Why AI Prompt Packs Fail PMs (and What Works Instead)

The prompt pack graveyard

Every PM has a version of this file somewhere: a Notion page, a downloaded PDF, a bookmarked Gist — a list of AI prompts for product managers, collected over a few weekends of LinkedIn scrolling. "Prompt for writing a PRD." "Prompt for a competitive analysis." "Prompt for turning research notes into insights." You used maybe three of them more than once. The rest are digital dust.

This isn't a discipline problem. It's not that you didn't try hard enough to build the habit. Prompt packs fail for a structural reason that no amount of willpower fixes: a prompt is a question, and a chat window has no memory of the answer. You paste the same "write me a PRD" prompt today that you pasted three months ago, and the model still doesn't know your product, your users, or what "done" looks like at your company. So it invents. You spend the next forty minutes correcting fiction instead of saving the time the prompt promised.

A prompt pack optimizes the question. It does nothing about the model's total lack of an answer to work from.

Why "just write a better prompt" doesn't fix it

The natural response is to make the prompt do more work: add more instructions, more examples, more "act as a senior PM at a B2B SaaS company" framing. This helps — briefly, marginally — and then hits a wall. There's a ceiling on how much a single prompt can compensate for a model that starts every conversation from zero:

• Facts don't fit in a prompt. Your actual activation metric, your three real customer segments, the reason your last redesign got scoped down — none of that lives in a prompt template. You'd have to retype it every single time, which nobody does past week two.
• Longer prompts still don't persist. Even a 500-word mega-prompt evaporates the moment the chat ends. Tomorrow's session starts from zero again, no matter how good today's prompt was.
• Prompts can't ask you questions. A static prompt template can't notice that you never specified a success metric and pause to ask. It just fills the gap with something plausible-sounding, and plausible-sounding is exactly how AI slop gets into a PRD.

The pack isn't badly written. It's solving the wrong layer of the problem. The bottleneck was never prompt quality — it's that a chat window is amnesiac by design.

The fix: give the model a memory, not a better question

Engineers ran into this exact wall years before PMs did, and they didn't solve it by writing better prompts — they solved it by changing where the conversation happens. Coding agents like Claude Code, OpenAI Codex, and Google Antigravity don't operate in a stateless chat box. They operate inside a folder, and they read the files sitting in that folder before they write anything.

That single architectural difference is what a context system is built on. Instead of re-explaining your product every time in an ever-longer prompt, you write it down once:

• context/product.md — what you actually ship, current focus, the metrics you actually track
• context/users.md — your real segments, their jobs-to-be-done, the pain points you keep hearing
• context/company.md — stage, how decisions get made, what "good" looks like in your org's documents

Now the prompt's job shrinks dramatically. It doesn't need to describe your entire product anymore — it just needs to say "read the context, then do this specific task." The facts live in files the agent re-reads every time, not in a prompt you retype.

Same prompt, two different outcomes — a side-by-side

Here's what changes in practice. First, the prompt-pack version — the kind you'd find in a downloaded list, pasted cold into a chat window:

Write a PRD for a new referral program feature.
Include problem statement, goals, requirements, and success metrics.

Reasonable-looking prompt. But the model has never seen your product, so it invents a referral program that resembles every other referral program it's seen in training data — generic goals, a made-up "20% increase in referrals" success metric, personas that don't match your actual users. You'll spend more time un-inventing this draft than writing one from scratch.

Now the context-system version — same task, run inside a folder that already has your product files in it:

Read the files in context/ first.

Write a lean PRD in markdown for the feature described below.
Structure: Problem & evidence · Goals and success metrics ·
Target users · Scope and explicit non-goals · Requirements
(written for engineers) · Risks · Open questions.

Ground every claim in my context files. If something important
is missing (a metric, a segment, a constraint), ASK me instead
of inventing it.

Feature: a referral program letting existing users invite others
for both sides to get a discount.

Same task, structurally similar prompt — but this one runs against real files. The agent pulls your actual user segments from context/users.md, checks whether you track a referral or invite metric at all, and if you don't, it asks you instead of making one up. That question — "you don't currently track invite conversion, what should the success metric be?" — is worth more than the entire rest of the draft. It's the agent finding a gap in your own thinking, not filling it with fiction. This is the exact approach covered step by step in how to write a PRD with Claude Code, and the same pattern extends to turning that PRD into Jira-ready user stories.

Building your own context system in three steps

You don't need to design a folder structure from scratch or write files by hand. The fastest path is to let the agent do the interviewing:

1. Open an agent in an empty folder. Claude Code, Codex, or Antigravity — any of them work, since the technique is plain markdown files, not a specific tool.
2. Ask it to interview you. Say: "Interview me to create context files about my product, my users, and my company. Ask one question at a time, then write the files." Fifteen minutes of answering questions, and you have a permanent foundation.
3. Point every future prompt at it. Start every task with "read the files in context/ first," and add one rule: instruct the agent to ask about gaps instead of guessing. That single instruction is what turns confident fiction into useful questions.

From here, your old prompt pack doesn't get thrown away — it gets upgraded. The same list of prompts you collected still has good bones; they were just missing something to read from. Point them at your context folder and they stop plateauing.

Common mistakes to avoid

• Collecting more prompts instead of building context. A 51st prompt doesn't fix what a memoryless chat window can't do. Context first, prompts second.
• Writing context once and never updating it. Your product changes; your context files should too. Stale context produces confidently wrong answers just like no context does.
• Letting the agent invent instead of ask. If you don't explicitly instruct it to flag missing information, it will fill gaps with plausible-sounding guesses. Say it out loud in the prompt, every time, until it's habit.
• Treating this as one big prompt to perfect. The unlock isn't a smarter single prompt — it's a system of files plus reusable instructions that compounds every time you use it.

FAQ