feelsfast.fyi

The AI cut

Building an
AI product?

AI waits are not the deterministic page-load waits the rest of the platform is tuned for. Duration spans two orders of magnitude. The shape is conversational, not navigational. The answer arrives mid-wait. Same toolbox, different problem.

Thinking…

Concepts · 2

AI-specific perceived performance in two reads

Why AI waits behave differently from web waits, and where perception engineering on these surfaces crosses into deception.

  1. AI changes the shape of the wait

    AI waits are not the deterministic page-load waits the rest of the platform is tuned for. Three properties change: duration variance across two orders of magnitude, conversational shape, and the answer arriving mid-wait. Which classical thresholds survive the move to AI, and which do not.

    Block & Zakay 1997Miller 1968Nielsen 1993Card et al. 1991Card, Moran & Newell 1983Doherty 1982Ziegler et al. 2022

  2. Honesty in AI UX

    Where AI perception engineering crosses into deception. Fake streaming, manipulated cadence, confident UX over uncertain output, hidden tool calls, cancellation theatre. The line between polish and lying — and how to stay on the right side of it.

    EizenbergBlock & Zakay 1997Anstis 2003Guo et al. 2017

Scenarios · 5

See it on every surface

Five places AI shows up in product UIs today, ordered by time-band — from sub-second autocomplete to multi-minute agentic runs. Each card carries the scenario card from /scenarios; the demo lives directly below it.

  1. Scenario

    Inline Completion

    Cursor / Copilot / v0 territory — typing-time inference. Debounced query, abort on next keystroke, optimistic accept on Tab.

    Ghost-text Completion

    User types in a notes editor; the AI suggests an ending for the current line as ghost text. Naive: fires on every keystroke, no abort. Tuned: 200 ms debounce, abort on next keystroke, Tab to accept.

    100 MS – 1 S
    Interactive

    Off

    Tab · accept suggestion

    On

    Tab · accept suggestion

  2. Scenario

    Chat Streaming Response

    AI streams tokens after a brief thinking state. Pace tokens to a natural reading rhythm; cancellation on every keystroke. (Live demo on /playground.)

    Animated Text Streaming

    A chat-style assistant returns a short answer. Naive: total wait, then the full response drops in. Tuned: ~600 ms thinking state, then tokens stream at a natural reading pace.

    1 – 10 S

    Off

    Press Run to start

    On

    Press Run to start

  3. Scenario

    Tool Execution

    Agent visibly does something (reads a file, runs a query, edits code). Tool-call transparency is the perception trick.

    Live Status Timeline

    An agent reads a file, searches the codebase, and runs a typecheck. Naive: opaque "Working…" spinner for the full duration. Tuned: each tool call streams in as the agent runs it, with its own running / done state.

    1 – 10 S

    Off

    Press Run to start

    On

    Press Run to start

  4. Scenario

    Long Compute Inference

    Long-running AI inference (image generation, complex analysis). Engagement, progress where measurable, tool-call transparency where applicable.

    Animated Text Streaming

    A chat-style assistant returns a short answer. Naive: total wait, then the full response drops in. Tuned: ~600 ms thinking state, then tokens stream at a natural reading pace.

    1 – 10 S

    Off

    Press Run to start

    On

    Press Run to start

  5. Scenario

    Agentic Workflow

    Multi-step agent that may run for minutes. The user can watch the trajectory; cancellation always available; determinate progress where the agent can estimate it.

    Multistep Status List

    A multi-step agent — plan, research, write — runs for ~12 s. Naive: "Running agent…" for the duration. Tuned: timeline of phases with current step highlighted, sub-step progress within the research phase, and a final success line.

    10 S+

    Off

    Press Run to start

    On

    Press Run to start

Skill

Bake this into your AI codegen

Once installed in your AI-assisted coding tool, the skill recognises which AI surface it is generating and ships the right pattern by default.

Install

One markdown file in Claude Code, Cursor, Codex, or any agent that reads skills. Once installed, your tool recognises the AI surface it’s generating — chat, inline completion, tool execution, long compute, agentic — and applies the right pattern by default: deliberate token cadence, motion-safe thinking state, tool-call transparency, honest cancellation. Without the skill, your AI generates loaders by vibes.

npx skills add andrzejdelgado/feelsfast-skill

Or visit the Skill page for direct-download and other-agent install commands.