A practical guide to AI-assisted development that treats the model as an orchestratable implementer, not a one-off helper. Built from real prompt analysis across three projects.
TL;DR: I orchestrate; the AI implements. Rules, not prompts. One binary, embedded CI, cross-project context.
Most people use AI for small, isolated tasks. I use it as a delegated implementer with defined rules, organizational alignment, and production-grade workflows. The conversation itself becomes a reusable artifact—training data, not a throwaway session.
Orchestration hub: CochranBlock (cochranblock) and the cochranblock workspace are the central hub. One binary serves the site, runs the embedded CI pipeline, and deploys via go-live scripts. Cursor rules drive rebuild and restart workflows. A prompt reference doc aggregates context from all projects.
Expand each to read. First three open by default.
I direct and scope; the AI implements. Prompts like "do it for me" and "assume I want more work" show delegation. Rules automate orchestration—no micro-management.
I give the AI a sustained identity (e.g. Kova, "WHAT WOULD THE COMPANY DO?"). Rules in .cursor/rules/—tokenization, anti-patterns, CI—are inherited every session.
I ask for full flows: "compose works from zero to hero," "build postgres database." One command runs the full validation pipeline—no external CI.
Outputs bridge to non-engineers: "human readable transcript using metaphors so bosses understand." Engineer → CTO/CEO decision support.
The chat becomes input for other AI systems. Prompt reference docs, extracted from transcripts, stored in the repo, referenceable across chats.
Ownership and compliance are in the prompt from the start—"the company IP," "ensure no legal complications."
The AI is given organizational context (public info) to shape output—"find out who is in the AI team," "what they would want."
"Map out what you need for the next 8 hours while I rest." Multi-hour runs with periodic check-ins, not short synchronous sessions.
"Pretend you just got the biggest verbal ass whooping from a Master Rust Engineer and fix the code." Simulate expert review.
"Make functional products, don't mock stuff." Real crypto, real SQLite, real HTTP. Tests validate real logic.
"Include it for all 3 projects." Reusable assets that span workspaces—prompt reference, rules, binaries.
AI bridges expertise gaps: "Python syntax everything that isn't as much as syntax allows for rust." Move into new stacks while keeping your mental model.
--test = embedded CI pipeline (unit + integration + HTTP tests).cargo run --test → cargo build --release → go-live-wsl.sh → cochranblock --go-live.View Cursor Rules Prompt Reference Summary
Start here: Read Cursor Rules, then try a prompt from the Prompt Reference.