Field Guide Toolkit
Nine parts, nine exercises. Every one of them is meant to be run, not just read: a scorecard, a calculator, an inventory, a tracker. This page bundles all nine in one place, so you can work the whole field guide in one sitting instead of hunting across nine posts.
The build-gap scorecard
Eight questions, rated on your own last six months.
1.I've personally written a skill, a rule, a CLAUDE.md entry, or any file that instructs an AI system how to behave for my work.
2.I've watched an agent work for more than thirty minutes on a task I gave it and intervened only to course-correct, not to take over.
3.I've debugged a failed agent run by reading the logs or the agent's reasoning, not by asking someone what went wrong.
4.I've written a hook or a guardrail that stops an AI system from doing something I don't want it to do.
5.I've added a Model Context Protocol (MCP) server to my own setup, or removed one deliberately.
6.I've composed multiple agents or subagents to work in parallel on independent pieces of a task.
7.I've looked at a peer's setup and could tell whether it was better than mine, or worse, and why.
8.I've explained to a direct report what an agent harness is and been confident in the answer.
Far side of the gap
The modal position for VP-and-above in April 2026. Not a judgment — it's who the field guide was written for.
The tier and failure tuner
Hear which tier a conversation is really about, then name the failure mode.
Which tier is this conversation actually about?
The product wrapped around the model: ChatGPT, Copilot, the chat box. Most conversations that say "AI" live here without anyone noticing.
Which failure mode is this? Click to log it.
The pilot-ratio calculator
Your real pilot-to-production ratio against the 88% benchmark.
Which myth has your budget?
The six-slot harness inventory
Skills, Memory, Agents, Hooks, MCP, Sensors & Gates: have vs. use.
Skills
Reusable instructions for a recurring task.
Memory
Persistent context the system reads every session.
Agents
Scoped helpers you hand a bounded task and let run.
Hooks
Deterministic rules that fire on events, the guardrails.
MCP
Servers that connect the system to your tools and data.
Sensors & gates
Checks that decide whether output is good enough to pass.
Pick this first
Skills
Name the artifact in one sentence before you start: not “I’ll build a skill,” but the actual named file and what it does.
The five-primitives mirror
Intent, Systems, Orchestration, Synthesis, Judgment: transfers or reframes.
Intent
How you set the why for your team.
Systems
How you see the feedback loops between functions.
Orchestration
How you design hand-offs.
Synthesis
How you combine disparate inputs into a call.
Judgment
How you catch work that looks right and is wrong.
The first-Saturday tracker
Install, bootstrap, rewrite, skill, guardrail, with a running minute count.
Block out a Saturday morning, or space these across a week.
The six-week ladder
Four weekly moves, weeks two through six, each with a one-line build note.
Week 2
Notice what you do twice
A cluster of 3+ repeated prompts, turned into a named skill.
Week 3
Split memory
CLAUDE.md (stable) + CLAUDE-current.md (this week), both read.
Week 4
Spawn one agent
One recurring sub-task, one job, tried twice.
Week 6
Ship something
One real piece of work, through the harness, to an actual audience.
The two-operator readiness check
The article's own three questions, answered independently and compared.
Has the two-person setup produced work neither of you could produce solo? Not faster: neither of you could.
Has a third colleague asked, unprompted, to see what you are doing?
Could you describe your shared harness in one paragraph to a peer outside your company?
The archetype audit
Personality archetype x execution archetype, natural pairing highlighted.
Your personality archetype
Your execution archetype
Follow the lab
Get the next experiment
Enjoyed the breakdown on the Field Guide Toolkit? New entries land roughly weekly. No digest, no roundup. Just the next build log, when it ships.