Skip to content
LedgerLedger

The library

Every move, in one place.

117 free lessons. The whole practice, tax through advisory, plus building your own tools.

Tap one, or describe what you're stuck on below.

New here? Start with Compliance & safe AI use (free) and go in order. Jump anywhere once it clicks. The Builder track is its own path. Start it whenever you want AI building FOR you.

§ Ask Ledger for a lesson on anything

Describe a real thing you need to get done. Ledger builds the drill from your exact situation.

The client work

Returns to notices to advisory. 12 desks · 32 lessons on the client work, all free.

1Compliance & safe AI use (free)7 lessons›

Client information under s. 7216, grounding, and what never gets pasted. The skeleton, practiced to reflex.

Show it one memo that cleared review›What client data can go in a prompt (almost nothing)›Build your firm's research master prompt›Delegate a full pre-signature review›Make AI defer to your workpapers, not its guess›Force the answer into a decision matrix›Convene an AI board of advisors›
2Workflows & automation5 lessons›

Design a repeatable task into an AI workflow. With a human approval gate exactly where it belongs.

Map a repeatable task into a workflow›Design the approval gate before anything leaves the firm›Spec the follow-up cadence for missing client documents›Automate the engagement-start checklist›Set the control dial: where a human stays in the filing workflow›
3Claude & Cowork mastery3 lessons›

One platform, mapped to the job: a Project for your standing context, a Cowork skill for a repeatable prompt, MCP connectors for your live inbox and calendar.

Give Claude your firm once›Turn your best prompt into a skill›Let Claude read your inbox and calendar›
4Client data and s. 72166 lessons›

What counts as a disclosure and why a paste is one, de-identifying a fact pattern in sixty seconds, the consent that satisfies the regulation, which tools need one, and red-teaming last week's prompt.

What counts as a disclosure, and why paste is one›De-identify a fact pattern in 60 seconds without losing the question›Write a §7216 consent that actually satisfies §301.7216-3›Which AI tools need consent and which do not (the control question)›When "taxpayer" beats the client's name, and when it does not›Red-team your own paste: run the scanner on last week's prompt›
5Tax research that survives review6 lessons›

Ask a question the model cannot answer from vibes, force the citation and open it, the three shapes of fabricated authority, settled against unsettled, the memo from a messy email, and when to stop and call it a position.

Ask a research question the model cannot answer from vibes›Force the citation, then verify it exists before you rely on it›The three ways a model fabricates authority (case, section, revenue ruling)›Distinguish settled from unsettled, and make the model say which›Build the issue/rule/analysis/conclusion memo from a messy client email›When to stop researching and call it a position, with substantial authority›
6The practitioner's week5 lessons›

Audit the week before you hand the first task over, catch the section that does not say what the answer says, cost the scripted run, wrap the day into the workpapers, and know where a return goes when you paste it.

Audit your week before you hand the first task over›Audit an answer before it goes in a memo›Cost the scripted run before busy season›Wrap the day into the workpapers›Where a return goes when you paste it›
7Notices, exams and resolution0 lessons›

Translate a notice into what the client has to do, draft the abatement request from the facts, build the exam response index, explain an assessment without promising an outcome. Written next.

Translate an IRS notice into what the client actually has to doSoon
Draft the penalty abatement request from the reasonable-cause factsSoon
Build the response package index for an examSoon
Explain an assessment to a panicking client without promising an outcomeSoon
First-time abate: qualify it before you write itSoon
What never goes in a letter to the ServiceSoon
8The close and client accounting0 lessons›

Turn a messy trial balance into a close checklist, list the exceptions the model cannot explain, draft the month-end narrative an owner will read, chase the missing forms without forty emails. Written next.

Turn a messy trial balance into a close checklistSoon
Reconciliation exceptions: get the model to list what it cannot explainSoon
Draft the month-end narrative a non-financial owner will readSoon
Chase the missing 1099s and W-9s without writing 40 emailsSoon
Categorize the uncategorized, and flag what needs a humanSoon
Build the recurring close template once, reuse it every monthSoon
9Financial statements and reporting0 lessons›

Footnote language from the facts, variance analysis that says why, the line between compilation, review and audit, management letter comments a client acts on, the no-assurance line and where it goes. Written next.

