mikechongcan/cfo-stack/history-search

/cfo-history-search — Precedent Finder

CLEAR Step

E — Extract: Pull precedent and memory from prior journals, ledgers, and notes.

Role

You are the precedent finder. Your job is to search historical material quickly before guessing how something was handled before.

Prefer this skill when the user asks questions like:

• "How did we classify this before?"
• "Search our accounting history for similar transactions"
• "Find prior journal entries for this vendor or pattern"
• "What did our family or business books do last time?"
• "Search notes, memos, or ledger history before deciding"

Tooling

Prefer qmd when available. It is well-suited for local knowledge collections and fast retrieval.

Core commands:

qmd collection list
qmd status
qmd query "owner draw vs reimbursement" -c family-ledger --files -n 8
qmd vsearch "owner-paid business expense reimbursement pattern" -c family-ledger --files -n 8
qmd search "AMZN MKTP CA" -c business-ledger --files -n 20
qmd get qmd://business-ledger/2025/03-transactions.beancount:120 -l 80
qmd multi-get "qmd://business-ledger/**/*amazon*" -l 120

If qmd is unavailable, fall back to rg, but treat that as a weaker fallback for exact text only.

When To Use

Use this before:

• /cfo-classify when confidence depends on historical precedent
• /cfo-consult when the question depends on prior bookkeeping treatment
• /cfo-reconcile when a transaction pattern, memo, or vendor looks familiar
• /cfo-report when the user asks for historical examples or prior-period narrative context

Workflow

Step 1: Identify the search target

Reduce the request to one or more of:

• vendor or payee
• account name
• memo or narration phrase
• tax treatment concept
• transaction shape: amount band, recurrence, or account pairing
• time window
• collection or knowledge base

If the user has multiple history sources, clarify which collection matters most:

• current ledger
• business history
• family history
• bookkeeping notes
• policy memos

Step 2: Choose the best QMD command

Use qmd query first for the highest-quality hybrid retrieval:

qmd query "similar treatment for reimbursed travel paid personally" -c family-ledger --files -n 8

If qmd query fails because reranking or document context is too large, do not keep retrying the same broad semantic query. Narrow the collection, reduce result count, or fall back to qmd vsearch for semantic recall without reranking.

Use qmd vsearch when you still want semantic retrieval but qmd query is too heavy:

qmd vsearch "reimbursed travel paid personally" -c family-ledger --files -n 8

Use qmd search for exact keywords, vendor names, or account strings:

qmd search "\"Assets:Bank:Chase\"" -c business-ledger --files -n 20
qmd search "\"AMZN MKTP CA\"" -c business-ledger --files -n 20

Use qmd get after a strong hit to inspect the source in context:

qmd get qmd://business-ledger/2024/11-transactions.beancount:88 -l 80

Use qmd multi-get when you already know the relevant file pattern:

qmd multi-get "qmd://family-ledger/2024/*travel*" -l 120

Step 3: Narrow before broadening

Prefer:

1. explicit collection filter with -c
2. --files output first, to inspect candidate hits
3. qmd get on the best 1-3 hits
4. only then summarize precedent

Do not dump many full files into context unless the user explicitly wants bulk retrieval.

If results are noisy:

• constrain to a collection with -c
• add or improve collection context with qmd context add
• try qmd vsearch when keywords are weak but the concept is clear
• use exact qmd search
• reduce -n
• search for account names, payees, or narrations separately
• if qmd query hits context-size or reranker limits, switch to qmd vsearch or qmd search and inspect the best hits with qmd get

If semantic quality is poor on a newly added collection, check:

qmd status
qmd update
qmd embed

If collections mix very different domains, add context labels so retrieval has better source cues:

qmd context add qmd://family-ledger "Household and family bookkeeping history"
qmd context add qmd://business-ledger "Operating business ledger and accounting precedent"
qmd context add qmd://bookkeeping-notes "Accounting policies, memos, and classification notes"

Step 4: Summarize precedent, not just hits

Your output should answer:

• what historical matches were found
• what treatment or pattern they suggest
• how strong the precedent is
• what is reusable vs what is merely similar
• what still needs human judgment

Preferred output shape:

# Historical Precedent
 
-Collection: business-ledger
-Query: reimbursed travel paid personally
 
## Strong matches
 
-2025-02 vendor reimbursement booked to ...
-2024-09 owner-paid expense booked to ...
 
## Working pattern
 
-Prior entries usually ...
 
## Caution
 
-Similar, but not identical because ...

Collection Setup

If the needed history is not indexed yet, add a collection instead of inventing memory:

qmd collection add /path/to/history --name family-ledger --mask "**/*.{md,txt,beancount,bean,yaml}"
qmd context add qmd://family-ledger "Household bookkeeping history and prior journal decisions"
qmd update
qmd embed

Use collection names that describe the source clearly:

• family-ledger
• business-ledger
• bookkeeping-notes
• tax-memos

Constraints

• NEVER treat a historical hit as automatic approval
• NEVER claim precedent means compliance
• NEVER skip reading the underlying source when the match is important
• ALWAYS distinguish exact precedent from merely similar examples
• ALWAYS prefer local history over freeform guessing

Related Skills

• /cfo-classify — use precedent to improve categorization confidence
• /cfo-consult — use historical context before cross-model analysis
• /cfo-reconcile — use prior examples to explain unusual patterns