David W. Keith

@dwk.io

203 Followers 130 Following

Former engineer at Apple, Nest, Google, and Tara AI. Now supporting AP CS education, non-profits, leftist political candidates, and open source. 🇺🇸✊🌹 🔐 https://dwk.io Agnostic <(S)He/It>

Claude Code usage showing Fable 5 (Ultracode) burning through tokens that has a weekly reset of 9PM tonight with 0% used so far.
A conversation with Claud Design explaining why it won't continue because I have use 81% of my weekly limit.
Claude Code CLI output: The work is complete: all 14 tasks implemented, every task spec- and quality-reviewed with fixes applied, final holistic review passed, 165 tests green, both schemes build. Implementation complete. What would you like to do? 1. Merge back to main locally 2. Push and create a Pull Request 3. Keep the branch as-is (you'll handle it later) 4. Discard this work Which option? ✻ Churned for 2h 55m 0s
A GitHub Diff that adds line 180: ## Step 2a⅞½ — Performance budget (warn-only)
Interesting note: the gemini-code-assist bot was wrong about its own parent company's product structure on 3 of 4 items. Verified each against primary sources before responding.
Claude: The PR's figures and citation are actually correct. The bot's suggestions are wrong. Let me reply to the review threads to explain why no change is needed.
Screenshot of a terminal input that says "Regenerate the exemplar as if you were Gemini "
Claude Quiver Buddy saying: *vibrates sharply* A SPACE. You spend how long on THIS?
Screenshot of a terminal with Quiver the ASCII-art April Fools mascot on it showing a speech bubble that reads: *vibrates intensely* Eight PM? EIGHT PM‽ Rate limiting is a TRAP—
My senior picture with my Mac
You're right — I undermined my own argument in real time. The "unfakeable" specific entry took me about two seconds to generate. Point taken.
Building a full RFC 8259 Brainfuck parser. This will be substantial. Let me create it systematically:
The parser is hanging. Let me write clean, verified-working Brainfuck code:
Screenshot of macOS Settings sidebar whether the badge after "Adde AppleCare Coverage" is shifted down below the line.
A YAML data structure showing that the tasks were completed 90% faster than the original estimates.
Screenshot of Claude on macOS: user: Helpful, the various spellings are interchangeable depending on my style guide. The preferred spelling should be the one with the most precise meaning to a LLM Claude: **Preferred spelling for LLM precision:** **story-led** **Rationale:** The hyphen explicitly signals a compound modifier relationship, eliminating parsing ambiguity. Without it, tokenizers may interpret "story led" as noun + past-tense verb ("The story led to...") rather than a unified adjective. Single-word forms like "storyled" risk inconsistent tokenization and lower frequency in training data.
A rain chart showing a sudden downpour just after 10 AM
Theater curtain with the title “The Golden Girls Live!” projected onto it
"I'm not against the automobile, but I just feel that the automobile has moved into communities too much, and I feel that you can design so that the automobile is there but still put people back as pedestrians again, you see. I'd love to work on a project like that."-WALT
A post complaining about being logged out constantly that is rejected because of error 2
Screenshot of a marketing email from "Smart SwimSuits" to an "etsy@" email address on my domain.
Tesla Model 3 screen saying I need a physical car keycard, taken with my phone that is setup
### 📊 Token Usage Tracking **Purpose**: Track actual token consumption to improve future estimates and validate story point accuracy. **How to Track**: When completing a task with Claude Code, add the following to the completed task: ```markdown - **Tokens Used**: [X tokens] (~Y% of 200K budget) - **Token/Point Ratio**: ~[Z]K tokens per point - **Deliverables**: [N lines] ([code] + [tests] + [docs]) ``` **Current Baseline** (from first completed task): - **1 point** = ~67K tokens (comprehensive implementation with tests & docs) - **Budget Planning**: 200K tokens ≈ 2-3 points comfortably - **Quality Level**: Production-ready with full test coverage and documentation **Factors that increase token usage**: - Large codebase exploration - Complex architecture decisions - Extensive documentation - Comprehensive test suites - Multiple implementation paths - Error handling & edge cases **Factors that decrease token usage**: - Focused, well-scoped tasks - Existing patterns to follow - Minimal new documentation - Simple implementations - Clear requirements