Product Deep Dive
How STEEPWORKS Compounds Your Knowledge
A deep dive into the knowledge system you're buying
Generic AI starts fresh every time. Yours remembers.
Same prompt. Different output.
Generic AI
"Here's a cold email for enterprise accounts using the AIDA framework with a pattern-interrupt opener..."
Good but generic advice. You'll customize it anyway.
Every session starts from zero.
Your AI
"Loading cold email writer using your 'challenge the status quo' opener that worked on Acme last quarter..."
Your words. Your wins. Remembered.
Every session builds on the last.
You're not teaching AI from scratch. You teach it YOUR business once. Then it works for you forever.
Solo founder wearing every hat? 20-person GTM team? This scales with you.
How the Compound Effect Actually Works
The more you use it, the smarter it gets. That's the whole point.
Four tiers. Zero bloat. Permanent memory.
Every correction you make compounds into learning rules. 500 sessions later, context loads in the same time as session one.
FEEDBACK
ActiveYour edits, ratings, corrections—the raw signal
WORKING MEMORY
HotLast 10 sessions per skill—always loaded
SUMMARIES
WarmCompressed patterns by theme—loaded on demand
FACTS
PermanentLearning rules and insights—queryable forever
Why layered memory matters
Most AI tools forget everything between sessions. This architecture captures your corrections at multiple levels. That means patterns you teach once become permanent improvements.
Hot tier catches everything
Every edit, rating, and correction is captured in real-time
Patterns compress automatically
3+ similar corrections become a summarized preference
Rules survive forever
Validated patterns become permanent learning rules
From edit to autopilot
Your third correction is your last.
Edit captured
T0: Feedback
3x threshold
Pattern detected
Rule generated
T2: Summary
Human review
Validation gate
Integrated
T3: Fact
Edit captured
T0: Feedback
3x threshold
Pattern detected
Rule generated
T2: Summary
Human review
Validation gate
Integrated
T3: Fact
The insight:
You're building a knowledge base that captures exactly how you want AI to work. Your workflow generates your context and compounds your competitive advantage.
You could figure this out yourself. In about six months.
Or start Monday with a system that already works.
Build It Yourself
Then: Solo maintenance. Every problem feels like your fault. No one to ask "is this normal?"
Start With This
To first productive session. Then: quarterly updates, Discord of builders, patterns stress-tested by others.
What you're actually buying:
- •42 working skills across 7 categories: research, content, planning, code, outreach, review, automation
- •Folder architecture designed for 10,000+ files without chaos (the structure I rebuilt 4 times)
- •Skill chaining protocol: one skill triggers another—research → draft → review in a single command
- •YAML-driven knowledge graph with auto-linking (stops the "where did I put that?" problem)
- •The CLAUDE.md patterns that prevent context loss, compaction disasters, and skill conflicts
30-day "actually works" guarantee. If the system doesn't click within 30 days, email me and I'll refund you. No elaborate justification needed—"not for me" is enough.
The question isn't whether you can build this yourself. You can.
The question is whether 6 months of your time is worth $299.
Auto-Loading Rules
Rules live in files. Files have path patterns. When you work on a matching file, rules load automatically.
The Context Loading Hierarchy
- → Loads on EVERY conversation
- → Who you are, folder structure, core patterns
"Your AI's permanent memory"
- → Each folder has its own README.md
- → Loads when working IN that folder
"Context that scales with your structure"
- → Pattern-matched rule files
- → Load based on file paths being edited
"Behavior that adapts to location"
The Mechanism
paths: frontmatter
Each rule file declares which file patterns trigger it. Glob patterns like "03_finance/**" match any file in that folder.
Pattern matching
When you open or edit a file, the system checks its path against all rule patterns. Matching rules activate.
Composable rules
Multiple rules can match. A file in 04_projects/steepworks/ might load both project-rules.md AND steepworks-specific rules.
Write once, apply forever.
Rules are composable infrastructure. You define behavior once, and it applies automatically every time you touch a matching file.
# Finance Domain Rules
---
paths: ["03_finance/**"]
last_updated: 2026-01-01
---
## Key Standards
### 1. Confidentiality (CRITICAL)
**ALWAYS** apply strict confidentiality rules:
- Never expose specific dollar amounts
- Use percentage changes over absolute numbers
- Redact account numbers, SSNs, or other PII
### 2. Precision & Accuracy
**ALWAYS** verify calculations:
- Double-check all mathematical operations
- Cite data sources for all financial figures
- Use consistent decimal precisionPlain markdown. No lock-in. Forever portable.
Your knowledge lives in files you own. Not a database you pay for monthly.
Plain Markdown
Human-readable files. Open in any editor. No proprietary format.
Git-Friendly
Version control that actually works. Diff, merge, branch freely.
Folder = Architecture
No database schema. Your structure IS the system.
Any AI, Any Tool
Works with Claude Code today. Cursor tomorrow. Whatever comes next.
You're not building in a walled garden—you're building a knowledge base.