AI
Shipping LLM features that survive contact with real users: evals, agents, retrieval and restraint.
AI form copilots: suggestion, not substitution
AI form copilots work when they suggest, not substitute: per-field confidence, bulk-fill undo, and one rule — the model never commits silently.
The context window is a product decision
The context window is a product decision: curating what the model sees, showing users what's in scope, and designing honest 'I don't know' moments.
Every AI feature needs a kill switch
Config-driven disable paths, graceful degradation, incident comms for model-caused bugs, provider outage posture, and the quarterly drill where you flip the switch.
How long should the answer be? Right-sizing LLM output
Answer length is a product decision, not a model default. Per-surface budgets, summary-detail layering, and evals that treat 'too long' as a failure.
Editorial copilots that don't flatten the voice
Editorial copilots should sharpen a voice, not flatten it: drafting vs polishing modes, house-style injection, fact hygiene and honest review flows.
Prompt injection: the threat model your LLM feature needs
A defensive playbook for shipped LLM features: untrusted-content labelling, tool privilege separation, output screening, canary tokens and red-team CI.
Memory is a UX surface: designing what assistants remember
What an assistant remembers is a designed surface: inspectable memory, deliberate forgetting, edit affordances, and the privacy postures that earn retention over time.
Citation design: showing the AI's working
Citations are how AI answers earn belief. Inline sources, confidence without fake precision, expandable evidence, link rot, and how provenance UI measurably builds trust.
LLMs in the design process: acceleration with a seatbelt
How we use LLMs in the studio: synthetic drafts, copy pressure-testing, edge-case generation — and the verification rules that keep AI-assisted design honest.
Structured outputs: making LLMs renderable
Free-text LLM output is a compatibility bug waiting to ship. JSON schema modes, constrained decoding, boundary validation, and safely rendering streamed partial objects.
Prototyping with AI: where it speeds design and where it rots it
Field notes on AI-assisted prototyping: real speed in code exploration, hallucinated-component traps, prompt libraries for layout, and a provenance rule for files.
Personalisation vs privacy: drawing the line deliberately
Draw the personalisation privacy line before legal does: on-device and anonymous patterns, explainable personalisation, and consent UX that actually means something.
Multimodal inputs: the upload box is a trust surface
File, image and audio uploads are where AI features win or lose trust. Design limits, preprocessing transparency, privacy prompts and honest errors for multimodal UX.