Capability is extended by handoff design.
The core question is not whether a model can generate text. It is which human function moved into the system and what new scaffolding made that move tolerable.
Extended Capability Library / field evidence
Applied AI Handoff Atlas
Curiosity Builds read as evidence: small transfers of judgment, memory, practice, access, and representation into AI-shaped systems. The point is to make the why work legible: what changed, what broke, and what pattern should compound into better applied AI engineering.
What the atlas argues
In this vocabulary, a handoff is a practical transfer: the system now carries a bit of judgment, memory, access control, explanation, or representation that a person previously carried informally.
Evaluation starts with the work being handed off.
Pronunciation, scheduling, tutoring, memorial search, and resume writing need different evidence. A single "AI works" claim is too broad to be meaningful.
Trust often comes from friction.
Staff confirmation, student approval, visible provenance, opt-in recipe generation, and "about, not as" archive boundaries are not limitations around the product. They are part of the product.
Small builds expose general patterns.
A family lending tool, a baseball game, and a student resume builder all reveal the same applied AI issue: systems encode authority, memory, standards, and disagreement.
The reusable lens
Every build below is read as a lesson in the extension around the model. That keeps the claims honest: when evidence is thin, the page says what artifact still needs to be preserved.
Input legibility
Can the user express the real situation before generation?
Grounding context
What does the system know, remember, cite, or exclude?
Validation
Who or what decides whether the output is acceptable?
Repairability
Can a wrong output be traced, corrected, and prevented next time?
Ratification
Who has to approve, own, or stand behind the result?
Build notes
These are not polished case studies yet. They are claim-bearing evidence notes, optimized to show what each build teaches and what proof still needs to be saved.
Wild Reader
early literacy
- Handoff
- A parent-set AI image gate became a visible product rule: five correct reading answers earns one image prompt.
- Evidence
- Public Hypandra handoff note; live build; screenshot; child feedback on mobile experience and dictation.
- Lesson
- The reward is not decoration. It teaches persistence, controls access, and changes who has to say no.
- Library claim
- Reward structure and interface harnesses extend capability before generation starts.
Evidence gap: Preserve short clips or screenshots of the reward loop and the dictation flow.
Pronouncle
speech practice
- Handoff
- Speech-to-text was asked to judge pronunciation rather than merely transcribe it.
- Evidence
- Public Hypandra handoff note; screenshot; design notes on STT limitations.
- Lesson
- A model optimized to understand users can be the wrong evaluator for a system that needs to notice errors.
- Library claim
- Validation is constitutive: if the measurement instrument is wrong, the AI capability claim is wrong.
Evidence gap: Save comparison cases: single words, full sentences, confidence scores, and human judgments.
Popmom's Lending Library
family inventory
- Handoff
- A loose family memory practice became a shared record of cars, loans, and responsibility.
- Evidence
- Public Hypandra handoff note; screenshot; story of naming and photographing cars with a child.
- Lesson
- A record is an authority move. The system does not just remember; it changes whose memory wins.
- Library claim
- Social and institutional harnesses matter even in small tools: ratification is part of function.
Evidence gap: Capture the before/after decision points: who updates the record, who checks it, and when disagreement happens.
Saw The Ball
sports judgment
- Handoff
- Baseball positioning advice became a timed, graded decision surface.
- Evidence
- Public Hypandra handoff note; screenshot; build with a 10-year-old player.
- Lesson
- Domain judgment needs contestability. A system answer may conflict with a coach, context, or local philosophy.
- Library claim
- Task structure extends capability, but it can also hide disagreement by forcing one answer mode.
Evidence gap: Add examples with alternate coach rationales and post-click explanations.
Scheduler Mark
constraint solving
- Handoff
- Several models solved the same messy schedule, then judged each other without a ground-truth checker.
- Evidence
- Public Hypandra handoff note; screenshot; benchmark-style model comparison output.
- Lesson
- AI critique surfaces disagreement, but disagreement is not evaluation without ground truth or a rubric.
- Library claim
- Rubric design and offline evaluability decide whether a benchmark produces evidence or theater.
Evidence gap: Preserve fixtures, model outputs, judge outputs, and a human-scored answer key.
TryAgBuddy
small agriculture planning
- Handoff
- The tool remembers farm context so the user does not have to restate setup, goals, and constraints each session.
- Evidence
- Public Hypandra handoff note; screenshot; context-update workflow.
- Lesson
- Persistent context helps only when users can inspect, accept, reject, and repair what the system remembers.
- Library claim
- Grounding context and repairability extend capability more than generic chat does.
Evidence gap: Save examples of accepted and rejected memory updates.
