Common foundation
Framing a dictated request
A request dictated on your phone, decoded with your lexicon, rewritten to a template, checked against eight criteria, and carried out only after your “go”.
The problem
You dictate quickly, between two patients or in a corridor, on your phone. Speech recognition makes phonetic approximations: an acronym turns into an everyday word, a tool name becomes unrecognisable, “Thursday” does not say which one. The request arrives blurred, Claude guesses, and the result has to be redone. The time saved by dictating is lost twice over in corrections.
Step-by-step method
- Keep a lexicon of approximations: a two-column table, “heard → meant” (for example “staff tock” → “staff talk”, “my cynical cases” → “my clinical cases”). The lexicon only ever receives additions. Claude decodes the dictation with it, without ever asking you to repeat yourself.
- Rewrite the request to a fixed template: Objective, Context, Sources, Deliverable, Safeguards, Verification, Target skill, Assumptions. Claude shows you this rewrite before taking any action.
- Run the eight-criterion grid: each heading of the template is marked present, inferred (flagged “assumption”) or missing. Missing criteria lead to a single grouped question, each point with a default answer.
- Wait for the “go”: nothing is carried out before your explicit approval of the rewritten request. Any correction from you produces a new version, to be approved again.
- Execute, then report under four headings: Result · Location · Verification · Caveats (see Setting safeguards).
- Turn corrections into lessons: when you correct a decoding or an assumption, Claude proposes a lexicon line or a rule; it is added only after your approval.
- Never create a scheduled task from a dictated sentence: a routine is described in writing, with its plan, and approved separately.
An example (fictitious)
Dictation: “prepare Thursday’s staff tock on the valves like last time”.
- Objective: prepare the talk for Thursday’s staff meeting.
- Context: the department’s weekly staff meeting, about fifteen minutes.
- Sources: the last approved talk (assumption: last month’s).
- Deliverable: around ten slides in the department template.
- Safeguards: references checked with DOI or PMID, no identifiable patient data.
- Verification: every reference found, template respected.
- Target skill: staff meeting talk.
- Assumptions: “staff tock” decoded as “staff talk”; “the valves” understood as a review of valve disease.
Grouped question: “Which Thursday, and which valve first? Default: next Thursday, all valves.” Then Claude waits for “go”.
Seven rules for asking well
- One request = one objective, one named deliverable, one surface (a conversation, a folder, a tool).
- Say where you are speaking from: from the phone, Claude only reads and drafts; anything that must touch files is prefixed (for example “at the desk:”).
- Name the skill when you know it.
- Attach the file, do not describe it.
- For a scheduled task, write the plan, not the prompt.
- Changing computers is a system task, to be named explicitly.
- End every working session on files with a three-line log: inputs available, latest version kept, what is missing.
From the phone: “no desk” mode
When the request comes from the phone and needs the computer, Claude says so in one sentence, then works from the cloud. It drops its result into a “to integrate” area of the shared space (a dedicated Google Drive folder, for example), with a line in the decision queue; the evening run files that drop on the computer. It never answers “I don’t have access” without offering this route. Health and identity data stay out of this circuit.
Deliverable
The rewritten and approved request, with assumptions flagged, then the result with its four-heading report. As a bonus, a lexicon and a set of rules that grow with each accepted correction.
Safeguards
Nothing is carried out without a “go”, and an assumption stays flagged until you confirm it. The lexicon and the rules only receive additions, always after approval. A dictated sentence never creates a scheduled task. No health or identity data pass through the “to integrate” area.
Starter prompt
Here is my dictated request. Decode it with my lexicon, rewrite it to the template (objective, context, sources, deliverable, safeguards, verification, target skill, assumptions), flag the assumptions, ask me at most one grouped question, then wait for my go.