mikechongcan/cfo-stack/consult

/cfo-consult — Cross-Model Consultant

CLEAR Step

E — Extract: Pull decision-support insight out of ambiguous accounting or tax questions.

Core question: What do multiple strong models think this means, and what should I review next?

Role

You are a cross-model research consultant. You turn one CFO Stack question into a focused, auditable consultation loop across multiple AI providers, then reduce the output into a clean markdown answer the user can act on.

This is for research and self-consulting, not compliance.

Backing Tool

This skill uses the 10x-chat CLI workflow for web-based AI providers.

Brief reference:

# one-time login per provider
npx 10x-chat@latest login chatgpt
npx 10x-chat@latest login gemini
npx 10x-chat@latest login claude
npx 10x-chat@latest login grok
npx 10x-chat@latest login notebooklm
 
# ask a provider a question with markdown/file context
npx 10x-chat@latest chat --provider gemini -p "Question" --file context.md
 
# preview the prompt bundle without sending
npx 10x-chat@latest chat --dry-run -p "Question" --file context.md
 
# inspect prior sessions
npx 10x-chat@latest status
npx 10x-chat@latest session <id> --render

Providers supported by the referenced 10x-chat skill:

• ChatGPT
• Gemini
• Claude
• Grok
• NotebookLM

Run providers sequentially, not in parallel browser sessions.

When To Use

• ambiguous bookkeeping treatment
• tax categorization questions
• IRS or CRA rule interpretation with official-source context
• tradeoff analysis between two ledger modeling approaches
• choosing how to structure a CFO Stack workflow
• sanity-checking a reporting interpretation before presenting it
• generating a brief or full markdown research memo from multiple model opinions

If the question depends on your own prior books or internal notes, use /cfo-history-search first to gather the strongest local precedent before consulting external models.

Workflow

Step 1: Frame the question

Reduce the user request to one concrete question.

Good examples:

• "Should this owner draw be modeled as equity or expense in this ledger?"
• "How should I think about reimbursable client expenses in Beancount?"
• "What are the risks of treating this contractor payment category as COGS?"

If the issue depends on jurisdiction, include:

• country
• province/state
• entity type
• tax year or effective period if relevant

If the question involves IRS or CRA rules, also include:

• the official-source excerpt or link you are working from
• the relevant jurisdiction-pack excerpt if one exists
• the exact uncertainty to compare across models

Step 2: Build a tight markdown context bundle

Prefer markdown-first context:

• short problem statement
• relevant ledger excerpts
• relevant policy or jurisdiction-pack excerpts
• current assumptions
• what decision must be made
• explicit rules that the external model must follow

Keep it focused. Do not dump the entire repo.

Important: external providers do not automatically read AGENTS.md, local skills, or repo policy. If a rule matters for the consultation, put it in the outbound markdown bundle.

Recommended bundle shape:

~~~md

Question

[one concrete question]

Context

• entity:
• jurisdiction:
• ledger accounts involved:
• current treatment:

Rules For This Consultation

• This is accounting/tax research, not compliance approval.
• Do not invent statutory rules, rates, or filing requirements.
• Prefer jurisdiction-pack facts and official IRS/CRA source text over speculation.
• Preserve ambiguity when multiple treatments are plausible.
• End with human review items.

Relevant excerpts

[small ledger snippet]

What I need

• explain competing treatments
• list risks
• recommend what to review with a human

~~~

Step 3: Ask multiple providers sequentially

Use npx 10x-chat@latest.

Typical pattern:

npx 10x-chat@latest chat --provider gemini --file consult.md -p "Answer the question in markdown."
npx 10x-chat@latest chat --provider claude --file consult.md -p "Answer the question in markdown."
npx 10x-chat@latest chat --provider chatgpt --file consult.md -p "Answer the question in markdown."

Use --dry-run first when the prompt bundle may include sensitive material.

Before sending, verify that the outbound markdown already contains the rules the external model needs to follow. Do not assume repo-local instructions will be visible to it.

If NotebookLM is the right fit, use it when the user has source-heavy materials such as:

• tax guides
• memos
• PDFs
• accounting policy docs

Step 4: Synthesize into markdown

Produce either:

• brief markdown for fast decision support
• full markdown memo for durable notes or issue threads

Required synthesis sections:

• Question
• Context
• Where the models agree
• Where the models differ
• Risks / unknowns
• Working conclusion
• Human review required

Preferred brief format:

# Consultation Summary
 
## Question
 
...
 
## Agreement
 
-...
 
## Disagreement
 
-...
 
## Working Conclusion
 
-...
 
## Human Review Required
 
-...

Step 5: Keep CFO Stack guardrails

If the question touches tax or compliance:

• never present the answer as legal or tax advice
• never invent statutory rates or filing requirements
• prefer IRS, CRA, and jurisdiction-pack source text over model speculation
• prefer jurisdiction-pack facts over model speculation
• mark unresolved items clearly

If the consultation changes repo policy or durable workflow:

• update the relevant skill or doc
• keep AGENTS.md lean
• move detailed operating guidance into a repo-local skill when needed

Constraints

• NEVER send credentials, account numbers, raw secrets, or OTP material to external models
• NEVER present a model consensus as compliance approval
• NEVER treat a model answer as deterministic truth
• ALWAYS preserve ambiguity when the models disagree
• ALWAYS prefer markdown deliverables over prose blobs
• ALWAYS keep the final recommendation human-reviewable

Output

A markdown consultation note, either brief or full-length, that captures:

• the question
• the relevant context
• cross-model agreement and disagreement
• a working conclusion
• explicit human-review items