Stop Motion Planner
creative production
- Handoff
- A rough film idea became scenes, shot lists, and production notes.
- Evidence
- Public Hypandra example and handoff note; screenshot; build with a young creator.
- Lesson
- A handoff attempt can reveal missing workflow knowledge: what the creator actually plans, films, edits, and ignores.
- Library claim
- Input formation is first-mile capability work, not a cosmetic prompt-writing problem.
Evidence gap: Collect before/after scripts and notes from an actual production attempt.
Spell Better Now
adaptive practice
- Handoff
- Spelling practice became small rounds with focused words and immediate letter-level feedback.
- Evidence
- Public Hypandra example; screenshot; homeschooling context.
- Lesson
- For skill practice, the useful AI move is often narrowing the loop, not making the lesson larger.
- Library claim
- Feedback latency and task grain are capability extensions.
Evidence gap: Preserve student session traces and examples of adaptation decisions.
Curious Fractions
math reasoning
- Handoff
- Recipe scaling became a reasoning surface with estimates, hints, and AI coaching.
- Evidence
- Public Hypandra example; screenshot; teacher context.
- Lesson
- The AI value is in eliciting and responding to reasoning, not in giving the final fraction.
- Library claim
- Input legibility and diagnostic feedback are what make tutoring claims credible.
Evidence gap: Save examples of student reasoning, hints shown, and teacher review.
OnceLook
memory and observation
- Handoff
- The system generates questions about what a user noticed after seeing material once.
- Evidence
- Public Hypandra example; screenshot; family build context.
- Lesson
- Generated questions shape attention. The system is not just checking memory; it is defining what mattered.
- Library claim
- Validation instruments carry values and task definitions.
Evidence gap: Create a calibration set showing question difficulty, relevance, and ambiguity.
Kieran's Algebra Game
algebra learning
- Handoff
- Emoji and scale puzzles introduce algebraic structure before revealing formal notation.
- Evidence
- Public Hypandra example; screenshot; build with a 10-year-old.
- Lesson
- Explanation timing matters. Reveal the formal system too early and you lose the bridge; reveal too late and transfer suffers.
- Library claim
- Interface sequencing is part of the capability, not merely presentation.
Evidence gap: Capture puzzle attempts and whether players transfer to equations.
Rippled Echoes
history learning
- Handoff
- A learner can meet a generated historical perspective and connect it to the present.
- Evidence
- Public Hypandra example; screenshot; history-learning prototype.
- Lesson
- Vivid generated perspective can make uncertain reconstruction feel more authoritative than the evidence supports.
- Library claim
- Source grounding and disclosure are capability boundaries in interpretive domains.
Evidence gap: Add citations, uncertainty labels, and examples where sources disagree.
HighSchoolResumes.com
student career prep
- Handoff
- Students move from interview answers to resume bullets and opportunity searches while AI contributions remain inspectable.
- Evidence
- Hypandra blog post; README; AI transparency page; prompt provenance screenshots; live site.
- Lesson
- Youth-facing AI writing needs elicitation, choice, provenance, and explicit approval rather than invisible rewriting.
- Library claim
- Transparency is a working interface, not a policy footer.
Evidence gap: Preserve anonymized student flows, teacher feedback, and examples of rejected suggestions.
Public screenshot pending
Bunny Biscuits
home inventory and recipes
- Handoff
- A pantry inventory can be updated locally, by voice, and only deliberately sent into recipe generation.
- Evidence
- Local repository changelog and roadmap; local-first inventory design; opt-in AI recipe change.
- Lesson
- The strongest AI design move was friction: the user has to articulate intent before the model sees inventory.
- Library claim
- Reversibility, consent, and local data boundaries are capability extensions in everyday tools.
Evidence gap: Add public screenshots, a data-flow diagram, and a before/after of the opt-in recipe change.
Public screenshot pending
Zanna Smith Archive / MelTemp
memorial archive
- Handoff
- An archive guide searches and synthesizes materials while speaking about Zanna, not as Zanna.
- Evidence
- Local HAIP audit; repository prompts; archive ingestion materials; public memorial-site work.
- Lesson
- Representation boundaries are product requirements. The guide can help visitors explore, but it must not impersonate the person memorialized.
- Library claim
- Social ratification, disclosure, and source grounding define capability in memory-sensitive domains.
Evidence gap: Make citation behavior more visible and preserve audited examples against source materials.
Applied AI engineering read
The strongest claim this corpus supports is not "I built AI apps." It is narrower and more useful: I repeatedly embedded AI into real human workflows, found where naive automation failed, and then treated harnesses, evals, feedback loops, data boundaries, and human ratification as part of the system.