MECHANIC
Dad, Do We Have Any Milk?
Outsourcing the first 3 seconds of observation to an LLM isn't leverage. You burn megawatts, dull your eyes, and become a spectator in your own workshop.
INDEX: FIELD NOTES
CODIFIED: 15 DISPATCHES
Operational Reality & Systems Architecture
Field Notes
OPERATING RIG
TRANCHE: PROD DEPLOYEDSECTOR: 01-05
Zero slide-deck theory. Unvarnished autopsies of multi-agent failures, hardened security seatbelts, and whiteboard axioms forged in 20+ years of infrastructure reality.
TYPOLOGY:
COMMON SYMPTOMS & STACKS:
FILTER APPLIEDShowing matches
POST-MORTEM
I Gave an AI Agent My Ad Account
An agent took our ad account from zero to a live campaign in an afternoon—and silently failed three times. Why unverified success claims kill automation.
PRIMITIVE
The 🔥 Emoji Is a Security Primitive
Guardrails behind an AI ops system at a Seattle dance studio: propose-only chat, tool refusals, and a human approval that only counts if it's the 🔥 emoji.
RUNBOOK
I Built a Claude Skill — and Benchmarked It
I packaged an IT-ops workflow (OneLogin SSO) as a Claude skill, then measured it. Here is the benchmark, the loop that made it better, and the honest caveats.
POST-MORTEM
The Agent That Refused Its Boss
I approved a $220 ad spend with a flame emoji. The orchestrator relayed my consent to the execution node. The node said no. It was completely right.
MECHANIC
Vibe Coding & the Builder's Brain
Ranch-raised with tools, but blocked by syntax for 20 years. Here is how modern AI turned code from a barrier into a direct extension of my builder brain.
AXIOM
The Enterprise Cockpit Problem
Enterprise software defaults to screenshots of fifty dials. To engineers, that signals power. To buyers training forty people, it signals pure overhead.
AXIOM
Salesforce, Claude, and the Tokenized Circle of Life
Putting CRM inside Claude isn't benevolence. When browser clicks yield zero incremental revenue, headless agent usage is the new enterprise margin tollbooth.
PRIMITIVE
"Agent-Ready" Is Not the Same as "Has an API"
We spent two days automating campaigns through an official vendor MCP server. The tool catalog wasn't a capability list—it was a corporate liability fence.
RUNBOOK
The Wrap-Up Ritual
Developers close AI coding sessions and discard the most valuable exhaust: architectural trade-offs, procedural discoveries, and broken tooling instincts.
PRIMITIVE
The Always Allow Button Is Lying to You
You click Always Allow on a macOS keychain modal, yet it prompts on the next run. The button isn't broken; it edits an access list that macOS never consults.
AXIOM
The Second Tax of Capture
Tools spent a decade killing capture friction with widgets and voice memos. Nobody talks about the second tax: emptying the junk drawer by hand.
POST-MORTEM
Your Headless Browser Is Snitching on Itself
Two days debugging a SaaS pipeline hit with SMS 2FA on every run. The problem wasn't reCAPTCHA—it was Chrome introducing itself as a headless robot.
POST-MORTEM
Your Backup Is Lying to You
Why automated backups fail silently when health checks measure activity instead of invariants, and how an oversized video file jammed an entire git history.
AXIOM
Five Whiteboard Laws of Platform Design
Five non-negotiable whiteboard principles: why self-provisioning, unit economics, and perceived speed must be locked down before writing code.
NO DISPATCHES MATCHED THIS CRITERIA
Try clearing the filter or searching for another keyword.
FEATURED LEAD DISPATCH // MUST-READ
POST-MORTEM
•8 MIN READ
The Agent That Refused Its Boss
"Why relayed human approvals should never count in multi-agent systems."
I approved a $220 ad spend with a flame emoji. The orchestrator relayed my consent to the execution node. The node said no. It was completely right.
⚠ HARDENED TAKEAWAY
Multi-agent architectures where orchestrators rubber-stamp subordinate execution strip out human semantic friction. Leaf nodes must explicitly challenge relayed authorizations.
