尚未检测到更新。GitHub 发布版本后会显示在这里。
按最近已完成的 8 小时观测回合间的 GitHub 星标增量排序
Open Net Worth is an open-source personal AI wealth assistant you run yourself: one ledger for your whole financial life, worked by an AI agent you control.
翻译准备中
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权，自行管理介绍，无需管理员批准。
登录并生成验证文件，在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件，即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录，将前往登录页面。登录后请回到此处，再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify，即可验证所有权并自行管理介绍。
一分钟完成验证Curated recipes (prompts and patterns) for financial work with AI agents. We share the recipe, not the black box.
翻译准备中
正在加载完整 README。
查看完整 README ↗Curated recipes for doing financial work with AI agents, the right way.
We share the recipe, not the black box.
Prompts and skills you can read, audit and make yours. Never a black box you install on trust.
Prompts and patterns for FP&A, modeling and reporting with Claude (and any capable agent). Each recipe is battle-tested in the field, opinionated, and honest about where a prompt stops being enough.
This is the cookbook. It sits in a small family of open entry points to Layerz:
| Project | Role | Link |
|---|---|---|
cookbook (you are here) | The recipes: how finance people actually use AI, day to day | the playbook |
finance-md | The standard: how an organization encodes its financial conventions for any agent | the spec |
| Layerz MCP | The tool: how an agent drives Layerz to build structured, versioned models | the integration |
slides-for-claude | The presentations: turn a model or a topic into a self-contained HTML deck | the skill |
The standard tells an agent what your numbers mean. The tool gives it a place to build that does not drift. The cookbook shows what to ask for in the first place.
New to the cookbook? Instead of scanning the whole list, follow a path: the recipes and skills already here, put in the order one kind of user actually reaches for them.
| Path | For | Starts with |
|---|---|---|
| The part-time CFO's kit | Fractional / part-time CFOs and finance advisors | Understand the business, then a flash audit, then the recurring engine |
| The founder's kit | Founders building and owning their own plan and numbers, pre-CFO | Get a rough model on the table, then make it defensible for a raise |
| The FP&A engine | In-house FP&A and controllers running the monthly cycle | Own the context once, then harden the close, variance and forecast |
| The deal desk | M&A, transaction advisory and deal teams working under deadline | Read a model you did not build, stress it, ship it clean |
More paths will land as the catalog grows. Each one is pure curation, it points to the recipes and skills below.
A good prompt can take you a long way. Then it hits a wall: nothing persists between sessions, the logic drifts the moment an assumption changes, and you cannot audit where a number came from. That wall is not a failure of prompting, it is the point where you need structure (a documented FINANCE.md) and persistence (a real model, e.g. Layerz).
So every recipe here is written to be useful on its own, and to name the wall it cannot cross. When you hit that wall, the recipe points you to finance-md or Layerz. No recipe pretends a prompt is a model.
We share the recipe, not the black box. Everything here is plain, readable Markdown you can open, audit, fork and make yours. The enemy is never the format, it is the black box: an opaque artifact that rots, hides its logic, and runs a canned process on you whether it fits your business or not. We do not ship those. That is the Layerz thesis applied to this repo: own your context, no black box, no drift.
So the cookbook has two kinds of entry, and both stay transparent:
recipes/): prompts you paste. Some do the task now (compare scenarios, audit a model you received), some forge an artifact you own and can read (a script, a FINANCE.md, a skill of your own). Read, paste, own it.skills/): small listen-first consultants you install. A skill here asks before it acts and adapts to your case, it never unrolls a template blindly. It is still open Markdown you can read and change, not a compiled box. Install it, then fork it.A skill in this repo is a recipe that listens. It earns the "not a black box" line by being auditable and interview-first, the opposite of a canned automation.
| Recipe | What it solves |
|---|---|
| Own your context | Stop re-explaining your conventions every session |
| Parametric scenarios | Generate and compare assumption-level scenarios without the model falling apart |
| Audit an inherited model | Understand and stress-test a model someone else built, fast |
| Pre-delivery review | Forge a pre-flight review skill, tuned to your conventions, so nothing ships broken |
| Harden your close skill | Audit a close-variance skill you already built and make it deterministic and gap-proof |
| Analytical review | Forge a recurring review that understands the business first, then surfaces variances, trends and the questions to raise |
Each recipe follows the same shape (see TEMPLATE.md): the problem, the prompt, where it breaks, and what to reach for when it does.
