↳lexi
02 / SHARED VOCABULARY

Define it clearly.
Then say it briefly.

Turn recurring work instructions into candidate entries with explicit meanings. These are designed examples, not frequency rankings mined from user conversations.

VERIFY · Verify sourcesCheck the original source of each conclusion, cite it, and label any unverifiable claim as uncertain.For repeated research handoffs; the label does not perform verification.
SCOPE · Respect scopeModify only explicitly authorized fields, retain the original records, and stop that operation if its scope is unclear.Keep the complete constraint; shorthand does not enforce permissions.
HANDOFF · Hand off workInclude the goal, completed work, unresolved issues, evidence locations, and authorization needed for the next step in the handoff.Explicit fields reduce repeated explanations of the handoff format.
RECONCILE · Reconcile figuresCheck the period, currency, units, and data sources, list each discrepancy, and do not invent missing amounts.A candidate definition for finance and spreadsheet workflows.
REPORT · Report resultsState the outcome first, then provide verifiable evidence, remaining limitations, and items requiring the recipient's action.Defines output structure; it does not guarantee factual accuracy.
CONFIRM · Review a consequential actionBefore an irreversible action, show the exact target, expected impact, and feasible recovery method, and confirm that existing authorization applies.Do not remove authorization requirements during compression.
COST EXPERIMENT / VERIFY

When does it pay off?

Functional names are human-facing labels; the experiment uses decodable references. Count the dictionary again for every separate request that must carry it.
o200k_base / TEXT TOKENS
Original input—
Dictionary + full input—

Waiting for measured counts

Loading tokenizer…

See the full model-facing message
What makes a shared dictionary usable?
ContractWhy keep it
Definition & versionThe receiver confirms the same version so identical shorthand cannot silently mean different things.
ScopeOne message, one task or a whole session? Do not assume the next request retains the mapping.
FallbackAsk for the complete definition of an unknown entry; retain negations, amounts, dates and authorization boundaries.
Fewer characters need not mean fewer tokens
A dictionary does not modify a model’s tokenizer. A new abbreviation can split into several tokens, so this page counts the complete sequence. If software expands shorthand before sending it to a model, it may save transport or storage, without reducing model input tokens.
Hugging Face · Tokenization algorithms ↗Machine-readable candidate dictionary ↗