Guide: Seeding Your Knowledge Files

Claude Cowork · Feature primer

The Problem This Solves

Every recipe in this library asks you to create knowledge files — our-standards.md, icp.md, brand-voice.md, customers.md, scoring-rules.md. These files are what make Claude’s output specific to your business rather than generic.

The problem: a blank file is hard to fill. Most people know roughly what should go there but freeze when they try to write it. What counts as a non-negotiable? What are industry-standard terms? What exactly makes a customer ideal vs. not?

Synthetic seeding solves this. Instead of writing the file from scratch, you let Claude research the space, ask you targeted questions about your specific context, and generate a draft — then you react to that draft rather than inventing from a blank page. The result is a populated knowledge file grounded in real norms and calibrated to your business, built in minutes rather than hours.

When to Use This

Use synthetic seeding any time you’re stuck on a knowledge file. It works especially well for:

• our-standards.md — contract non-negotiables and industry norms (see: Contract Review recipe)
• icp.md — ideal customer profile (see: Lead Triage recipe)
• brand-voice.md — how you sound and what you avoid (see: Campaign Execution recipe)
• scoring-rules.md — what signals matter for your business
• Any knowledge file where you know roughly what belongs but struggle to articulate it

How It Works

The technique runs in two phases, each using Claude’s AskUserQuestion tool to present structured choices rather than open-ended blanks.

Phase 1 — Discovery Claude asks you a short set of targeted questions about your context — your industry, deal sizes, counterparty types, business model. These aren’t open-ended essays. They’re multiple-choice prompts where you pick the answer that fits, with a free-text option if nothing quite matches. Three to four questions is typical.

Phase 2 — Review Claude generates a draft of the knowledge file based on its research combined with your answers. Then it walks you through each item — one question at a time — and asks you to confirm, modify, or reject it. Again: structured choices, not blank text boxes. You’re editing, not authoring.

By the end you have a populated, calibrated knowledge file. You didn’t invent it from scratch — you shaped a well-researched starting point into something that fits your business.

The Seeding Prompt

This prompt works for almost any knowledge file. Replace the bracketed parts with the specific file and context you’re working on.

I need to create [file name, e.g., our-standards.md] for my [recipe name, 
e.g., Contract Review] workflow, but I'm not sure what to put in it.

Here's what the file should contain: [paste the file description from the recipe].

Before generating anything, ask me 3-4 discovery questions about my business 
and how I typically operate — use the AskUserQuestion format so I can select 
from options rather than write from scratch.

After I answer, research the relevant industry norms for my situation and 
generate a complete draft of the file. Then walk me through each item in the 
draft using AskUserQuestion so I can confirm, modify, or remove each one 
before we finalize it.

Example: Seeding our-standards.md for Contract Review

The discovery phase looks like this:

Claude presents something like:

What types of contracts do you sign most often?

• (A) Client service agreements and MSAs
• (B) Vendor and software agreements
• © Both in roughly equal amounts
• (D) Other

What’s a typical deal size for contracts you sign?

• (A) Under $10K
• (B) $10K–$50K
• © $50K–$250K
• (D) Over $250K

How would you describe your counterparties?

• (A) Mostly larger companies presenting their standard templates
• (B) Roughly equal — sometimes we send the template, sometimes they do
• © Mostly smaller vendors and contractors where we’re the larger party

After you answer, Claude researches standard terms for your combination (e.g., professional services company, $10K–$50K deals, mostly receiving counterparty templates) and generates a draft.

The review phase looks like this:

Claude walks through each generated item:

For your non-negotiables, I’ve suggested: “Uncapped liability is always a red flag.” For your deal sizes and contract types, this is standard. Does this fit?

• (A) Yes — keep it as written
• (B) Yes, but only for deals over $25K
• © No — liability caps aren’t a concern for us
• (D) Modify: [free text]

I’ve suggested: “Auto-renewal windows shorter than 60 days should be flagged.” Is this right for you?

• (A) Yes — 60 days is the right threshold
• (B) Make it 30 days — we monitor renewals closely
• © Make it 90 days — we need more lead time
• (D) Remove this — not relevant

Each item gets confirmed, adjusted, or removed. You finish with a our-standards.md that reflects real industry context calibrated to how you actually operate.

Tips for Getting Better Results

Be specific in the seeding prompt. The more context you give Claude before the discovery phase, the sharper the research. Mention your industry, your typical deal type, and roughly how complex your contracts tend to be.

Don’t skip the review phase. The generated draft is a starting point, not a finished product. Some items will be exactly right. Some will be close but need adjusting. A few may not apply to your business at all. The review phase is where the file becomes yours.

Revisit after your first real use. After you run the recipe with the seeded file for the first time, you’ll have a much clearer sense of what’s working and what’s missing. Update the file based on what you actually saw — gaps, over-flagging, or items that never came up.

Use it for any knowledge file, not just standards. The same technique works for icp.md (“here are the ICP characteristics Claude researched for your industry — which ones apply?”), brand-voice.md (“here are five tone profiles — which is closest to yours?”), and scoring-rules.md (“here are the urgency signals common in your sales context — which do you actually see?”).

What This Teaches You About Claude Cowork

Reacting is easier than inventing. The cognitive load of filling a blank file is much higher than reviewing a populated one. Synthetic seeding exploits this — Claude does the first draft, you shape it. This applies far beyond knowledge files: any time you’re stuck, try asking Claude to generate a starting point and react to that instead.

AskUserQuestion structures the conversation. In Cowork, the AskUserQuestion tool lets Claude present choices rather than open-ended prompts. This is the difference between “describe your ideal customer” (hard) and “which of these best describes your ideal customer?” (easy). Look for opportunities to use this pattern any time you’re asking for input that could be framed as a choice.

The discovery phase earns the review phase. The reason the review phase produces good results is that the discovery phase gathered the right context. If the generated draft feels off, the fix is usually to add more context in discovery — not to correct individual items one by one.