A shared skill library that maps an acquisitions team's AI work to every deal stage

The client's acquisitions team had started writing Claude skills: instruction files that teach Claude to do one job the way the team does it, such as loading a model into a spreadsheet, assembling an investment memo, or pulling market demographics. Each analyst built their own. Nobody could see the whole set at once.

That caused quiet problems. Two people could write skills for the same step, a skill could depend on a colleague's skill that had since changed, and steps with no skill at all were invisible. In a working session with the team, the ask was one shared place where the deal process and the skills meet.

A web app, built from that session and live the same day. Our own review of version one was harsh: it was hard to read, and its process map came from meeting notes rather than the team's real process. Version two shipped six days later, built on the acquisitions lead's own process map.

• The map is the data. The app reads the lead's map export exactly: 4 phases, 13 stages, 90 skills, 60 flagged gaps, 11 human steps, 4 automations, and 156 relationships between skills. Each item gets a stable id, so importing the next version of the map updates in place instead of duplicating.
• A graph workspace. Modeled on the graph view in Obsidian: an explorer on the left, the graph in the middle, the selected note on the right. Phases run left to right in the map's own order, the map's own legend is drawn as-is, and a board view lays phases out as columns of skill cards.
• A versioned library. Every skill shows whether its file is in the library, who owns it, and which version is current, with history, line-by-line diffs, discussion, and a download Claude can install.
• Review. Skills move from idea to draft, in review, approved, and retired. Only admins approve, and a member who changes approved instructions sends the skill back to review.
• Audit. A rules check across the library for gaps, overlaps the map already names, stale copies of a colleague's skill, and names Claude will refuse to install, plus an optional AI review that reads the instructions themselves.
• A connector to Claude. Twelve tools to read the map, pull a skill ready to install, and push a new version, from Claude's own chat.
• An activity log of every change, from the web app and Claude alike, that cannot be edited.

Every relationship on the graph has evidence. The links were extracted from the text of the map, then checked by a second pass whose job was to reject them. Three were thrown out. Hover over any line and you see the sentence from the map that supports it.

Permissions are enforced twice: once in the app, and again in the database's own row-level rules. The connector signs Claude in as the person using it, so Claude gets exactly that person's rights, and every change Claude makes lands in the same activity log. There is no master database key in the app, so nothing bypasses those rules.

Skill files carry a header with settings such as license and which tools the skill may use. Early versions made people strip it before pushing. The header is now stored exactly as written for every version and returned on every pull. When a skill pre-approves tools, the connector says so before anyone installs it, because installing it in Claude Code lets it use those tools without asking.

The audit earned its place on its first run. Against the draft library, the AI review returned 13 findings, and several were real flaws the rules check could not see: a memo assembler whose section order disagreed with its own stage, two memo skills reading a returns summary that no skill produced, and a skill filed under underwriting that needed a diligence report first. Those were fixed in the draft.

The build passes 95 unit tests and 50 database checks, and every table in the cloud copy was checksummed against the local one to prove the load was exact.

The team's full process map is live as a working graph, with every skill, gap, and handoff in one place. The web app is live for the team to browse the library, download skills, and push new versions. When the lead revises the map, an import shows what changed and applies it without touching anyone's skill files or ownership.

Two pieces are still switching on. The Claude connector is built and passes its outside checks; Claude can sign in once one setting is turned on in the sign-in service. The optional AI review in the audit waits on an API key, and the rules check runs on every page load in the meantime.