Ai Workflow
How I adapt my AI coding workflow: model choices, handoffs, local experiments and the limits of what I can measure.
These are personal case studies, including experiments I stopped and preferences I changed. The retrospective explains what the records can tell me, and what they can't.
What my AI logs can actually tell me
Looking back through my workflow history showed how much activity I could count, and how much measurement I still needed.
Stopping a context optimization
A representative compaction test lost important state, so I stopped the experiment and returned to shorter work stages.
Choosing models by role
Trying Opus for orchestration and Sol for independent review, while keeping model preferences separate from measured results.
Keeping a small source of truth
Restoring an owner-controlled instruction file and separating important operating rules from generated configuration.
Checking the age of the evidence
A stale error changed a workflow diagnosis, and exposed why I need timestamps and attribution alongside usage numbers.
Defining finished for an AI agent
Building a catch-up workflow that can resume unfinished work without turning every session into another inventory.
Matching models to bounded tasks
Narrowing delegated work after reliability concerns and looking at the context around model calls.
Taking ownership of my AI instructions
Simplifying the operating instructions I wanted to understand and edit myself.
Making model routing visible
Asking for inspectable delegation history and checking that configured routes could actually answer.
Checking capacity before routing work
Revising my routing preferences after correcting an assumption about available model usage.
A local model test I decided to stop
One recorded attempt, an uneven comparison, and the decision to park further experimentation.
Measuring the cost of a finished result
Planning a small workflow comparison that counts retries and review as part of the work.
Checking what my AI usage numbers mean
An unexpected quota reset led me to examine pricing fallbacks, source freshness and the difference between estimates and spend.
Routing work with a review step
I explored task-specific model choices while keeping bounded tickets, verification and independent review in the workflow.
Keeping my agent setups consistent
Matching configuration files exposed the gap between instructions being saved and a new session actually following them.
Measuring why local AI felt slow
A slow greeting pushed me to separate model loading, instruction processing and queued work from response generation.
Writing a better handoff between AI models
I started separating planning from implementation and giving the next model a bounded change with acceptance checks.
Choosing what my AI session loads
An archived cleanup request led me to look more carefully at plugins, skill descriptions and session context.