digest

name	digest
user-invocable	true
description	Re-layer the previous response (or a named file) into a fast-to-skim form: answer first, a review-map or key-findings short version, action-changing blockers only, detail pulled on demand. Restructures for fast reading WITHOUT dropping substance. Use when a reply came out dense, or the user says "tl;dr", "too long", "wall of text", "make this skimmable", "digest that", or invokes "digest". Do NOT use to summarize-away detail, on already-short replies, or to clean code (use deslop).
allowed-tools	Read, Write

What this does

Turn a dense answer into a fast-to-skim form the reader can absorb in seconds, with all detail kept and pulled on demand. This is restructuring, not summarizing: every decision, number, file path, command, caveat, risk, and citation in the source must survive. Write it in plain language, and make it stand alone.

Target

Default: the immediately preceding assistant message. If the user names a file or path, Read it and use that instead. Emit only the re-layered version—no preamble, no commentary.

Output shape

Re-emit in this order (mirrors the Response Contract in CLAUDE.md / AGENTS.md):

1. Answer — one line: the outcome, answer, or decision. Never buried.
2. Short version the reader can act on without scrolling:
   • Code / changes → a Markdown table with real |-delimited columns, one row per file, with whatever columns fit the change (never a single column with ·-joined fields).
   • Research / analysis → the key findings as tight bullets, most-important first.
3. Detail — keep logs, long rationale, alternatives, and full citations below the short version; do not delete them. Add pull handles ("expand the X risk", "show the full diff"). If the detail runs past ~one screen, Write it to ~/.ai/reports/<timestamp>-<task-slug>.md and leave only an index + path inline.

Rules

• Preserve substance. Cut only filler: preamble, restating the request, hedging, narration ("I then searched…"), and redundant explanation. Never cut a decision, risk, path, command, number, or source.
• Blockers, not doubt. Carry a blocker inline in the answer or the short-version bullet it affects—a failed check, an answer-flipping assumption, or an irreversible action taken—and drop it when none is decision-relevant.
• Cap the visible layer, not the content. Answer ≤1 line; short version ≤~7 bullets or one row per unit. Don't compress the content into vagueness—a digest that hides needed detail fails as badly as the wall of text it replaced.
• First screen complete. The short version must be enough to act on, not a teaser that forces a re-read.
• Terminal-safe layering. Section order + scrollback, not <details> folding (it doesn't collapse in a plain terminal). Reserve <details>/HTML for browser- or GitHub-rendered output.
• Compose, don't duplicate. digest only restructures. For AI cadence run humanizer; for code-diff slop run deslop—each a separate pass.