn
What Opus 5 Changed
VERSION 2.2 SAID Hand-build the scaffolding: task lists, worker states, progress tracking, approval steps.
NOW The tool does this natively. Task tracking, isolated workers, completion alerts, and plan approval are built in. Rebuilding them in a skill just spends tokens describing features you already have.
VERSION 2.2 SAID Delegate aggressively by default , push as much work as possible to cheaper models.
NOW Delegation is deliberate, not automatic. The agent does not spin up a team of sub-agents unless you ask for it. Default to one worker doing the job properly, and ask for parallel agents when the work truly splits.
VERSION 2.2 SAID Never name specific models, because the lineup keeps changing.
NOW The lineup is stable enough to name: Opus 5 for architecture and independent review, Sonnet 5 for bounded execution, Haiku 4.5 for mechanical work. Vague tiers made people guess.
VERSION 2.2 SAID The way to save money is to use a cheaper model.
NOW We measured it, and that was the smaller lever. The dominant cost is the conversation being re-sent on every single step. Model choice matters; what you drag along with you matters more.
Pick the lightest thing that can work: a script, one direct edit, one specialist, one bounded loop, or full orchestration. Do not scale up just because you can.
Check what the tool already does before writing process for it. A skill should carry judgment, not restate built-in features.
Opus 5 for architecture and independent review. Sonnet 5 for bounded execution. Haiku 4.5 for mechanical work. Bulk input gets summarised locally first.
Open the file, click the button, run the command. An executor’s summary is a claim. Whoever did the work never signs it off alone.
Send back only the failed evidence and the affected scope. One evidence-driven fix, then stop and raise it rather than looping.
Publishing, sending, spending, deleting, and deploying stay behind human approval, whatever the agent believes it has finished.
The real saving is context, not model choice.
We audited our own usage across 907 sessions. About 94% of all tokens were not new work at all , they were the existing conversation being re-sent on every step, and the five heaviest sessions alone accounted for a third of everything. Anything pulled into a long session is paid for again on every later step, so a file read early in a big job can cost hundreds of times its own size. Practical version: read the part you need instead of the whole file, summarise bulk input before it enters the conversation, prefer a script that returns a number, and finish one job per session instead of carrying five jobs’ worth of history forward.
How To Install It
.claude/skills/intelligence-layer/SKILL.mdAGENTS.mdPaste the full skill into project instructions or a reusable system prompt. It is written to stand alone.
Useful triggers: “Use the intelligence layer,” “Route this properly,” “Architect and QA this,” “Verify this against the real output,” or “Orchestrate this in parallel” when you genuinely want several agents. Note the change: the last one is now a request you have to make, not a default.
Upgrade 03
Prepare for AI search.
People are starting to get answers before they reach a website. AEO helps ChatGPT, Gemini, Perplexity, and Google’s AI results understand what a company does, trust the information, and cite the right page.
of representative real-user queries triggered a Google AI Overview in a 2026 benchmark study.
daily traffic reached matched English Wikipedia pages after exposure to Google AI Overviews.
revenue per visit came from US shoppers referred by AI tools versus non-AI sources.
What This Means
Search is becoming answer-first.
AI answers now appear across Google and standalone tools. They can reduce ordinary website traffic, but the visitors they do send may arrive with stronger intent. The goal is no longer only to rank. The goal is to become a source that AI systems can understand, trust, cite, and recommend.
beknownonline.com
Top 15% of sites scanned · Source: Framer AEO
Category scores: Findable 25/25 · Quotable 25/25 · Understandable 25/25 · Trustworthy 23/25
Main improvement: Add more credible external citations. The detailed report gives external citation links 2/7.
Overall score: 85/100
marketerprompts.com
Top 5% of sites scanned · Source: Framer AEO
Category scores: Findable 25/25 · Quotable 25/25 · Understandable 25/25 · Trustworthy 10/25
Main improvement: Add at least three contextual internal links in the page body. The detailed report gives internal links 0/18.
Overall score: 98/100
Test one workflow this week.
Choose one repeated task
Use a real workflow such as campaign research, content production, or website updates.
Add clear checks
Define what “good” means before the agents start.
Measure the gain
Track time saved, corrections avoided, defects caught, and final output quality.
AI upgrades: frequently asked questions
What are the three AI upgrades and why do they matter?
First, save useful corrections in a lessons file so the AI stops repeating the same mistakes. Second, run agents in a loop that plans, does, checks, corrects and learns. Third, make your website easy for AI search tools to quote.
What is a lessons file for AI?
It is a short file, such as CLAUDE.md, where you record one clear, reusable rule each time the AI gets something wrong, so the next session starts with that lesson already in context.
What is an AI agent loop?
A loop where the agent completes one bounded task, checks its own work, corrects mistakes and records what it learned before repeating. It applies to content, research, website work and lead generation.
How do I prepare my website for AI search?
Make it findable, quotable and easy to understand: clear answers near the top of the page, specific facts, credible external citations and good internal links. The AEO scan results in this article show how two real sites scored.
Where should I start with these AI upgrades?
Test one repeated workflow this week: choose a task you do often, add clear checks for what good looks like, and measure the time or quality gain before expanding.
undefined
Keep reading: see how we build custom AI tools & apps, browse more guides, or talk with us about your own setup.
Further reading: NIST AI Risk Management Framework