thinking-out-loud skill

Thinking Out Loud

A ten minute voice ramble transfers more context than any prompt a person would type, and models reconstruct rambles well. The failure is downstream and invisible: the model fills every gap in the ramble confidently. "The usual model" silently becomes a specific model. "The standard size" becomes a specific viewport. A position the user reversed mid-ramble survives as fact. None of this registers as uncertainty from the inside, so none of it ever becomes a clarifying question. The model then acts on a misreading it fully believes, and the user discovers it an hour of generated work later.

This skill is the fix: before acting on any ramble, produce an echo, a short structured audit of everything absorbed, with the model's own additions quarantined from the user's words. The user corrects three lines instead of debugging a built artifact.

Why an echo instead of follow-up questions

Asking clarifying questions is good, and the interview below does it. But questions alone cannot secure a ramble, for two structural reasons:

model believes.** A clarifying question requires felt uncertainty, and confident misreadings feel like knowledge. The echo forces every inference and gap-fill into the open whether or not it felt uncertain.

• **Questions verify what the model doubts. The echo verifies what the

facts and half-decisions. Even good questions probe three or four; the rest of the model's understanding goes unverified into action. The echo inventories the entire transfer, and it works by recognition, not recall: the user reads and spots what is wrong, which is far cheaper than producing answers, and ramblers often do not know their answer until they see the wrong guess written down.

• Questions sample; the echo audits. A long ramble carries dozens of

The contract

fragments, until the echo is approved. Reconstruct first.

1. Act on nothing. No file edits, no code, no plans, no solutions to

section, apart from the user's own content. Never present a guess in the user's voice.

1. Label every addition. Inferences and guesses live in their own

flip. Never silently average or pick.

1. Surface every reversal. Adopt the later position, but flag the

and restarts are resolved silently from context. Keep the user's own vocabulary and project names.

1. Lose nothing. Tangents get parked, not dropped.
2. Never remark on dictation artifacts. Typos, homophones, filler,

saved unprompted.

1. Ask before persisting. The approved brief is offered a home, never

When to use

restarts, filler, and mid-message reversals ("actually no, scrap that")

• A message is a long, weakly punctuated stream of consciousness with

recognition, sorry for any typos", "dictating this")

• A message opens with a voice preamble ("switching to speech

• The user says they want to ramble or think out loud
• The user asks to be interviewed to untangle a fuzzy idea

When not to use

while keeping their exact words

• Short requests that are already clear
• The user wants a verbatim transcript, minutes, or cleanup of dictation

• Long but already structured text, such as a pasted spec or document
• The user asked a direct question and wants a direct answer

The echo

One structured reply. Dense, scannable, and short: the user should find and fix an error in seconds. Full template with a worked example in references/echo-format.md.

achieve. Often this differs from what they said first; that is fine.

1. Mission: one sentence stating what the user is actually trying to

list. Mark anything they called a top priority.

1. Locked: the user's decisions and constraints, merged into one

parked tangents (one line each).

1. Open: questions the ramble raised but did not answer.
2. Ledger: flips (both positions in one line, later one adopted) and

(strongly implied but never stated) and "Guessed" (gaps you filled). Tell the user to correct these first.

1. My additions: the only interpretation callouts. "Inferred"

Compression rules, non-negotiable:

own content by definition; only the model's additions get called out.

• Nothing appears twice. Every fact lives in exactly one section.
• No "you said" recap. Everything outside My additions is the user's

or it is bloat.

• One line per bullet. If a bullet needs two lines, it is two bullets

"standard size", "soon": each lands in Open or Guessed, never absorbed into a locked item as if it were specified.

• Vague quantifiers are never silently resolved. "The usual model",

Close by inviting corrections and offering the interview.

The interview (optional)

Follow-up questions have their place: after the audit, not instead of it. Only if the user accepts the offer, or asked to be interviewed up front.

• Ask only about items flagged in Open or Guessed
• One question per message, highest information gain first
• Each question states in one clause why it matters
• Cap at five questions; stop early once answers stop changing the brief
• After the interview, restate only the sections of the echo that changed

Capture mode (multi-message rambles)

Not needed for dictation tools, where the whole ramble arrives as one message. Use it when the user invokes the skill before rambling and then adds thoughts across several messages, possibly over a long stretch.

'done' when you want the echo.")

• Acknowledge once, in one short line ("Go ahead, I'm listening. Say

("Listening."). Vary it slightly so it does not feel robotic.

• For every following message, reply with a single minimal line

sentences, then return to listening.

• Do NOT solve, praise, summarize, analyze, or ask questions mid-stream.
• If the user asks a direct question mid-ramble, answer it in at most two

any clear equivalent, then deliver the echo.