15 sessions on PowerShell with a coding agent: the agent failed at least once in 10, and 2 finished without a human stepping in.
Takeaway: Next time I'd just tell it upfront that this is a Claude Code skill repo (not a plugin marketplace) and that it needs SKILL.md copied into ~/.claude/skills/, since the agent burned several turns discovering that itself.
Takeaway: Next time I'd specify upfront that I want the split and training pipeline to preserve the conversation metadata (system/human/gpt) rather than letting the agent default to the flattened corpus.txt — that correction is what forced the whole pipeline to be rebuilt properly.
Takeaway: Next time I'd flag up front that any CLI flags removed during a refactor need a compatibility shim, since removing --recover silently broke one test fixture.
Takeaway: I'd paste the exact error message next time so the agent doesn't have to guess which line was actually broken.
Takeaway: Next time I'd ask for the stack-size and static-storage constraints to be checked up front, since two of the three crashes came from oversized structs living on the stack rather than logic bugs.
Takeaway: I'd start the first prompt by explicitly naming the root directory and the include/src split up front, since correcting those after the fact cost two early interruptions.
Takeaway: Next time I'd put the milestone list, the GitHub-vs-Codeberg cryptoscam split, and the Python-version target (3.7+) all in the first prompt instead of drip-feeding them — the agent had to backtrack on dataclasses and kw_only fields because 3.7 support came in late.
Takeaway: Next time I'd just say up front that it's a skill repo, not a plugin marketplace, since the agent burned a few steps trying the marketplace/plugin install path before checking the repo contents directly.
Takeaway: Next time I'd specify the persistent-arena/variable-registry design up front — the query-vs-clause arena confusion cost the most back-and-forth.
Takeaway: Next time I'd start by describing my actual corpus structure (the JSON with system/human/gpt roles) up front instead of letting the agent assume a flat text file — that would have saved a whole detour through split_corpus.py v1.
Takeaway: Next time I'd specify upfront which CLI flags must stay wired through even if the C API doesn't support them yet, since dropping --recover during the refactor silently broke a fixture test and cost a round of debugging.
Takeaway: I'd paste the exact error PowerShell throws next time instead of just saying 'fix this' so the agent doesn't have to guess which line is broken.
Takeaway: Next time I'd ask upfront for the stack-usage rules (no post-init malloc, bounded everything) to be double-checked against test harness patterns, since two of the three real bugs were stack overflows from large structs being copied or placed on the call stack instead of static storage.
Takeaway: Next time I'd specify the exact layout rule (like the 80/20 split) up front instead of letting the agent guess at pixel dimensions per resolution — that's the correction that actually simplified the code.
Takeaway: I'd open with the exact directory layout (include/ for headers, src/ for cpp, root at d:\Yggdrasil) since that's the first thing I had to correct.