Product Context

Product context

Every skill in this pack reads one file before it does anything. Without it an agent writes plausible marketing for a company it has invented from your landing page, and every draft after that inherits the invention. This skill builds that file.

It is the least glamorous thing here and it is the one that decides whether everything else is any good.

The rule that makes this worth doing: ask, do not guess

An agent handed a website can produce a confident description of a business in seconds, and it will be wrong in ways nobody catches for months, because the wrong parts read fine. So this skill works in two passes and the second one is not optional.

1. Read what exists. The site, the docs, the README, anything the founder points at. Extract only what is actually stated.
2. Ask about the rest, then mark what is still unknown. Every field you could not verify gets asked as a question, in one batch, in plain terms. Anything the founder does not answer stays in the file marked UNKNOWN rather than filled with something reasonable. A blank the drafting skills can see is safe. An invention they cannot see is not.

Never write a claim into this file that nobody said out loud.

What the file answers

Nine questions, in this order. The order matters: the later ones are only answerable once the earlier ones are honest.

1. What is it, in the words a customer would use. Not the category you would like to be in. If your customers say "a tool that tells me why my site is slow", write that, not "a performance observability platform".
2. Who is it for, specifically enough to exclude people. A description that excludes nobody is not a description. Name who this is not for.
3. What did they use before. The status quo you are displacing, whether that is a competitor, a spreadsheet, an agency or nothing at all.
4. What can you prove. Anything you can point at: a number you measured, a customer who will say it out loud, a public result. Each one gets the caveat that travels with it. This is what stops replies being opinions. The claims skill owns the detail; this file holds the pointer.
5. What are you claiming that you cannot yet prove. Being explicit here is what keeps it out of the copy by accident.
6. The words you keep. The five or ten terms you want to own, spelled the way you spell them.
7. The words you refuse. The category words, cliches and competitor framings you will not use, with a one-line reason each. This list does more work than the one above.
8. Where you publish, and where you do not. Channels you are actually present on, plus the ones you have deliberately ruled out, so no skill proposes them again.
9. What you will never say. Hard lines: claims you will not make, comparisons you will not draw, things that are legally or personally off limits. Every drafting skill treats this as absolute.

Where the file lives

One file, PRODUCT.md, in product-context/ under the estate this pack shares: SKILLS_ESTATE, which defaults to ~/.claude/content. On a default install that is ~/.claude/content/product-context/PRODUCT.md. Create the directory if it is not there.

The exact path is the whole handoff, so it is worth being boring about. Every skill below reads that string and no other. A context file written anywhere else is a context file nothing ever opens, and the drafts go back to being generic without anyone being told why.

Treat the file as private working material, not as a public document. It holds things you have not published and possibly things you never will. Nothing in this pack sends it anywhere.

How the other skills use it

Each of these opens product-context/PRODUCT.md at that path by name.

• voice reads it before calibrating, so the forced choices are about your register rather than about writing in general.
• draft reads it first, always, and refuses to write generic copy when the file is missing: it asks you to run this instead.
• tells reads the refused-words list and treats each one as a check.
• The desks read the publishing channels and the never-say list before anything is drafted for a venue.

A field marked UNKNOWN propagates honestly: a skill that needs it says so rather than filling the gap.

With AfterLaunch connected (five answers arrive drafted, four are asked)

Everything above needs no key. With one, most of this file already exists, because the product read your site and extracted the business context it drafts from. Connected, the record is home and PRODUCT.md is a cache of it: read the record first, and fall back to the local file only when there is no key.

Your brand voice is already here from your site; the voice skill trains your own.

All nine still go to the user: the drafted five are for confirming or correcting, not for skipping.

Read first, in this order. list_kb_pages to see what is there, then get_kb_page on each of the four pages by slug, then get_snapshot for the positioning and the competitor set with the difference statement carried with each rival.

Drafted from the record, questions 1, 2, 3, 6 and 7. Show what you found, question by question, as a draft answer to correct. Never present a page as settled: the pages are templated once from an automatic extraction and are not re-derived afterwards, so a live one can be stale or truncated mid-sentence. A drafted answer the user does not explicitly confirm is never written into PRODUCT.md, and that field stays UNKNOWN. Silence is not a yes.

• Question 1, what it is in customer words. about-my-business holds the category, the one-liner and the value proposition. Say plainly that those three are written in marketing register, then ask for the sentence a customer would actually use.
• Question 2, who it is for. my-audience.
• Question 3, what they used before. my-competitors, which is short, plus the competitor set already read above.
• Question 6, the words you keep. The "how I talk about it" phrases on about-my-business, plus the characteristic phrases on about-my-voice.
• Question 7, the words you refuse. The avoid list on about-my-voice covers part of it. Ask for the rest.

Asked cold, questions 4, 5, 8 and 9. What you can prove, what you claim and cannot prove, where you publish and where you do not, and what you will never say. The record cannot answer any of the four. Nothing read off a website knows what you would stand behind in public or what you refuse to say.

Send every answer back. All nine, whether confirmed, corrected or answered cold, go back with record_insight, kind fact, one plain sentence of 10 to 500 characters, one insight per answer. That is what makes the product draft from the corrected version instead of re-learning the wrong one. record_insight is the only write this skill has: no tool writes a page, so a stale page is corrected by an insight rather than edited from here. Write the same answers into PRODUCT.md as well so the offline skills keep working, and treat that file as the copy rather than the original.

Without a key nothing above changes: all nine questions get asked, anything unanswered stays UNKNOWN, and the file is the only home there is.

What this skill cannot do

It cannot check whether any of this is true in the market. It cannot ask an answer engine what it thinks your product is, hold this file steady across sessions on its own, or tell you whether your positioning is working. It writes down what you know. Measuring whether the world agrees is a different job and it needs something with a meter and a memory.

House rules

British English. No em-dashes, no exclamation marks. Plain words over category language, in this file above all: a context file written in marketing prose teaches every downstream skill to write marketing prose.