01
Architecture erosion through
tool gravity
An architecture erodes through a thousand micro-decisions unless it is actively enforced in every operational step.
Scope summary How I work with AI. An honest maturity view of my own AI working practice, tested while building sem-agentic — separate from the commercial evidence delivered by the system itself. Evidence-backed: Own maturity tracker, repeatedly re-audited. Seven documented engineering patterns from the Augur project. Dual-tool working practice (Claude Chat + Claude Code) with clear role separation. Not evidenced: Complete agentic system architecture (currently 2/20 — deliberately low). Organisation-wide scaling of this working practice. External validation of the maturity framework.
AI engineering practice
How I work with AI
An honest maturity view of my own AI working practice, tested while building sem-agentic — separate from the commercial evidence delivered by the system itself.
- Objective
-
Track my own AI engineering practice as honestly as a business metric.
sem-agentic (Augur) was the test case for examining how I work with AI tools: not only whether a system emerged, but how disciplined the path towards it was. The result is a maturity tracker, not a polished portfolio.
- Approach
-
Two tools, two roles: Claude Chat for strategy and sparring; Claude Code for architecture and implementation.
Repeated audit against a fixed maturity framework (Engineering Discipline, Knowledge & Workspace, Agentic System Architecture, Quality & Observability), with the status re-dated after every material update. Low scores are decisions, not omissions: the work implemented what proves the point, not what fills the table.
- Why this is a separate page
-
Campaign steering and an AI working practice are two different forms of evidence.
sem-agentic as a commercial case demonstrates what the system does for Google Ads campaign steering. This page demonstrates something else: how I work directly with AI tools as a leader, and what that teaches me about organisations facing the same transformation.
Evidence-backed
- Own maturity tracker, repeatedly re-audited
- Seven documented engineering patterns from the Augur project
- Dual-tool working practice (Claude Chat + Claude Code) with clear role separation
Not evidenced · out of scope
- Complete agentic system architecture (currently 2/20 — deliberately low)
- Organisation-wide scaling of this working practice
- External validation of the maturity framework
Lessons from engineering
Seven patterns from two weeks of the Augur project.
Pitfalls are not catalogued in advance; they are collected during engineering. Thirty to sixty minutes of reflection per day, consolidated into seven broader patterns. Each pattern draws on at least two independent observations.
Architecture erosion through
02
Senior attention as a
sharpening instrument
Senior sponsors do not decide the content; they ask the right question at the right points.
03
Verbal outline ≠
operational reality
What sounds clear in a plan is not yet defined for the team. Force specificity before additional work begins.
04
Evaluation discipline against
your own desired result
Without hard measures, the most recent output inspection decides. It is always coloured by the result you hoped to see.
05
Learning culture, not error culture
An organisation does not learn through workshops, but through permanent anchoring at the point where the next decision is made.
06
Pattern reuse
and the library effect
The most valuable investment is not the next tool, but a library of internal patterns treated as intellectual property.
07
Operability and
operator empathy
Marketing AI does not fail because of the model, but because it is not usable for the person who has to work with it every day.
Working organisation: how I use AI tools and track my own maturity honestly.
Status as of 1 June 2026.
working organization
Karsten
senior · approval · vp-story owner
claude chat
Strategy & Sparring
VP-Story Material
Reflection & Memory
Documents & Visuals
claude code
Review code state
Architecture outline
Develop services
ADRs & Tests
AI engineering practices · maturity
The low scores are not an omission, but a decision: the work implemented what proves the point, not what fills the table. An honest status says more than a polished one.
0 not present 1 aware 2 repeatable 3 defined 4 managed 5 optim.
A · engineering discipline 15/20
stage-1/2 build 3
pre-flight audit 4
discrepancy-driven 4
hard-fail senior 4
B · knowledge & workspace 16/30
cross-chat memory 2
conversation recall 2
audit + re-audit 3
reusable prompts 3
shared workspace 4
tiered knowledge 3
C · agentic system architecture 2/20
skills-system 1
goal-setter 1
background-routines 0
mcp-integration 0
D · quality & observability 5/20
eval-driven ci 2
observability+cost 2
a/b prompt-testing 0
self-maintenance 1
shared workspace · 15 ADRs
project.mdstatus.mdadrs/ ·15reflections/audit.mdbacklog.md
tools & platforms
claude.ai · claude code · github · localhost:5678 · browser
The Agentic Blueprint, insights from the engineering work, rendered in the
browser with PDF.js. Loaded on the first tab change (~19 MB).
Loading PDF…
The actual subject is not the code. It is what the development work teaches about leading an organisation through agentic AI: the bottleneck is not in the model, but in the hundred decisions the system returns to people. The person accountable for the outcome shapes this phase. The system can generate work. Responsibility remains human.
The corresponding commercial evidence: sem-agentic — Google Ads SEM campaign steering →