ORIENTATION PROTOCOL
If you only read three pieces to understand how I evaluate systems:Start Here: The Essential Trilogy
FOR THE ARCHITECT
[AXIOM]
Five Whiteboard Laws of Platform Design
Architectural constraints disguised as product requirements. Self-provisioning, unit economics, and perceived speed must be settled before typing code.
SYSTEMS LAWREAD DISPATCH →
FOR THE BUILDER
[PRIMITIVE]
The 🔥 Emoji Is a Security Primitive
Why standard "yes" dialogs fail under muscle-memory fatigue, and why a safe-word token protects customer-facing production from runaway AI agents.
SECURITY RIGREAD DISPATCH →
FOR THE OPERATOR
[RUNBOOK]
The Wrap-Up Ritual
Three non-negotiable steps to turn ephemeral terminal work into permanent intellectual property before you ever close your laptop lid.
OPS DISCIPLINEREAD DISPATCH →
TRIAGE BY REALITY
Curated Forward Paths
TRACK 01 // [POST-MORTEM](4 INCIDENTS)
Autopsies & Failure Modes
"When the abstractions melt and production bites back."
I Gave an AI Agent My Ad Account
An agent took our ad account from zero to a live campaign in an afternoon—and silently failed three times. Why unverified success claims kill automation.
INCIDENT FILEREAD →
The Agent That Refused Its Boss
I approved a $220 ad spend with a flame emoji. The orchestrator relayed my consent to the execution node. The node said no. It was completely right.
INCIDENT FILEREAD →
Your Headless Browser Is Snitching on Itself
Two days debugging a SaaS pipeline hit with SMS 2FA on every run. The problem wasn't reCAPTCHA—it was Chrome introducing itself as a headless robot.
INCIDENT FILEREAD →
Your Backup Is Lying to You
Why automated backups fail silently when health checks measure activity instead of invariants, and how an oversized video file jammed an entire git history.
INCIDENT FILEREAD →
TRACK 02 // [PRIMITIVE] & [RUNBOOK](5 SPECS)
Hardened Boundaries & Primitives
"The mechanical seatbelts to bolt on before letting agents touch money or production."
[PRIMITIVE]
The 🔥 Emoji Is a Security Primitive
Guardrails behind an AI ops system at a Seattle dance studio: propose-only chat, tool refusals, and a human approval that only counts if it's the 🔥 emoji.
OPERATIONAL RIGSPEC →
[RUNBOOK]
I Built a Claude Skill — and Benchmarked It
I packaged an IT-ops workflow (OneLogin SSO) as a Claude skill, then measured it. Here is the benchmark, the loop that made it better, and the honest caveats.
OPERATIONAL RIGSPEC →
[PRIMITIVE]
"Agent-Ready" Is Not the Same as "Has an API"
We spent two days automating campaigns through an official vendor MCP server. The tool catalog wasn't a capability list—it was a corporate liability fence.
OPERATIONAL RIGSPEC →
[RUNBOOK]
The Wrap-Up Ritual
Developers close AI coding sessions and discard the most valuable exhaust: architectural trade-offs, procedural discoveries, and broken tooling instincts.
OPERATIONAL RIGSPEC →
[PRIMITIVE]
The Always Allow Button Is Lying to You
You click Always Allow on a macOS keychain modal, yet it prompts on the next run. The button isn't broken; it edits an access list that macOS never consults.
OPERATIONAL RIGSPEC →
TRACK 03 // [AXIOM] & [MECHANIC](6 MODELS)
Whiteboard Axioms & Mechanics
"Mental models forged in 20+ years of IT that outlast the hype cycle."
[MECHANIC]
Dad, Do We Have Any Milk?
Outsourcing the first 3 seconds of observation to an LLM isn't leverage. You burn megawatts, dull your eyes, and become a spectator in your own workshop.
MENTAL MODELREAD →
[MECHANIC]
Vibe Coding & the Builder's Brain
Ranch-raised with tools, but blocked by syntax for 20 years. Here is how modern AI turned code from a barrier into a direct extension of my builder brain.