Draft footnote language from the underlying factsSoon
Variance analysis that says why, not just how muchSoon
Compilation vs review vs audit: never let the output blur the lineSoon
Management letter comments that a client will actually act onSoon
Turn a 40-page statement package into a one-page owner summarySoon
The 'no assurance is provided' line, and where it must appearSoon
10Advisory and planning0 lessons›

Entity selection the client can follow, the quarterly estimate explainer with assumptions on the page, scenarios without an invented rate, the year-end letter by client type, and pricing the advisory engagement in writing. Written next.

Entity selection comparison the client can followSoon
Quarterly estimate explainer with the assumptions on the pageSoon
Model a scenario without letting the model invent the rateSoon
The year-end planning letter, segmented by client typeSoon
Turn a tax return into three advisory conversationsSoon
Price the advisory engagement, and scope it in writingSoon
11Practice, staff and season0 lessons›

Engagement letter scoping from a discovery call, the intake questionnaire that prevents February, review notes that teach, triage by risk, the fee increase letter, and the firm's own AI use policy. Written next.

Engagement letter scoping from a discovery callSoon
The client intake questionnaire that prevents FebruarySoon
Review notes that teach instead of just correctingSoon
Triage the extension list by risk, not by alphabetSoon
Draft the fee increase letter you have been avoidingSoon
Write the firm's own AI use policy, and make staff read itSoon
12Independence and ethics in practice0 lessons›

Independence in one question, nonattest services and the self-review threat, conflicts between two clients in one deal, what the Confidential Client Information Rule adds, Circular 230 due diligence on a model's draft, and whether to disclose your AI use. Written next.

Independence in one question: are you auditing your own workSoon
Nonattest services and the self-review threat, in plain languageSoon
Conflicts between two clients in the same dealSoon
What the Confidential Client Information Rule adds on top of s. 7216Soon
Circular 230 due diligence when the model wrote the first draftSoon
Disclose your AI use to clients, or do not. Decide it deliberatelySoon

Builder track

Stop using AI. Start building with it. Claude Code, skills, agents, and systems that run without you. 13 areas · 85 lessons, capability-first, all free.

B1Setup & first contactInstall Claude Code, meet the terminal, and have your first real conversation about the work on your desk.›
Four words before you startBuild →Pick a plan & installBuild →Terminal tour: read your dashboardBuild →IDE setup: give it a windowBuild →Your first real task, iteratedBuild →Your files as contextBuild →Talk to it: voice inputBuild →
B2The project brainCLAUDE.md sets the trajectory: your firm's standing context, your voice, your client-data rules, baked in.›
CLAUDE.md: set the trajectoryBuild →/init on a real folderBuild →The 200–500 line ruleBuild →Split into rules/Build →Auto-memoryBuild →Secrets & .env hygieneBuild →
B3Steering & safetyA minute of planning saves ten of building, and it is the difference where return information is involved.›
Permission modesBuild →Grill, then planBuild →/permissions rulesBuild →Bypass mode, safelyBuild →Cost capsBuild →The human-in-the-loop matrixBuild →
B4Context engineeringKnow what fills the window before you paste a 300-line trial balance into it.›
/context anatomyBuild →Handoff, clear, compactBuild →Token densityBuild →Specific beats vagueBuild →Pick model & effortBuild →The overhead auditBuild →
B5The verify loopTask -> do -> verify. On a memo, the verify step IS the work; skip it and you have signed a guess.›
Task → do → verifyBuild →Screenshot loopsBuild →Tests as verificationBuild →Fresh-data testingBuild →Cold-start runsBuild →Self-annealingBuild →
B6Skills: package your processThe close checklists and review procedures you already keep become software the AI runs the same way every month.›
SOP → SKILL.mdBuild →Frontmatter anatomyBuild →Scripts as musiciansBuild →Iterate to 99%Build →Self-healing skillsBuild →Your skill libraryBuild →
B7Connect your tools (general-ledger exports, calendar, your data)Connectors that let it read the general-ledger export and the calendar, plus skills other people built for you: install, vet, audit. Nothing identifiable goes through one you have not read.›
What MCP isBuild →Install & vet a serverBuild →Audit the token costBuild →Paste the API docs insteadBuild →MCP → skill conversionBuild →Build a tiny serverBuild →
B8The DO frameworkCode executes, the LLM judges. Compound error math explains why a long chain drifts, and why a figure never comes from the judge.›
Docs → chatbots → agentsBuild →The PTMRO loopBuild →Directive anatomyBuild →Code executes, the LLM judgesBuild →Compound-error mathBuild →MetadirectivesBuild →
B9Subagents & parallel workFan out when it pays; feeling productive isn't being productive. The reviewer agent is the one worth running.›
When to fan outBuild →Reviewer & QA agentsBuild →Summaries, not raw textBuild →Git worktreesBuild →3 approaches in parallelBuild →The 3–4 agent capBuild →
B10Design: proposals & client one-pagersEngagement proposals, the client one-pager and the year-end letter that look designed, built against a reference you supply.›
The screenshot-compare loopBuild →Reference-driven designBuild →Component librariesBuild →Decks & one-pagersBuild →Brand consistencyBuild →The design QA criticBuild →
B11Firm & data agentsReconciliations, close packages and the month-end narrative that produce themselves, from exports you de-identified first.›
Excel that explains itselfBuild →docx & pptx skillsBuild →Sheet-connected pipelinesBuild →The weekly market report factoryBuild →Ask-your-workbook agentBuild →Deliverable factoriesBuild →Build your market second brainBuild →
B12Ship a real toolWebhooks, schedules, observability, one durable tool for the firm, demoed.›
Deploy a static siteBuild →Wrap a skill in a URLBuild →WebhooksBuild →Cron in plain EnglishBuild →Secrets & observabilityBuild →Design pass: ten checks on a phoneBuild →Capstone: ship your practice toolBuild →One real build, start to finishBuild →
B16The work folderFrom chat to a folder of your own files that one agent reads and writes. Keep a list of your corrections, turn it into a skill, shape the folder around the work, and hand it on.›
Which layer are you onBuild →Test your setup with one tiny jobBuild →Keep a list of your correctionsBuild →Turn corrections into a skillBuild →Start with the outcomeBuild →Shape the folder around the workBuild →Give each stage a fileBuild →One model, different jobsBuild →Keep it current and hand it onBuild →

