Appearance
Learn
Collaborator errors → distilled into guide updates. It's a step in our roadmap for improvement
Process
as we roll along, we hit a lot of bumps. i've noticed that i get fed up and stop dead. clean house. takes time. need a better triage system. Let's start with:
- list mistakes as they happen (oldest last)
- hyphen-N date title
- distill: identify pattern, write rule, add to guide
- research: better tools, clever ideas
- track for escalating need:
- fed up
- stop dead
- clean house
To distill an entry:
- Identify the pattern (what went wrong, repeatedly?)
- Write a rule (imperative, actionable)
- Add rule to the appropriate guide file
- Remove the raw entry from this file
Raw Log
- 2026-03-06 scope creep on fix: user fixed a bug (is_diagonal derived + migration). I then edited 5 .di data files that the user never asked me to touch. When the user fixes something, don't cascade changes into other files unless explicitly asked. The user's fix (migration + version bump) was self-contained — the .di files would go through migration on load. Altering data files introduced a new bug and wasted time.
- 2026-02-23 competing ideas: user knows what they're doing. Don't volunteer competing ideas or alternative approaches without being asked. If the user states a direction, follow it — don't second-guess.
- 2026-02-13 back render mode: got lost in sign conventions. Repeatedly got the "in front of" / "behind" logic wrong for back-facing faces (normals point away from camera, so signed distance signs are inverted vs front faces). Added then removed flip_depth. Core lesson: before writing code, write down the sign convention on paper — what does d>0 mean, what does d<0 mean, for this specific face orientation. Don't guess.
- 2026-02-13 back render mode: user had to say "explain, not do" and "STOP" twice. When stuck on a hard geometry problem, explain your understanding first instead of immediately coding a fix. The user knows the domain better.
Distilled
| Pattern | Rule added to |
|---|---|
| Stale reads | always.md — "Re-read before editing" |
| Wrong year/path assumptions | always.md — "Verify Before Writing" |
| Tool failure deflection | workarounds.md — "Tool Failure Recovery" |
| Incomplete rename | workflow.md — "Rename with mv, then search" |
| Project-specific swap | workflow.md — "Remove, don't swap" |
| Worktree paths | always.md — already covered in "All file paths" |
| Drifting from requirements | always.md — "Requirements Echo" |
| Ignoring shorthand | always.md — "Shorthand First" |
| Revisit as info not action | shorthand.md — revisit now includes "ask Work on?" |
| Incomplete journal | shorthand.md — journal now says "Execute ALL parts" |
| Contradicted self | always.md — "Before saying No, verify" |
| Misread exact name | always.md — "Exact names matter" |
| Checked off without testing | always.md — "Code written ≠ feature complete" |
| Asked which task first | always.md — "Task lists are priority-ordered" |
| Used Bash ls instead of Glob/Read | always.md — "Use Glob/Read, not Bash" |
| Speculated instead of observing | pitfalls.md #12 — "Observe before speculating" |
| Abbreviated code names | pitfalls.md #13 — "No abbreviations in code" |
| Misread "here" as file reference | pitfalls.md #14 — "'Here' means chat output" |
| Wrote to Claude memory dir | CLAUDE.md — already added |
| Worktree paths (repeated) | pitfalls.md #7 — already covered |
| npm instead of yarn (repeated) | always.md — already covered |
| Skipped always.md reads (repeated) | CLAUDE.md — already covered |
| Miscounted entries (stale read) | pitfalls.md #1 — already covered |
| Ignored shorthand command (repeated) | pitfalls.md #8 — already covered |
| Logged without checking existing rules | pitfalls.md #15 — already covered |
| Treated question as instruction | pitfalls.md #16 — "Questions aren't instructions" |
| Promised beyond this chat | pitfalls.md #17 — "Scope promises to this chat" |
| Trimmed rejected idea instead of restarting | pitfalls.md #18 — "Rejection means start over" |
| Treated analysis shorthand as action | pitfalls.md #19 — "Analysis shorthands are not action requests" |