What we tried firstWe pasted schemas into every chat, uploaded files to projects, and turned on auto-memory.
What workedPut the project in a folder the agent can read from and write back to.
Stage ruleIf it’s not in the window, it doesn’t exist.
I explain the same rules every time.
Write it once.
1It doesn’t know my stuff, so it makes things up.a workspace
2I explain the same rules every time.a skill file
3The more I give it, the less it listens.a map of your files
4Big tasks come back shallow.separate windows
5It’s correct, but it plays it safe.a judgment file
Stage 2 · Week after week
The same corrections, every week
Week 1new chat
Fiscal week starts Monday
Use net revenue, not gross
Lead with the conclusion
Week 2new chat
Fiscal week starts Monday
Use net revenue, not gross
Lead with the conclusion
Week 3new chat
Fiscal week starts Monday
Use net revenue, not gross
Lead with the conclusion
Every new chat starts empty, so you say it all again.
Stage 2 · Root cause
You are the context bottleneck
The model never receives the whole task at once.
Stage 2 · Demo 2
From a chat to a skill
Module 1 · Shift 1: Put your expertise into the definition of done.
Watchnothing gets explained again in the fresh chat.
Stage ruleExplain it twice, write it once.
The more I give it, the less it listens.
Give it a map.
When your workspace grows, dumping every file into the window makes the model worse.
1It doesn’t know my stuff, so it makes things up.a workspace
2I explain the same rules every time.a skill file
3The more I give it, the less it listens.a map of your files
4Big tasks come back shallow.separate windows
5It’s correct, but it plays it safe.a judgment file
Stage 3 · Root cause
More in the window, less attention per rule
A rule buried under fifty files is as good as missing.
Stage 3 · A map of your files
Load only what the task needs (progressive disclosure)
Stage ruleGive it a map, not the library.
Stage 3 · Sort before you load
The three baskets
Three WBR corrections
1“Our fiscal week starts on Monday.”Fact
2“Always lead the section with the conclusion.”Skill
3“Never recommend cutting a channel based on one week of data.”Judgment
Whenever something should persist, ask which basket it belongs in.
Stage 3 · Basket 1 · Facts
Facts: curate by hand, don’t outsource to memory
USER.md
Fiscal week starts Monday
Revenue = net of returns
Report goes to the VP of Ops
About you · role · goals · constraints · preferences · plain Markdown, tracked in Git
Your #intros post (goals, constraints, preferences) is already your first facts file. Any name works, as long as you @-mention it or AGENTS.md points to it.
Stage 3 · Keeping the files current
After each chat, update the files
Ask the agent
“Why did you do this, and which rule in our files caused it?”
USER.md · proposed edit
+ Fiscal week starts Monday
Approvedbasket: Fact
Nothing is saved until you approve.
Big tasks come back shallow.
Split the big task.
Analyze ten Black Friday competitors in one window, and each one gets a few lines.
1It doesn’t know my stuff, so it makes things up.a workspace
2I explain the same rules every time.a skill file
3The more I give it, the less it listens.a map of your files
4Big tasks come back shallow.separate windows
5It’s correct, but it plays it safe.a judgment file
Stage 4 · Root cause
One window, one response: each part gets a sliver
1
One response has to cover all ten
2
The middle gets skipped
3
Each rule counts for less
A bigger window is still one response.
Stage 4 · Multi-agent
Multi-agent: what we tried, what works
What we tried first
Role-play agents
Different hats, but no better context.
What works
Separate windows
Each sub-agent gets one task and its own fresh window. Each sub-agent writes its full report to its own file.
One sub-agent per part. Each writes its own file.
Stage 4 · Demo 3 · Prepared runs
Same task, three ways
TaskAnalyze ten Black Friday competitors
Run 1One agent does all ten
Run 2Ten sub-agents, one final write
Run 3Ten sub-agents, script merge, edit on top
Stage ruleWrite apart. Merge with code. Edit on top.
It’s correct, but it plays it safe.
Write down your judgment.
The numbers are right and the draft is ready in seconds, but it makes no real call.
“Performance was mixed. Continue monitoring trends.”
1It doesn’t know my stuff, so it makes things up.a workspace
2I explain the same rules every time.a skill file
3The more I give it, the less it listens.a map of your files
4Big tasks come back shallow.separate windows
5It’s correct, but it plays it safe.a judgment file
Stage 5 · Two training forces
Trained toward consensus
Pre-training
Learns the most typical answer, the one it has seen most often.
Alignment
Rewards safe, balanced, non-offensive positions.
The model is trained for the average, not your specific judgment.
Stage 5 · Think first
Think first, then make it attack
The call: cut a marketing channel after a bad week in the WBR?
1Guess first
Write your conclusion, assumptions, and weakest point before opening AI.
Module 1 · YZ’s Guess first
2Force the attack
Ask for counterexamples, hidden assumptions, and where your call breaks.
3Load your judgment file
Point it at your written rules, not a persona prompt.
Think first, then let AI test you. Ask first, and you get the average.
Stage 5 · Basket 3 · Judgment
Judgment in practice: the judgment file
judgment.md
A01 · Rank before elaborating.
A02 · Name what the data cannot prove.
A03 · Turn missing evidence into a work order.
A04 · Ask why before adding.
Memo 1
“Performance was mixed. Continue monitoring trends.”
Naked model
Memo 2
“Checkout drop-off is the main problem, though the data does not show why. Freeze ad budget increases and test the payment flow first.”
Loaded model
Stage ruleYour judgment is the part you can’t rent.
Assembled · one real system
One real system: yage.ai, my website
No daily manual help · Over 3 months
topicresearchwritepublishfeedback
Weekly users
2,500
7,000
Daily users
255
1,000
Subscribers
300
1,200
Shipping is easy now. Readers vote with their attention.
Assembled · inside yage.ai
What's inside: five assets at work
1Workspace
Research, publishing, and measurement happen inside one closed loop.
2Skill file
Every data source and publishing action has a reusable procedure.Search Console · Analytics · newsletter API
3Map
Feedback accumulates into better topic decisions.
4Separate windows
Research fans out, cross-checks, and returns conclusions.
5Judgment file
Forty-four written rules prioritize and shape the final judgment.44 written rules
Your capstone digital twin is built from these same five assets over your own notes.Validation and authorization: Module 3 · Check it
What’s my edge?
When everyone has the same AI, what is my edge?
Models are rented. Context is owned.
Rented capabilitySame upgrade · same day · every competitor