Vendor notes, dated

Prices and plan names move. Each note says when it was looked up, whether a person has confirmed it, and the day it comes due. Past that day, treat it as a question.

  • Claude. Claude Pro is $20 a month billed monthly, or $17 a month billed annually. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • ChatGPT. ChatGPT Plus is $20 a month. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Google Gemini. Google Gemini's paid consumer plan is Google AI Pro at $19.99 a month; Google AI Ultra starts at $99.99 a month. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Anthropic. Anthropic's consumer plans (Free, Pro, Max) may use chats to train models under a setting the user controls, with five-year retention when training is on and 30 days when it is off; Claude for Work, Education, Government and the API are excluded. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Anthropic. Anthropic's usage policy, effective 2025-09-15, requires a qualified human reviewer and disclosure of AI use for housing, legal and financial decisions. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • OpenAI. OpenAI's consumer plans train on chats unless the Improve the model setting is turned off, and a Temporary Chat is not used for training; Enterprise, Business, Edu and API data are not used by default. [S] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Google Gemini. Google Gemini Apps keeps activity 18 months by default, keeps human-reviewed chats up to three years, and tells users not to enter confidential information. [P] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Microsoft Copilot. Microsoft Copilot on a personal account may use chats to train models unless the user opts out; Entra and Microsoft 365 work accounts are excluded. [S] Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Claude. Claude Team, Enterprise and API inputs and outputs are not used for training by default; a BAA is available only on Enterprise and the first-party API after admin action. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Claude. Claude Free, Pro and Max inputs can be used for training if the user allows model improvement, with retention up to five years if allowed and 30-day deletion otherwise (pages dated 2026-03-16 and 2026-07-01). Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • OpenAI. OpenAI API inputs are not used for training by default; abuse-monitoring logs are kept up to 30 days; zero data retention is available by approval. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • OpenAI. OpenAI Business, Enterprise and Edu no-training defaults are reported by secondary sources only; the vendor trust pages returned 403 on the check date. Read the vendor page yourself. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Microsoft. Microsoft Copilot and Copilot Chat with Enterprise Data Protection: prompts and responses are not used to train foundation models, the DPA applies, and Microsoft acts as processor (Learn page updated 2026-08-18). Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.
  • Google. Gemini in Google Workspace content is not human-reviewed or used for training outside the customer's domain; default retention is 18 months. Researched Oct 6, 2026, not yet confirmed by a person, recheck by Dec 15, 2026.

Every area gives away its first moves free. Each track spins endless personalized lessons.