What to feed a project — and how it searches.

The teach nobody else gives you. Claude doesn't always read your project — past a point it SEARCHES it. Build for the search and your outputs stop going generic. Resource 02 is the structure. This is the content rules.

How it actually works · two modes

Small · reads everything

Under ~12 files, Claude loads every file's full text into context, every message. Best fidelity — but every file taxes every turn.

Big · RAG search (paid only)

~12+ files (or a few huge ones) and Claude switches to searching — it pulls the chunks that look relevant. Scales 10x — but retrieval can miss your gold.

Don't count words — read the badgeIn the right-side knowledge panel: a capacity bar, a status line like "Claude will look up info as needed" or an Indexing marker. The dead giveaway — in chat you'll watch Claude run a project-knowledge-search step. Search step = big. Just answers = small.

The token truth

Reading everything IS more tokens — but a chat with the same files pasted in costs the exact same tokens. A project isn't more expensive than a chat. It's the same cost, automated.

So: one-off, won't reuse the files → chat. Reuse the files or instructions → project. The thing to avoid isn't "project" — it's carrying files you don't need, because a small project has no search to skip them.

"Add everything" vs "add the right ones" — resolved

Anthropic's tip

"Add all relevant docs — more context, better." True for RECALL. It assumes RAG sorts the pile once you cross the threshold.

The operator rule

Add everything RELEVANT — and relevant means distilled, topic-named, earning its place. Dump raw + unorganized and the search scans a mess; it can't organize what you never organized.

Feed it / starve it

Feed it	Starve it
Distilled, topic-organized files	Raw transcripts, raw exports
Your 3 best samples	Your entire archive
Current pricing, current offers	Anything you'd caveat with "might be old"
One file per topic, named for search	`notes-final-v2-FINAL.md`
Files the JOB requires	Files uploaded "just in case"

The transcript workflow · distill before you upload

Run it through a chat first

Pull what you need: direct quotes, testimonials, decisions, objections, action items.

Organize by topic, not by call

Not call-march-14.md — instead testimonials-distilled.md, pricing-objections.md. One topic, one file.

Keep the raw transcript OUT

Archive it in Drive. The project gets the distilled layer only.

Direct the search · copy these phrases

# In the prompt — per task Pull direct quotes only — no paraphrasing — from testimonials-distilled.md. Check pricing-objections.md before drafting. Mirror the client's own language. Use gold-sample.md as the bar. Match its structure, not its topic. Ignore the offer files — this task is voice only. # In project instructions — to make a file win EVERY time Every email output follows gold-email-standard.md. Always check brand-voice.md before any client-facing copy.

The priority leverInstructions are read EVERY prompt. Files are searched WHEN relevant. If a file must apply every time, name it in the instructions. That's the whole trick to getting a template to prioritize.

Claude reads it small, searches it big. Either way, build for retrieval.