ThomasMoreAI tone-modes
Two registers, one test, one rule about scripts. This skill is a portable codification of a pattern: keep the internal reasoning candid, keep the outbound language defensible, and never let a script pretend to be a lawyer.
When this skill fires
- About to draft any user-facing artifact: letter, complaint, position paper, email, exhibit cover.
- About to paraphrase something the user said into paperwork.
- About to present the output of an intake / triage / packet script as the answer rather than as a scaffold.
The two modes
Lawyer mode. The register for anything that leaves the private workspace. Properties:
- Claims are stated as positions, not as feelings. Every factual assertion is traceable to an exhibit or a cited rule.
- Adversarial language ("they're trying to re-underwrite the policy after the loss") is rephrased ("the insurer's conduct is inconsistent with the agreed-value endorsement at policy inception").
- Hedges where warranted — "the record shows," "the file reflects," not "obviously" or "clearly."
- Tight. The read-aloud test (below) is the pressure check.
Casual mode. The register for internal reasoning, comments in drafts, strategy discussions, and the user's own notes. Properties:
- Honest about weak legs of the argument. If the specialist-labor-rate premium is the weakest line item, casual mode says so.
- Uses the user's own idiom, not courthouse idiom.
- Free to name what the counterparty is doing in plain language.
- Must not leak into the final artifact. Anything written in casual mode needs a translation pass before it goes outside.
The read-aloud test
Before any lawyer-mode paragraph ships, read it out loud (literally — subvocalize the words). Two filters:
- Would a hostile reader quote this sentence back at you? If the sentence contains a word that reads as snide, speculative, or emotionally loaded, it's a gift to opposing counsel. Rephrase.
- Does every noun and verb carry its own weight? Adverbs ("grossly," "blatantly") and intensifiers ("clearly," "obviously") almost always weaken the sentence they inhabit. Cut them.
If a sentence passes both filters out loud, it's ready.
Scripts as scaffolds, not oracles
The toolkit's scripts — situation_classify.py,
authorities_lookup.py, pii_scrub.py, build.py — are
deterministic rules engines against YAML tables. They are good at
what they do. They do not:
- Know whether the user's framing is correct.
- Know whether an authority's jurisdiction actually covers this matter.
- Know whether a PII substitution preserves meaning.
- Know whether a packet is persuasive.
The rule: every script output is a starting point the assistant reviews with the user, not a final answer the user signs. When a script emits a classification, name it aloud and let the user confirm. When PII scrub proposes a substitution, show the diff and let the user confirm. When the packet builder emits a PDF, open it and read it.
Synthetic example
Casual-mode note in the Maryland-Mustang file:
They're pretending the specialist-shop rate is a customary-rate problem. It's not. It's them trying to re-underwrite an agreed-value policy after the loss.
Lawyer-mode rendering in the MIA complaint:
The insurer's 2025-05-09 position letter deducts $5,280.50 from the agreed value for "non-customary charges." This deduction is inconsistent with the agreed-value endorsement (CIM-AV-ENDT-2023) in effect at policy inception, which sets the schedule of value absent fraud or material misrepresentation — neither of which is alleged.
Same fact, two registers. The casual version stays in the internal notes and is never copy-pasted into paperwork.
Do not
- Do not try to write in both registers at once. Pick the mode, write the thing, then translate if it needs to move to the other register.
- Do not let a script's output pass straight into a filing. Every output is a draft for the user, not a conclusion.