Capture Caddy
Voice-guided AI prototype for live portrait photography
Custom GPT with voice mode tested real-time posing support during live shoots. Professional presence—not information access—proved to be the binding constraint. Real-time AI competes with the trust conditions it's meant to support.
-
Purpose: Test whether real-time AI support can coexist with professional presence in live creative work
-
Role: Solo practitioner: problem framing, prompt design, real-world testing, pivot decision
-
Team: Solo
-
Context: Personal exploratory prototype for part-time photography workflow
-
Duration: 2 weeks
-
Status: POC complete; pivot direction identified but unvalidated; core constraint discovered
Confidence fades between shoots with no practical fallback
Skill decay through intermittency. As a part-time photographer, weeks or months pass between shoots. Each gap erodes posing fluency—the muscle memory that makes directing subjects feel natural.
Support exists, but outside the moment of need. Reference materials are technically accessible mid-session, but using them requires stopping, searching, and visibly consulting something—transforming support into exposure.
Loss of credibility, not just flow. Hesitation mid-session. Awkward pauses. Subjects sensing uncertainty. The photographer's internal experience (lost fluency) becomes visible to the client.
Presence is the constraint, not access to information
Accessibility was assumed to be the limiting factor. The hypothesis was that accessibility was the problem. If posing support were hands-free and conversational, confidence would hold.
Professional credibility depends on uninterrupted presence. Professional credibility depends on unbroken presence. Any visible consultation signals uncertainty to the subject.
Visible consultation breaks the authority loop. When the photographer pauses to receive guidance—even hands-free—the subject registers it. Trust shifts. The dynamic changes.
Real-Time Intervention Competes with the Conditions It's Meant to Support
Real-time AI coaching fails for the same structural reason reference boards fail: it competes with presence. Solving accessibility didn't help because the intervention timing was wrong.
Performance pressure amplifies interaction cost. Environmental noise interferes with voice capture. Earbuds fall out during movement. The assistant can't be interrupted mid-speech. Long audio instructions compete with active task management—camera, subject, composition. These frictions weren't incidental—they surfaced because real-time support structurally conflicts with presence-dependent work.
Breakdown is social before it is technical. When the AI mishears, pauses, or requires correction, the awkwardness becomes visible to the subject. Fumbling with technology signals the opposite of command.
The AI doesn't need to fail technically to fail socially. Real-time support is structurally incompatible with presence-dependent work.
Why the Constraint Matters
In presence-dependent creative work, professional credibility is the fragile thing. Trust collapses faster than AI can deliver value if consultation becomes visible.
Constraint discovery precedes solution direction. Identifying where real-time AI structurally fails prevents building tools that work technically but fail socially. This constraint applies broadly: any workflow where authority depends on unbroken attention.
Intervention timing is the failure mode. Real-time support promises help exactly when you need it. But "when you need it" may be precisely when absorbing guidance competes with the conditions that make the work succeed.
This isn't fixable through better AI, faster responses, or clearer prompts. The structural incompatibility exists at the level of trust dynamics and workflow constraints—not technical capability.
Key Strategic Decisions
These decisions acted as risk gates—explicit choices that shaped what could be learned and how constraints were discovered.
- Observed: Desk testing validated technical function but couldn't surface workflow-level or social constraints.
- Decision: Test in live outdoor portrait session with real subject and performance pressure.
- Tradeoff: No experimental control; findings based on single session; couldn't isolate variables or iterate on design.
- Safety Implication: Real-world social dynamics could only be observed under authentic pressure—simulated testing would have missed the binding constraint entirely.
- Observed: Existing solutions (phone, paper) required stopping and visibly consulting—making lost fluency obvious to the subject.
- Decision: Use Custom GPT voice mode for hands-free, real-time direction rather than phone-based reference.
- Tradeoff: Introduced latency, mishearing, noise interference, no mid-response interruption; couldn't be refined iteratively.
- Trust Implication: Made AI friction audible and visible to subject; exposed social failure that silent text reference would have masked.
- Observed: Earbuds created private channel between photographer and AI; subject was excluded from the interaction.
- Decision: Switch to speaker mode so subject could hear all AI direction.
- Tradeoff: Lost audio privacy; ambient noise interfered more; AI friction and corrections became visible to the subject.
- Trust Implication: Exposed awkwardness but prevented secret consultation; made social failure observable rather than hidden behind private audio.
Impact At a Glance
Single live session revealed workflow-level constraints that desk testing missed.
Quantitative Signals
- n=1 live outdoor portrait session with real subject and performance pressure
- Real-time intervention tested under authentic conditions (ambient noise, movement, subject presence)
- Pivot direction surfaced but unvalidated beyond initial testing
This was discovery-oriented prototyping, not adoption validation. No iterative refinement or multi-session validation. No baseline comparison. No controlled testing of the pivot direction. The value is the constraint surfaced — not a shipped product or proven solution.
Qualitative Impact
Testing revealed that AI friction (latency, mishearing, corrections) became visible to the subject, undermining professional presence. The tool worked technically but failed socially — validating where real-time AI intervention structurally conflicts with trust-dependent workflows.
How This Sharpened My Judgment
These lessons generalized beyond Anne-bot. They've shaped how I approach AI tooling, organizational readiness, and scope decisions in subsequent work.
Lessons from Real-World Use
- Context is the primary constraint on AI relevance in judgment-heavy domains. I designed around missing organizational knowledge by making context-gathering interactive. Human-in-the-loop became the architecture.
- Scope reduction can be more impactful than comprehensive solutions. Full PRD feedback created noise. Problem-statement focus created signal with fewer context dependencies.
- Acknowledging uncertainty produces more trusted guidance. Iteration 2 succeeded because the assistant admitted what it didn't know instead of guessing.
- The leverage point is often upstream of where you're building. The real constraint was shared criteria, not feedback speed. A foundational problem, not a tooling problem.
- Testing can reveal what not to build, not just what to build. The prototype's failures taught me that comprehensive coverage was strategically wrong.
A Transferable Pattern:
In judgment-sensitive workflows, context (not model capability) is the primary constraint on AI relevance. When organizational knowledge isn't available in the system, make context-gathering part of the interface: explicit checkpoints where AI surfaces gaps and humans provide missing information.
Simultaneously, scope your intervention to the constraint with the fewest context dependencies and verify whether foundational alignment would solve the problem more efficiently than automation.