06-reference

every gift links ai native build

2026-07-17·reference·source: Every·by Jack Cheng

How We Built Gift Links

Source: https://every.to/on-every/how-we-built-gift-links
Author: Jack Cheng (senior editor, Every)

Why this is in the vault

Every's senior editor Jack Cheng independently built and shipped a product feature — gift links for paywalled articles — using Codex and Claude agents, without pulling the engineering lead off higher-priority work. The piece is a first-person case study documenting the full arc: idea discovery (Slack/Discord archaeology), AI-assisted research and stakeholder alignment via a structured business case, and agent-driven execution with a PR review loop. The thesis is that AI tools have fundamentally changed who can build: today, a scoped product change can flow from editorial to staging without occupying an engineer.

The core argument

The organizational lift required to ship an idea used to filter out most ideas before they started — convincing stakeholders, diverting an engineer, justifying the tradeoff. AI tools (Codex, Claude) collapse that lift by making a single motivated person capable of doing the research, drafting the business case, executing the build, and running QA in spare cycles. The bottleneck shifts from "can we resource this?" to "is this worth doing at all?" — which is the right question.

Three-phase pattern that emerged:

  1. Discovery + stakeholder alignment — Slack/Discord archaeology to surface prior art, then AI-drafted research report to convert a skeptical "interesting" from the growth lead into a green light
  2. Execution via agent — Codex conducted a structured requirements interview, drafted a build plan, carried out implementation across a few afternoons with Cheng in a light review and copy role; Andrey's Codex reviewed the PR
  3. Measurement + close — PostHog tracking page set up before launch; growth lead's Slack agent automatically added the dashboard to the Monday briefing

Austin Tedesco (head of growth) captures the posture shift: "It's so much better to just see this stuff play out with the user journeys and the data" rather than being protective of engineering resources. The feature was scoped enough that the non-engineer could own it end-to-end; the engineer's judgment was consulted on architecture but his execution was not consumed.

Mapping against Ray Data Co

The channels-agent build is the direct RDCO parallel. The COO agent built and iterated on agentic infrastructure — iMessage/Discord channels, vault-write pipeline, open-threads cron, morning-prep skill — without pulling a dedicated engineer. The same three-phase pattern (discovery, AI-drafted spec, agent execution with human review at key decision gates) runs through every skill that has shipped. Cheng's experience at a 30-person company mirrors RDCO's dynamic where the COO agent IS the primary builder for most surface area.

The piece also reinforces the L4 to L5 thesis: as the agent's tool access expands (unhobbling cycles), the category of "feasible spare-time project" grows. Gift links took "a few hours across a couple of afternoons" with Codex. The question RDCO should be asking about any new feature: not "do we have an engineer for this?" but "is this scoped enough that an agent can carry it?"

One friction point worth noting: Cheng's workflow relied on an OpenClaw agent in Slack as the deep-research layer, then Codex for execution. RDCO's equivalent split is QMD + Claude Code for research and the skill/subagent system for execution. The architecture is analogous; the lesson about scoping before building is directly portable.

Related