Half my AI sessions start wrong — and that's fine
TL;DR
Across 128 auto-scored Claude Code sessions, 68 had a "wrong_approach" friction — 53%. The agent's first instinct is wrong half the time. This isn't a problem to fix, it's a pattern to integrate into the workflow.
Across 128 coding agent sessions analyzed automatically (post-session model scoring), 68 were flagged with a “wrong_approach” friction. 53%. Half the time, the agent heads in the wrong direction on the first try.
What the data shows
Claude Code’s post-session scoring classifies frictions into categories. Worth clarifying: even during my OpenCode/Kilo Code period, Claude Code kept running in parallel — that’s where this data comes from. Over several months of daily use:
wrong_approach 68 occurrences (53%)
buggy_code 33 occurrences (26%)
excessive_changes 8 occurrences (6%)
misunderstood_req 7 occurrences (5%)“wrong_approach” dominates. The agent understands the request but picks an implementation strategy that doesn’t fit — a pattern that doesn’t match the project context, an unnecessary abstraction, a refactor too ambitious for the scope.
Why this is normal
The agent has no memory of the project between sessions. It discovers context fresh every time. Its first instinct is based on the most common patterns in its training data — not on your codebase’s specific conventions.
It’s exactly what happens when a skilled developer joins a new project: the first few days, they propose “standard” solutions that don’t fit local conventions. The difference is that the agent does this every session.
What this changes in the workflow
Knowing that the first attempt has a 50% chance of being a wrong_approach changes how you structure sessions:
Brief constraints, not just the task. “Don’t modify the public interface”, “Use the existing helpers in AbstractServiceTest” — negative constraints reduce wrong_approach more effectively than detailed instructions.
Cut early. If after 2-3 exchanges the approach isn’t converging, opening a new session with a reformulated prompt is more effective than course-correcting mid-stream. The context rot of a derailed session is hard to recover from.
Don’t interpret wrong_approach as failure. 62 of 128 scored sessions are “fully achieved” despite frictions. The agent corrects course — as long as you guide it.
The real signal
The number that matters isn’t the wrong_approach rate — it’s the fully_achieved rate: 48%. Nearly half of all sessions fully reach their goal. 35% partially achieve it. Only 2 sessions out of 128 are complete failures.
The agent often starts wrong but usually finds its way. The efficient workflow isn’t about eliminating the initial error — it’s about detecting it fast and pivoting.
It’s also the number I show first to someone starting out with agents — so they don’t read their first wrong_approach as a sign they’re doing something wrong.
Guide: Agentic workflow principles Guide: Prompt engineering for coding agents