Small listen-first consultants you install. Unlike a recipe you paste, a skill loads itself when the moment fits, then asks before it acts. It is still open Markdown you can read, audit and fork, never a compiled box.
| Skill | What it does |
|---|---|
| Finance flash audit | Walks a company's finance and back-office like a fractional CFO on day one: interviews first, maps the flows, tells the truth about how reliable the numbers are and what that costs in cash, then hands back a prioritized roadmap of where to act and in what order |
| Business plan sparring partner | Sets the foundations of a founder's plan before any model: asks the stage and the purpose first, short Socratic questions, the three to five drivers that move the outcome (operational ones included), where cash and revenue drift apart, and ends with a one-page plan brief the founder owns |
| Plan to model | Builds the plan brief into a first model, in one pass, that someone else can trust: assumptions per year with their sources, revenue and cash kept apart (deferred revenue, payment terms, grants), three statements on their own tabs, checks that flag when the plan breaks, and two sober dashboards, at the depth the stage calls for |
| Finance context workspace | Sets up the working folder your finance agent reads from, so you stop re-explaining your setup every month: interviews your real close, scaffolds a context workspace you own (conventions, cost centers and owners, sources, mapping), and wires two loops, the deterministic close checks and a variance pass that drafts the question to send each cost-center owner |
| Three-statement builder | Builds a transaction-ready integrated model, but interviews you hard first (the revenue engine, the accounting specifics, the deal structure and adjustments), then builds it so the P&L, cash flow and balance sheet tie out, and hands the tie-out to structure because a prompt cannot hold a balancing model |
| Shadow model |
Each skill is a single SKILL.md. Download it, then in Claude: + > Skills > Manage skills > Add > Upload skill. Read the file first, that is the point.
Claude Code, other agents, and how to make a skill yours: skills/INSTALL.md.
The cookbook above is what we wrote. The library below is what others wrote and we found worth your time. Curation, not authorship. Suggestions welcome (see Contributing).
| Resource | Author | Good for | Link |
|---|---|---|---|
| Anthropic Cookbook | Anthropic | Foundational prompting and agent patterns (not finance-specific) | https://github.com/anthropics/anthropic-cookbook |
| your suggestion | you | what it is good for | open a PR |
This table is intentionally short. A library of 10 vetted resources beats a dump of 200. If you know one that earns its place, open a PR.
Recipes tell you what to ask for. These say why, with sources and method:
On practice: building a model that doesn't drift · token efficiency in finance · auditable models
Recipes and library entries are welcome. Read CONTRIBUTING.md and copy TEMPLATE.md. The bar is simple: it has to be something you actually used, and it has to be honest about its limits.
MIT, see LICENSE.
We share the recipe, not the black box.
Maintained by Layerz. The financial memory of Claude: a structured model it builds, edits and exports to Excel, without drifting.
| Rebuilds a model someone handed you (a seller's model, a management plan) independently, then reconciles the two and tells you which assumption drives every gap and the question to put to the other side |
| Deal red flags | Learns the specific business first, then surfaces the red flags a buyer should not miss (is the EBITDA real, working capital stretched, concentration, debt-like items), ranks them by impact on value, and turns each into the diligence question to ask |
| Account mapping | Maps a raw accounting export (a trial balance or a general ledger) onto clean, structured financial statements, under the right framework (French PCG, German SKR, a company COA): confirms the framework, reuses the context you own, maps the standard accounts in a batch, and raises every judgment call instead of dropping it in "other" |
| Software spend review | Turns raw bank or card transactions into a defensible view of software, cloud and AI spend, starting with the part naive categorisation gets wrong: one vendor line is not one cost, so it splits the composite bills (seats, usage, tokens, payment fees, marketplace resale) before counting anything, asks for the billing detail instead of guessing an allocation, and reports what it cannot resolve |