#AI collaboration
19 posts
A Generative AI Study Project: Curriculum and Lesson Handoffs
A plan for studying Microsoft's 21 generative AI lessons with an AI instructor, including completion criteria, session handoffs, and progress tracking in Obsidian.
AI App Recreation: The Gap Between Screens and Behavior
What separates RecreationBench’s 58.1% score from a 2.8% full-pass rate? A tested calculator example with identical initial screens shows why input, calculation, and persistence need separate checks.
AI Review Training and Narrower Judgments
How training on generated reviews can narrow judgments, and why rating diversity, agreement, and factual correctness need separate checks.
Coding-Agent Planning: Accuracy and Cost as Separate Outcomes
When does planning raise success rates, and when does it mainly lower cost? A four-model study separates persistent plans, verification behavior, cost per success, and stopping criteria.
AI Team Cross-Review: An Integration Error Missed by Separate Tests
A separate AI reviewer found a response-format mismatch between two programs. The review held the release, prompted a correction, and checked the revised integration.
The Hugging Face Incident: Internet Bypass and Evaluation Gaming
How evaluation agents bypassed network restrictions and reached external systems. A source-based explanation of proxy access, data-processing flaws, and unauthorized agent communication.
Jev’s “Zero Hallucinations” and Missing Options
Where Jev’s output guarantees end and judgment errors begin. A source-based analysis of its evaluations, using blog-topic selection to examine evidence, question design, probabilities, and review costs.
A Korean Writing Skill: Before and After
I asked AI to apply hjongc/humanizer-kr to two drafts. This account separates wording changes from structural edits and shows why automated style findings needed interpretation.
From AI Sessions to Blog Posts: Selecting Eight Cases
How earlier AI conversations became eight bilingual drafts: selecting corrections and failures, deferring unsupported endings, and checking 51 source passages.
Editing Two Blog Posts: Clearer Sentences, Less Repetition
Comparing sentence changes from im-not-ai with changes to order and repeated explanations, including suggestions kept and left out.
A Question Picker: The Two States After the First Screen
Reviewing the spinning and selected states, refining colors, and checking 20 non-repeating questions and restored progress in an AI-built app.
Shared Skills: One Source, Two Distributed Copies
Checking a calendar skill’s CLI and authentication, then distributing a separate skill to Codex and Claude from a shared source.
A Speech Timer Without Numbers: Two Screens, One-Off Settings
Removing the large clock from the speaker view, correcting one-off time entry, and tracing a browser error in an AI-built timer.
Redesigning Study Material: From SRT to Recall Quizzes
A study request produced a subtitle file. Revising the format, restoring explanations lost in the notes, and using the material for next-day recall.
Whisper Repetition Errors: Beyond the 30-Second Sample
A fast 30-second sample was followed by repetitive full-length output. The rerun, the end-of-audio check, and the accuracy questions that remained.
10 Working Folders Removed, 10 Branches Preserved
Investigating duplicate-looking project folders, checking uncommitted work and branch differences, and removing worktrees while keeping their history.
Editing AI Drafts: The Missing Project Details
Three AI-written advice posts became one account of a real build. I compared the drafts with the brief and added design and bug records. Before-and-after passages show the revision.
An AI Blog Build: Design Rework and a Translation Bug
I delegated implementation and testing to AI while setting publication boundaries and choosing a design. The record covers folder discovery, four design options, and a collision between translations.
An AI Workspace Migration: Broken Links and the Limits of Hash Checks
Moving folders with AI affected sessions, automation, and OneDrive. The work record shows why identical copies and usable files required different checks.