How to use the Thinking Out Loud agent skill

What the skill makes the agent do

When a long, rambling message arrives, the skill tells the agent to act on nothing: no file edits, no code, no plans, and no solutions to fragments. First it writes an echo, a one-reply audit of everything it absorbed.

The echo has five parts: Mission (what you are trying to achieve), Locked (your decisions and constraints, with any top priority marked), Open (questions you raised but did not answer), Ledger (reversals and parked tangents), and My additions. The last part keeps the agent's inferences and guesses separate from your own words, so you can correct those first. The agent proceeds only after you approve the echo.

• When you change your mind mid-message, both positions are shown and the later one is adopted, so the reversal is visible rather than silently resolved.
• Tangents are parked, not dropped.
• Dictation typos and filler are resolved from context without comment, and your project names and vocabulary are kept.

Install the thinking-out-loud skill

Download the skill ZIP from this page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

The skill's README also lists an install command using the npx skills add tool with the path to the thinking-out-loud folder in the awesome-llm-apps repository. Use it if you prefer a command-line install.

Trigger the skill with natural requests

You do not need a special command. The skill is meant to trigger when your message matches its description. The source names these cases: you say you want to let your thoughts out loud or ramble before building anything; your message opens with a speech-to-text preamble such as 'switching to voice, sorry for typos'; the input is a long, weakly punctuated stream with restarts and reversals; or you ask to be interviewed about a fuzzy idea.

If you want the agent to listen first, say so directly, for example 'let me think out loud for a bit'. Keep dictating, then wait for the echo.

• 'Let me think out loud before we build anything.'
• 'Switching to voice, sorry for the typos. I want to rethink the sign-up flow...'
• 'Interview me about this half-formed idea before you touch the code.'

Review and approve the echo

Read the echo and correct the My additions section first, since those are the agent's own inferences and guesses. Fix any locked decision that is wrong, and answer the Open questions. The agent is designed to close by inviting corrections and offering an interview about the open items.

After you approve, the agent offers three options: append the brief to CLAUDE.md, save it to docs/rambles/ with a dated file name, or keep it in the conversation only. It does not save anything on its own. The approved brief then guides the rest of the session.

• Correct My additions first, because a wrong guess there shapes everything after it.
• Vague words such as 'the usual model' or 'soon' are listed under Open or Guessed rather than treated as settled.

Optional capture mode for spread-out rambles

Capture mode is for when you invoke the skill and then add thoughts across several messages. The agent replies to each message with a single short line and does not analyze or ask questions until you say 'done', 'echo', or a similar phrase. Dictation tools that send the whole ramble as one message do not need it.

What it needs and safety notes

The source lists no required dependencies. The skill does not run scripts and does not use the network. It consists of instructions in SKILL.md, with a format reference in references/echo-format.md and a README.md.

Because it runs no scripts, its only file effects come from you choosing a save option after approval, such as appending to CLAUDE.md or saving a file under docs/rambles/. Check those paths before you approve the save.

• Requires: nothing listed in the source.
• Runs scripts: no.
• Uses network: no.
• Licence: Apache-2.0.

When not to use it and tips

Skip the skill for short requests that are already clear, since the echo only adds delay. Also skip it when you need a verbatim transcript or a cleanup of dictation that keeps your exact wording, because the skill tells the agent to resolve dictation artifacts silently. It is not meant for pasted specs or documents that are already structured, and it is not meant for direct questions that need direct answers.

• Keep your rambles in one message when you can, since that is the main use case.
• Expect the echo to be short. The skill targets well under 250 words for a ten minute ramble.
• Use the interview only on the items the echo flags, and expect at most five questions.