Insights
Founder and builder perspectives on AI tools, thinking patterns, and the new way of working
Showing 349-360 of 485
30-Day Retention and the ZDR Cliff: What Fable 5 Means for Enterprise Skill Deployments
Fable 5 requires 30-day data retention and is not available under zero-data-retention. A ZDR org gets a 400 on every request. Here's the compliance decision guide for teams shipping skills.
Designing for Interruptions and Backchannels: UX Patterns for Full-Duplex Voice Skills
GPT-Live can say 'mhmm', stay quiet while you think, and let you cut in mid-sentence. Turn that full-duplex behavior into concrete UX rules and a checklist for your own voice-agent skills.
Guardrails for Voice: Impersonation Limits, Safety, and Compliance in Voice Skills
OpenAI's GPT-Live ships predefined voices only, blocks cloning, and builds in teen safety. Translate those safeguards into a concrete publishing checklist for responsible voice skills.
The $10/$50 Question: Cost-Aware Routing Between Fable 5 and Opus 4.8
Fable 5 costs roughly 2x Opus 4.8. That doesn't mean run everything on the cheaper model. Here's a routing heuristic for deciding which turns deserve the premium tier.
The 25% Latency Cut: Why Caching and Model Choice Decide If a Voice Skill Feels Human
gpt-realtime-2.1 cut p95 voice latency by 25%+ via caching. Here's the latency budget of a voice turn, the cached-vs-fresh audio economics, and when to reach for the mini variant.
Give Your Agent a Memory File: Persistent-Lesson Skills for Long-Horizon Runs
Fable 5 performs strongly with a markdown memory system. Here's the exact format Anthropic recommends, how to package it as a reusable skill, and why one lesson per file beats a growing scratch log.
Parallel Subagents by Default: Orchestration Patterns Fable 5 Rewards
Fable 5 dispatches parallel subagents more readily than older models. Spawn-and-block leaves that capability on the table. Here are the orchestration patterns that actually pay off.
Minutes-Long Turns Break Your Harness: Designing Async Agents for Fable 5
Fable 5 turns can run for minutes and autonomous runs for hours. If your harness assumes sub-30-second replies, it will time out mid-thought. Here's how to redesign for long turns.
Effort Is the New Dial: Routing Skills and Subagents from low to max
Fable 5 replaces the token budget with five effort levels. Here's how to route skills and subagents — routine work at low-medium, long-horizon at xhigh-max — plus what Claude Code's ultracode really is.
Don't Tell Claude to Show Its Work: The reasoning_extraction Trap in Skill Prompts
On Fable 5, asking the model to explain its reasoning or show its work can be refused with category reasoning_extraction and elevates fallbacks. Read the summarized thinking blocks instead.
Building Fallback-Aware Agents: Handling Fable 5's Refusal Stop Reason
Fable 5 refusals arrive as HTTP 200 with stop_reason refusal — not an error. Check it before reading content, and wire a server-side or client-side fallback to Opus 4.8. Code and a checklist inside.
Your Claude Code Skills Are Too Prescriptive for Fable 5
Anthropic says skill prompts written for older models are often too prescriptive and degrade Fable 5's output. Here's how to audit yours and swap step-by-step scripts for goal-and-constraints.