A regulation changes; your P&Ps have to change to match. Sometimes that's one section in one P&P. Sometimes it's a firm-wide sweep across dozens of P&Ps where you don't even know the full scope up front. Same kind of job — three levels of leverage. Pick the level that fits the size and messiness of the change.
You have a small CSV (or just the text) of the changed rule, and you already know which P&P it touches — your loss-mit P&P. You bring both into a chat and prompt:
"Here's the updated loss-mitigation rule. Here's my loss-mit P&P. Find the sections that reflect the old rule and rewrite them to comply with the new one."
Claude rewrites the affected sections. Done.
When it's enough: one rule, a target you can identify yourself, scope you can eyeball. Fast, zero setup.
Where it runs out: you do the targeting and the prompting every time; nothing is reused; if the rule text changes you re-bring it by hand; and it can't answer "which of my 40 P&Ps does this even touch?" For that, you need the next level.
This one isn't "edit a section." It's a whole workflow:
Two ingredients make this possible: the Skill and the Project.
The skill carries every staged prompt, the discipline rules (cite the rule verbatim; never assert AI that isn't evidenced; stay within scope), and the exact output format. You don't write prompts — you just say "inventory this bundle" or "draft the policy," and the skill runs the right stage the right way, the same way every time. Consistency and rigor come baked in, not improvised per chat.
A single chat can't hold a firm-wide job. The Project adds two things:
Here's how the sharing actually works (this is the part worth getting right):
┌─ Chat: scan Bundle 1 (fair lending) ─┐
├─ Chat: scan Bundle 2 (valuation) ─┤ each chat emits a
├─ Chat: scan Bundle 3 (QC / servicing) ─┤ structured FINDINGS block
└─ Chat: scan Bundle 4 (infosec) ─┘
│
│ ▲ going UP: the synthesis chat SEARCHES the
▼ │ project's chats to gather every findings block
┌────────────────────────────────────────────┐
│ Chat: SYNTHESIS │
│ • assemble all findings │
│ • chase the missing info (who / what) │
│ • draft the MASTER AI governance policy │
└────────────────────────────────────────────┘
│
│ save the master policy → PROJECT KNOWLEDGE
▼ (now every chat can read it)
┌─ Chat: realign Bundle 1 P&P ─┐
├─ Chat: realign Bundle 2 P&P ─┤ each P&P rewritten to match the
└─ … ─┘ master policy + vendor governance + change log
So it's fan-out → fan-in → fan-out: scan each bundle, pull them all together into one master policy, then push that policy back out so each P&P realigns to it.
The honest mechanics (so you're not surprised): the chats don't hold a live conversation. They share two ways — the synthesis chat searches the other chats to pull findings up, and the finished policy is saved to Project knowledge so the P&P chats can read it on the way back down. Cross-chat search can lag right after you run the scans, so the reliable habit is to also paste each chat's FINDINGS block into the synthesis chat. Same outcome, just don't rely on search alone.
When it's enough: complex, firm-wide changes where you don't know the full scope up front and you need consistency and memory across many documents.
Level 2 has one soft spot: the regulation CSV is still a file. You can swap it to stay current — but someone has to swap it. When the rule corpus updates, a new CSV gets sent around and you drop it into the project.
MCP closes that gap. Instead of a file you swap, the regulatory corpus is delivered into Claude as a live feed (a connector). The skill/project pulls the rules from that feed, so when the source is updated, every project is current on the very next run — no file swap, no reinstall, no "is everyone on the latest version?"
When it's enough: ongoing operation where currency matters and you don't want to babysit updates. The trade-off: a connector has to be set up once (usually a quick IT step), where a CSV file does not.
Think of Levels 2 and 3 as two rungs of the same "stay current" goal: swap-the-file (manual, works anywhere, today) → live-feed (automatic, needs a one-time connect).
| Prompting | Skill + Project | + MCP | |
|---|---|---|---|
| Best for | one known change, one P&P | firm-wide change across many P&Ps | always-current, ongoing operation |
| You provide | the rule + the P&P, every time | your P&Ps (the skill + CSV do the rest) | your P&Ps |
| Finding what's affected | you pick the P&P | the scan finds where it touches | same |
| The workflow | you prompt each step | the skill runs every stage | same |
| Many-document memory | none | project knowledge + cross-chat findings | same |
| Keeping rules current | re-bring them each time | swap the CSV knowledge file | delivered live, automatically |
| Setup | none | install the skill / make a project | + connect the feed (one-time, IT) |
It's the same job at every level — make your P&Ps match the rules. Reach for a prompt when it's a one-off you can target yourself. Reach for the skill + project when the change is big and messy and has to stay consistent across many documents. Add MCP when you want the rules themselves to stay current without anyone tending the file. You climb the ladder only as far as the job requires.