Cogloop

Inspiration I took AP Psychology this year and something clicked. I started understanding why I was inconsistent, unmotivated, and struggling to follow through — not because I was lazy, but because I never had the time or structure to actually think. Most people don't. We're reactive by default.

I realized that self-awareness isn't a personality trait — it's a skill, and it has real barriers. You can't reflect on a thought you didn't notice you were having. You can't name feelings you don't have vocabulary for. And insights evaporate before you can act on them. I wanted to build something that removed those barriers without removing the thinking itself.

Part of it was personal too. I've struggled with a lack of executive function my whole life, and AP Psych gave me the first framework to actually understand why. That self-compassion — understanding myself instead of blaming myself — is what I want to give other people.

What it does CogLoop is a daily AI-coached loop: Plan → Track → Reflect → Revise.

Most people think of these as separate activities. CogLoop treats them as one connected system — because that's what they are. Your morning plan should shape how you execute. What happens during execution should feed your reflection. What you reflect on should revise tomorrow's plan. Break any link and the whole system collapses back into good intentions.

Morning Clarity: You speak or type a messy brain dump — no formatting, no filter. The AI doesn't immediately organize it. It first reflects back the tension it hears underneath everything, asks if that's right, and only then builds a ranked plan from your own words. Your reasons come from you. The AI gives them structure.

Execution: A minimal single-task workspace showing your current priority, its subtasks, and your morning "why" — persistent at the bottom of the screen throughout the day. When you drift, you tap in. The app asks how your focus is — you self-assess first. Then it offers one small intervention, resurfaces your own focus rule and morning reason, and asks if you want the moment noted for tonight. Nothing is logged without your explicit consent.

Night Reflection: Built from your real drift log — not memory. The AI surfaces one pattern neutrally, proposes one rule for tomorrow while the insight is warm, and hands a draft plan into the next morning. You never start blank, and nothing is forced.

How I built it Frontend: Vanilla HTML, CSS, and JavaScript — a single interactive demo with three phases, screen-by-screen navigation, live typewriter transcript animation, and modal overlays for drift check-ins.

Backend: Node.js and Express with three real API endpoints that call the Anthropic Claude API (claude-sonnet-4-6) with carefully constrained system prompts per phase.

Voice: Web Speech API for live transcription, with Deepgram for higher-accuracy conversational voice input.

Storage: IndexedDB for local-first storage of the drift log and plan carryover — the memory that makes the night reflection grounded in real data rather than reconstructed memory.

The system prompts in src/prompts/ are the real product IP. They enforce the core principles: reflect before organizing, one intervention not ten, user self-assesses before the app shares what it observed, and consent before any logging.

Challenges I ran into The hardest challenge was getting the AI to do less. Claude's natural personality is warm, diplomatic, and generous with advice — exactly what I needed to dial back. Every system prompt had to explicitly constrain it: reasons must come from the user's own words, the AI cannot fabricate motivation, interventions are always optional. Getting that balance between helpful and overbearing took real iteration on prompt design.

Integrating Deepgram for conversational voice was also non-trivial. I wanted it to do more than transcribe — specifically to recognize when someone trails off mid-thought and gently nudge them to continue. Getting that interrupt timing to feel natural rather than intrusive was a real design challenge.

The third challenge was knowing where to draw the line between AI guidance and user autonomy — especially for a product that touches mental habits and self-perception. That line shaped every decision.

Accomplishments that I'm proud of What I'm most proud of is turning a vague, oversaturated problem space into something specific. I spent a lot of time with PKMs, Second Brain tools, journaling apps, and AI assistants — and I kept finding the same gap. Not that these tools were bad, but that they each missed something important and none of them talked to each other.

What most solutions miss is one of two things. Either they're fragmented — great at one thing (planning, or reflection, or note-taking) but disconnected from the others, so the insight you had at 11pm never makes it into tomorrow morning's plan. Or they're overinvolved — they tell you what to prioritize, generate your reasons for you, fill in your reflection before you've had a chance to think. Both patterns quietly undermine the user's own thinking over time.

What I wanted to build was something that integrated planning, self-awareness, real-time problem solving, and reflection into one connected system — so the user could tackle their life like a scientist, not a lab rat. Observing their own patterns, running small experiments, updating their approach based on real evidence. The tool holds the structure; the thinking stays theirs.

I'm also proud that this is my first hackathon, as an incoming freshman with limited dev experience, and I was able to build something with a real philosophical foundation — not just features.

What I learned The biggest thing I learned: a clear product vision is more valuable than technical execution speed. Knowing exactly what problem you're solving, for whom, and why your solution is structurally different from everything else — that clarity propelled every hard moment when the code wasn't working or time was running out.

I also learned that the most important AI design question isn't "what can the AI do?" It's "what should the AI not do?" The constraints matter more than the capabilities.

What's next for CogLoop Weekly and project-based reflection — the same loop at longer timescales, for tracking progress on bigger goals over time.

Calibration over time — showing users where their self-assessment was wrong. You predicted you'd finish the essay; you drifted for an hour. You dreaded the investor call; it went great. A tool with longitudinal memory can surface that miscalibration. You can't see it alone.

User research before deployment — the philosophy only works if it actually helps real people think. That comes first.

Built With

• claudeapi
• deepgramapi
• html/css
• indexeddb
• javascript
• webspeechapi