The struggle

You ask for a competitive angle and get one that ignores the segment you actually sell to. You ask it to draft a spec and it proposes the integration you killed last quarter, for the second time. So you type the background again — what the product does, who the users are, the constraint that always applies, the thing you already decided — and the next answer is better. Then you open a new tab and it’s gone.

The instinctive diagnosis is that you need better prompts, and there’s a whole genre of advice built on that premise. It’s mostly wrong, and you can prove it to yourself: take a generic answer, paste in three paragraphs of real product context, ask the identical question again. The improvement is enormous, and none of it came from phrasing.

Prompt craft resets to zero every session. Context, written down once, compounds — across your sessions, and across anyone else’s. The reason nobody has one is that building it feels like documentation rather than leverage, so it never wins against whatever is due Friday.

The chain

Step 1 — Harvest what’s already written · Notion AI

Your context already exists, scattered across shipped PRDs, retro notes, the strategy doc nobody reads and last quarter’s board update. Use Notion AI — or whatever holds your docs — to pull it rather than writing from memory, which is how you end up with an aspirational description of your product instead of a true one.

Ask it for: what the product does in plain terms, who the segments are and how they differ, the constraints that always apply, the things you decided not to do and why, and the vocabulary your company uses for things. That last one matters more than it sounds: house vocabulary is the fastest tell between an answer written from inside your company and one written from outside it.

Honest trade-off: if your docs are thin or badly out of date, skip this and write the pack from scratch in step 2. Harvesting bad context is worse than starting clean — you’ll launder stale decisions into something that now looks authoritative.

Step 2 — Compress it into a pack · Claude

The whole point is compression. A pack that’s 40 pages doesn’t get loaded and doesn’t get maintained. Paste the harvest into Claude:

Below is raw material about my product, pulled from internal docs.

Turn it into a CONTEXT PACK: the standing briefing an assistant should have
before answering any question about this product. Target 120 lines or fewer.

Structure it as:
- What the product is and who it's for (5 lines max)
- Segments, and how their needs differ
- Constraints that always apply (technical, commercial, regulatory)
- Decisions already made and NOT up for debate, each with its one-line reason
- House vocabulary: the words we use, and what we mean by them
- Open questions we genuinely haven't resolved

Hard rules:
- Every line must be a specific claim, not a value statement. "Enterprise buyers
  need SSO before they'll trial" is context; "we're customer-obsessed" is noise.
  Cut all noise.
- If the source material is contradictory or ambiguous, put the item under a
  "NEEDS A DECISION" heading instead of picking one. Do not resolve it yourself.
- Do not add anything that isn't in the source. If a section is thin, leave it
  thin and say so.

Source material:
"""
<paste the harvest>
"""

Then edit it yourself, and expect this to be the slow part. Not because the draft is bad, but because the “NEEDS A DECISION” list is where you discover your team never actually agreed on who the primary segment is. Resolving those is the real work, and it’s worth more than the pack.

Step 3 — Make it load automatically · Claude Code

A pack you paste is a pack you’ll stop pasting. Two ways to make it automatic, and you can do both:

Claude Projects — put the pack in project knowledge; every chat in that project starts with it. Zero setup, right answer for most PMs.
CLAUDE.md — Claude Code reads a markdown file at the start of every session. This one is version-controlled, reviewable in a pull request, and shareable with your team, which makes the pack a team asset rather than your personal trick.

Anthropic’s own guidance is to keep CLAUDE.md under about 200 lines — longer files consume context and, counterintuitively, reduce adherence. If yours is growing, split it with @path/to/file.md imports rather than letting one file sprawl.

Step 4 — Point it at live systems (optional, highest ceiling)

A static pack goes stale. MCP connects Claude Code to Jira, Figma, Slack, Google Drive or a database, so instead of describing your roadmap you let it read the roadmap. Anthropic’s docs give the shape directly: “Add the feature described in JIRA issue ENG-4521 and create a PR on GitHub.”

Skip this until steps 1–3 are habit. It’s a setup cost per integration, and a live connection to the wrong system is worse than a good static file.

Where it breaks

Writing the product you wish you had. The single most common failure. A pack full of positioning language makes every answer more confidently generic, because you’ve handed the model marketing copy and asked it to reason from it. Specific claims only.
Letting it go stale. A pack asserting a decision you reversed in March is worse than no pack — you’ll now get wrong answers with your own authority behind them. Ten minutes a month, or delete it.
Bloat. Past a couple of hundred lines, adherence drops and you’re paying context for instructions being ignored. If it’s growing, split it, don’t pad it.
Laundering assumptions into facts. Everything in the pack gets treated as true forever. Keep “NEEDS A DECISION” and “Open questions” as real sections, and resist the urge to tidy them away — an honest gap produces better answers than a confident guess.
Confidentiality. A context pack is, by design, the most concentrated description of your strategy that exists. Know where it’s stored, know your company’s policy on what may go into a cloud tool, and keep customer names and unreleased plans out unless you’ve checked.
Assuming it fixes reasoning. Context fixes relevance. A model with perfect context can still be wrong about your market. The pack removes an excuse; it doesn’t remove your judgment.

“Every AI session starts with the same 400 words about what your product is, who uses it, and what you already decided not to build — and the output is still generic.”