A ready-to-use project template for AI-assisted accounting workflows. Clone it, fill in your plan, drop in your data, and start working.
Built on the PythonMuse methodology for controlled, repeatable, and audit-ready AI workflows.
On GitHub, click "Use this template" to create your own copy. Or clone it directly:
Windows (PowerShell):
git clone https://github.com/PythonMuse/pythonmuse-workflow-kit.git ~\Documents\my-project
cd ~\Documents\my-project
code .Mac / Linux / Git Bash:
git clone https://github.com/PythonMuse/pythonmuse-workflow-kit.git ~/Documents/my-project
cd ~/Documents/my-project
code .Prefer to watch instead of read? Full walkthrough below:
pythonmuse-workflow-kit/
.vscode/
extensions.json recommended VS Code extensions
settings.json workspace settings
CLAUDE.md instructions for AI (rules, data locations, tone)
plan.md project blueprint (fill this in first)
status_update.md rolling progress tracker
data/
raw/ original source files (never modified)
processed/ cleaned or transformed data
src/ scripts and logic files
outputs/ generated reports and results
evidence/
run-logs/ saved AI run logs for audit trail
skills/
bank-reconciliation/
SKILL.md reusable bank reconciliation workflow
margin-analysis/
SKILL.md reusable margin analysis workflow
pdf-extract/
SKILL.md PDF-to-Markdown extraction workflow
docs/ notes, memos, reference material- Edit
plan.md-- describe your objective, source data, steps, and assumptions - Drop data into
/data/raw/-- masked or sample data only - Start an AI session -- point it at
plan.mdandCLAUDE.md - AI follows the plan -- proposes steps, waits for approval, saves outputs
- Review and document -- AI updates
status_update.mdwhen done
The folder structure is not just organization. It is the context that makes AI effective.
Your folder structure is your prompt.
GitHub Copilot reads your project structure, file names, and Markdown files to understand context. This template is designed to give Copilot everything it needs.
Set up a new project:
"I just cloned the PythonMuse Workflow Kit. Read CLAUDE.md and plan.md. Help me fill in plan.md for a bank reconciliation for March 2026."
Import the structure into an existing project:
"Look at the PythonMuse Workflow Kit at github.com/PythonMuse/pythonmuse-workflow-kit. Set up the same folder structure and template files in my current workspace."
Use a skill:
"Read the bank reconciliation skill in /skills/bank-reconciliation/SKILL.md. Apply it to the files in /data/raw/ and save results to /outputs/."
Start a new session (after a break):
"Read plan.md and status_update.md. Summarize where we left off and confirm the next step."
Save progress before ending:
"Update status_update.md with what was completed, where files are saved, any issues, and next steps."
Copilot prioritizes what it can see in your workspace:
| What Copilot Reads | What It Learns |
|---|---|
Folder names (/data/raw/, /outputs/) |
Where to find and save files |
CLAUDE.md |
Rules and constraints to follow |
plan.md |
What the project is trying to accomplish |
status_update.md |
Where things stand right now |
skills/*.md |
How to execute specific workflow types |
A well-structured project replaces dozens of repeated prompts.
This kit ships with the two sample CSVs from Article 01: AI Co-Pilot for Accounting already in data/raw/. You can reproduce the entire margin-analysis story from the article without bringing your own data.
Walk through it: open docs/demo-prompts.md and run the prompts in order in Copilot Chat with Claude selected.
| Step | What you'll prompt | What you'll get |
|---|---|---|
| 1 | Orient on the data | Structural read of both files |
| 2 | Calculate margin per order | Low-margin orders flagged |
| 3 | Show formulas, tie to source | Validation you can defend |
| 4 | Export joined dataset to Excel | Workpaper for manual review |
| 5 | Compare 2024 vs 2025 | YoY margin compression view |
| 6 | Rank reps, vendors, labor | Concentration and inflation patterns |
| 7 | Visualize the story | The five article charts as PNGs |
| 8 | Promote charting to a skill | Reusable /skills/visualize-margin/ |
| 9 | Strategic summary | 5-point CFO-ready memo |
| 10 | Explain it back | Reproducible narrative |
| 11 | Validate before repeating | AI validation report + human sign-off |
| 12 | Convert to a repeatable script | Production Python in /src/ |
Step 11 is the gate -- do not script anything you cannot defend by hand. Step 12 is the inflection point: once validated, you prompt Claude to turn the conversation into a versioned, commented, human-editable script -- with a CONFIG block, # >>> HUMAN INPUT NEEDED: markers, input validation, and a paired runbook. See the full prompts in docs/demo-prompts.md.
Real client data? The demo prompts are deliberately short. Before running this flow on anything sensitive, read the Before You Start -- Real-World Safety Note at the top of docs/demo-prompts.md. Mask first, state materiality and period explicitly, and forbid silent assumptions.
A script is the first level of automation, not the last. After Step 9, here is the maturity ladder you can climb -- pick the next move that solves your real pain:
- Promote to a Skill -- move the logic into
/skills/<name>/SKILL.mdso any future session can invoke it by name. - Pin the prompt -- version-control the prompt sequence under
docs/prompts/so it survives sessions. - Externalize config -- lift the CONFIG block out of Python into
config.yamlso non-coders own the settings. - Add a smoke test -- have Claude diff this run's totals against last month's approved output.
- Add a CLI wrapper --
python -m src.<script> --period 2025-Q4so it can be scheduled. - Generate the runbook -- ask Claude to produce
RUNBOOK.mdnext to the script, not just the code. - Lock the contract -- a schema check that fails loudly when next month's file format drifts.
- Capture a golden output -- save one approved run as
/evidence/golden/and diff future runs against it.
Full prompts for each move are in docs/demo-prompts.md under Beyond the Script.
For the underlying methodology see Article 11: From One-Time to Repeatable Workflows.
If you use Claude Code (inside VS Code), the experience is similar. Claude reads CLAUDE.md automatically and uses plan.md and status_update.md as project context.
Start a session with:
"Read plan.md and status_update.md. Summarize current state and confirm next step."
| Skill | Use Case ID | Description |
|---|---|---|
| Bank Reconciliation | UC-001 | Match bank to GL, classify exceptions, produce audit-ready summary |
| Margin Analysis | UC-003 | Gross margin by segment, period comparison, concentration risk flags |
| PDF Extract | UC-004 | Convert PDF to Markdown, mask sensitive data, extract structured fields |
To add your own skill, create a new folder under /skills/ with a SKILL.md following the same format.
| File | Purpose | When to Update |
|---|---|---|
CLAUDE.md |
Instructions for how AI should behave | At project start; adjust as needed |
plan.md |
Defines the project scope, rules, and steps | At project start; revise as scope changes |
status_update.md |
Tracks completed work, outputs, issues, next steps | After each major milestone |
This template is part of the PythonMuse AI Ledger series -- practical Python, AI, and automation for accounting and finance teams.
- Article 01: AI Co-Pilot for Accounting -- the worked example reproduced by the demo prompts in this kit
- Article 08: Why Claude "Forgets" -- why these three files matter
- Article 11: From One-Time to Repeatable Workflows -- the nine-step workflow pattern
- Article 14: Stop Using One AI Like It Is Excel -- how to use Claude, ChatGPT, and Copilot together
- Article 16: The PDF Token Trap -- why PDF-to-Markdown saves tokens and how the pdf-extract skill works
- AI Accounting Framework -- the full 13-section learning framework
Found a useful skill? Built a workflow template? Contributions are welcome.
See CONTRIBUTING.md for details.
From PythonMuse -- Practical Python, AI, and automation for accounting and finance teams.