76% of PM job postings want AI skills: a visual glossary of 15 agent terms

aakashgupta · x · 2026-10-25

Aakash Gupta published a glossary of 15 terms every AI product builder should know, noting that 76% of the 113 PM job postings he surveyed now require AI knowledge — concepts that only engineers needed a year ago.

The terms are grouped into five categories:

Core: Harness (everything around the model — Anthropic's harness improvements added 6 points on Terminal-Bench 2.0), Agent Loop (plan → act → check → decide), and Tools (files, terminal, browser, MCP apps the agent can reach).

Context: Instructions File (CLAUDE.md/cursorrules/AGENTS.md — the one file to fix first), Rules (standing do-not-violate instructions, backed by hooks), Context Library (knowledge the agent reads first), and Memory (cross-session team and product preferences).

Action: Skills (playbooks loaded only when a task needs them) and Subagents (helper clones with fresh context returning short answers).

Safety: Permissions (what runs without asking — he cites PocketOS losing a production database in 9 seconds to a token made for custom domains), Sandbox (an agent ran rm -rf in a developer's home directory testing a sandbox it built on Aug 26), and Human-in-the-Loop (stopping before sending, deleting, publishing).

Quality/Build: Checks (evals, linters, second-model review kept outside the agent), Hooks (code-enforced rules the model can't skip), and Feature Harness (the instructions, data, tools and checks inside a shipped AI feature).

Related event: AI Skills Now Essential for PMs as Agent Terminology Guide Goes Viral(2 posts)→

Original post →

More from coding & agent

coding & agent channel →