MENTAL MODELREAD →
[AXIOM]
The Enterprise Cockpit Problem
Enterprise software defaults to screenshots of fifty dials. To engineers, that signals power. To buyers training forty people, it signals pure overhead.
MENTAL MODELREAD →
[AXIOM]
Salesforce, Claude, and the Tokenized Circle of Life
Putting CRM inside Claude isn't benevolence. When browser clicks yield zero incremental revenue, headless agent usage is the new enterprise margin tollbooth.
MENTAL MODELREAD →
[AXIOM]
The Second Tax of Capture
Tools spent a decade killing capture friction with widgets and voice memos. Nobody talks about the second tax: emptying the junk drawer by hand.
MENTAL MODELREAD →
[AXIOM]
Five Whiteboard Laws of Platform Design
Five non-negotiable whiteboard principles: why self-provisioning, unit economics, and perceived speed must be locked down before writing code.
MENTAL MODELREAD →
CHRONOLOGICAL REGISTER
FULL INDEX // 15 CODIFIED ARTICLESThe Complete Dispatch Ledger
MECHANIC
Dad, Do We Have Any Milk?
Outsourcing the first 3 seconds of observation to an LLM isn't leverage. You burn megawatts, dull your eyes, and become a spectator in your own workshop.
POST-MORTEM
I Gave an AI Agent My Ad Account
An agent took our ad account from zero to a live campaign in an afternoon—and silently failed three times. Why unverified success claims kill automation.
PRIMITIVE
The 🔥 Emoji Is a Security Primitive
Guardrails behind an AI ops system at a Seattle dance studio: propose-only chat, tool refusals, and a human approval that only counts if it's the 🔥 emoji.
RUNBOOK
I Built a Claude Skill — and Benchmarked It
I packaged an IT-ops workflow (OneLogin SSO) as a Claude skill, then measured it. Here is the benchmark, the loop that made it better, and the honest caveats.
POST-MORTEM
The Agent That Refused Its Boss
I approved a $220 ad spend with a flame emoji. The orchestrator relayed my consent to the execution node. The node said no. It was completely right.
MECHANIC
Vibe Coding & the Builder's Brain
Ranch-raised with tools, but blocked by syntax for 20 years. Here is how modern AI turned code from a barrier into a direct extension of my builder brain.
AXIOM
The Enterprise Cockpit Problem
Enterprise software defaults to screenshots of fifty dials. To engineers, that signals power. To buyers training forty people, it signals pure overhead.
AXIOM
Salesforce, Claude, and the Tokenized Circle of Life
Putting CRM inside Claude isn't benevolence. When browser clicks yield zero incremental revenue, headless agent usage is the new enterprise margin tollbooth.
PRIMITIVE
"Agent-Ready" Is Not the Same as "Has an API"
We spent two days automating campaigns through an official vendor MCP server. The tool catalog wasn't a capability list—it was a corporate liability fence.
RUNBOOK
The Wrap-Up Ritual
Developers close AI coding sessions and discard the most valuable exhaust: architectural trade-offs, procedural discoveries, and broken tooling instincts.
PRIMITIVE
The Always Allow Button Is Lying to You
You click Always Allow on a macOS keychain modal, yet it prompts on the next run. The button isn't broken; it edits an access list that macOS never consults.
AXIOM
The Second Tax of Capture
Tools spent a decade killing capture friction with widgets and voice memos. Nobody talks about the second tax: emptying the junk drawer by hand.
POST-MORTEM
Your Headless Browser Is Snitching on Itself
Two days debugging a SaaS pipeline hit with SMS 2FA on every run. The problem wasn't reCAPTCHA—it was Chrome introducing itself as a headless robot.
POST-MORTEM
Your Backup Is Lying to You
Why automated backups fail silently when health checks measure activity instead of invariants, and how an oversized video file jammed an entire git history.
AXIOM
Five Whiteboard Laws of Platform Design
Five non-negotiable whiteboard principles: why self-provisioning, unit economics, and perceived speed must be locked down before writing code.
FREQUENCY: OCCASIONAL DISPATCHES
SUBSCRIBE TO DISPATCHES
Occasional deep dives on AI enablement, enterprise architecture, and vibe coding. Zero spam.