Layer 2: Context Injection and Inference-Time Presence

Developer workspace with code editor and notes — context packs as infrastructure

PMM Mindset · May 2026 · Agent Discovery series · 3 of 4

Layer 1 makes the public truth findable. Layer 2 makes your truth inevitable at answer time — by shipping the short, versioned blocks people and tools inject before the model responds.

If you have been investing in docs, llms.txt, and evaluation-shaped content, you have been building Layer 1. Layer 2 is what stops the right answer from depending on whoever wrote the prompt that morning.

This is the layer where PMM owns approved claims inside the context window — not the long-form story, but the pack that sales, SE, and agents actually load.

What “good” looks like

-->Official rules, skills, and prompt packs — not tribal knowledge in random Slack pins.
-->Tight length — under roughly 1–2k tokens for the core pack; link out to Layer 1 for depth.
-->Explicit boundaries — when not to recommend you; when to escalate to human.
-->Two working examples — “happy path” and “known sharp edge.”
-->Changelog — treated like code review: who approved copy changes?

PMM responsibilities

-->Approved claims — single source for superlatives, numbers, customer proof.
-->Objection snippets — security, pricing motion, migration risk — honest, not clever.
-->Competitive talk track — what we concede, what we win on, what we ignore.
-->Review cadence — quarterly minimum; after every major launch.

If PMM writes Layer 2 in a vacuum, it rots the day after launch — same failure mode as Layer 1. Partner with DevRel and Eng on limits, auth flows, and error strings that must never be paraphrased.

Where this ships

-->IDE: project rules, org-wide defaults where policy allows.
-->Chat tools: shared project instructions, approved system prefaces for support and SE.
-->Starters: template repos with README plus embedded “how we use AI with this stack” for your product.

The surface is not one channel. It is every place a human or agent loads context before the model answers.

A minimal context pack (example shape)

You do not need a 10-page library. A shippable Layer 2 pack often looks like this:

-->Product in one paragraph — who it is for, what it is not.
-->Three proof points — each with a source link to Layer 1 docs.
-->Two objection handlers — security and pricing, written for inference not applause.
-->One “do not say” block — deprecated positioning, old SKU names, retired claims.
-->Escalation rule — when to stop answering and route to a human.

That is enough for a side-by-side eval: same buyer questions with and without the pack. Measure correctness and hallucination rate, not vibes.

Metrics

-->Side-by-side evals — same tasks with vs without the pack; measure correctness and hallucination rate.
-->Drift incidents — how often sales or field pastes outdated claims? Track and fix at source.
-->Adoption — percent of customer-facing technical staff on latest pack version.

Common mistakes

-->10-page “prompt libraries” nobody uses.
-->Copying Layer 1 long-form into the context window — noise replaces signal.
-->Letting security-sensitive detail live in prompts instead of procedures.

Series navigation

This post is Layer 2 in the Agent Discovery series. Layer 1 covers AI-readable docs and training-time visibility. Layer 3 covers the agentic surface — where recommendation turns into action.

Next: Layer 3 — Agentic surface (vault draft in progress).

PMM Mindset · Agent Discovery series · Track 1 · 3 of 4