A Jenn Field Guide
The Agentic Builder's Guide
How I learned to get unusually good with Claude, what actually transfers to Codex, and the operating system I'm building around both.
Journey map
From one tool to a durable operating system.
This guide is not just a manual. It is the story of how the work matured: first in Claude, then in the transferable layer, and now under Codex pressure.
Original proving ground
Claude
Where the first serious lessons on context, skills, validation, and voice were earned.
Durable layer
Transfer
The part worth keeping is the operating system behind the tool, not just the tool itself.
Current frontier
Codex
The present test: what changes when the agent gets sharper on implementation, review, and background work.
Part I
Claude was the proving ground
This is where I learned context architecture, session hygiene, skills, validation, and how to turn repeated corrections into durable instructions.
Part II
The transferable layer matters more
The real lesson is not one vendor tool. It is the craft of building a working relationship with agents, with proofs, artifacts, and feedback loops.
Part III
Codex is tightening the system
Now I am tracking what survives, what breaks, and what needs sharper workflows when the agent is stronger on implementation and bug-finding.
Learned in public
This guide is built from real projects, failures, fixes, and deliverables rather than generic best practices.
Structured around proof
Each durable lesson should eventually connect to a project artifact, a failure mode, or a visible shift in output quality.
Model-agnostic by design
Claude-specific mechanics matter, but the highest-value layer is the part that transfers to Codex and whatever comes next.
How this stays alive
A guide with a memory system behind it.
The public guide should not be a frozen manual. It should keep absorbing what I learn in projects, in session logs, and now in Codex.
1
Session log
Capture the raw story of the build: what I asked, what the agent got right, what it missed, and what I had to learn.
2
Lesson
Promote anything that repeats. Tag it as Claude-specific, Codex-specific, or transferable so the guide keeps comparative fidelity.
3
Pattern or skill
Turn stable lessons into reusable operating behavior: skills, runbooks, folder rules, validation checklists, or workflow prompts.
4
Guide evidence
Attach the lesson to projects and artifacts so the public guide can teach through proof instead of abstraction.
Where the lessons already show up
These are starter proof threads. The next step is letting agents continuously map chapters to project evidence so the guide stays specific, personal, and difficult to fake.
claude
Corrections should become loadable context, not disappear into chat history.
A failed tariff exposure analysis turned into a reusable skills system that changes outputs before the build even starts.
Shifted work from generic analysis toward action-oriented deliverables and stakeholder-aware outputs.
Skills as Loadable Contexttransfer
Durable instructions need a hierarchy and a home outside the conversation.
Running two sites with conflicting palettes forced a three-tier context system: global, project, and session.
Reduced repeated corrections and made first-pass output more reliable across separate codebases.
The CLAUDE.md Hierarchytransfer
Clean output is not the same thing as correct output.
Job-title exposure analysis failures led to proof chains, oracle checks, forbidden strings, and seam testing.
Moved review from aesthetics toward evidence, which is exactly the layer that transfers across models.
Validating AI Outputcodex
When the cost of being wrong is high, disagreement between models is a finding.
A second-opinion tool made model disagreement explicit so the review queue becomes targeted instead of vague.
Created a bridge from Claude-era workflows into a broader multi-agent, multi-model operating system.
Cross-Model Interlocutorcodex
The durable edge is not the shell. It is owning memory, skills, and review outside the vendor boundary.
The vendor-stack comparison hardened into a concrete local operating layer with one narrow loop, package-first wrappers, and explicit project records.
Turned the guide’s transferable thesis into a product direction that can survive model churn instead of reading like a comparison essay only.
Jenn OS v1