Chapter
Common failure patterns — catch them before they cost you
The coach plugin catches these in real time. The table is the fast index; each "Tip T.N" links to the full Instead/Prefer in the guide.
| The habit | The fix | Tip |
|---|---|---|
| Handing the agent a file list ("create these 15 files…") | Hand over the outcome + constraints, not a file list | Tip 1.1 |
| Writing a vague ask ("clean up", "improve", "make it work") | Be specific — name the symbol/file/change | Tip 1.2 |
| Phrasing it as "don't do X" | Say what to do, not what to avoid | Tip 1.3 |
| Giving a rule with no reason | Add the why — motivation makes it generalize | Tip 1.4 |
| Describing a bug or layout issue in prose | Show, don't tell — paste the screenshot + @Component.tsx; pipe raw logs straight in |
Tip 1.8 |
| Begging "think hard / ultrathink" everywhere | Dial effort instead; reserve ultrathink for one hard turn | Tip 1.13 |
| Firing a one-liner and steering the whole task turn by turn | Say it all in your first message — one complete brief beats ten corrections | Tip 1.14 |
| Jumping straight to code in unfamiliar areas | Investigate first (read-only), then edit | Tip 2.1 |
| Letting the agent touch many files unsupervised | Force an approval checkpoint + blast radius | Tip 2.3 |
| Asking for one big feature in one pass | Slice vertically — thin end-to-end increments | Tip 2.5 |
| Dumping the whole repo / huge context | Feed high-signal context, not everything | Tip 3.1 |
| Keeping secrets/keys in the repo, or pasting a key | Keep secrets out of git and context (.env + gitignore) |
Tip 3.2 |
| Re-fixing the same bug over and over in one session | /clear and rewrite the opening prompt |
Tip 3.3 |
| Re-explaining the same conventions each session | Put them in CLAUDE.md | Tip 3.5 |
| Working with no checkable "done" | Make the Definition of Done executable | Tip 4.1 |
| Implementing without a failing test first | Do TDD — failing test first; don't edit tests to pass | Tip 4.2 |
| Accepting "done, it works" with no proof | Demand evidence — the test command + its output | Tip 4.5 |
| Asking for a "conservative" review | Ask for all findings, severity-labeled — a conservative prompt suppresses real ones | Tip 4.6 |
| Running long work that loses progress | Commit every green step — checkpoints to revert to | Tip 5.1 |
| Putting one agent on a huge audit/refactor | Let it self-orchestrate parallel subagents | Tip 6.1 |