← ClaudeAtlas

transcript-hygienelisted

Bound what subagents return and what tool output enters the transcript; load when briefing a subagent or reading large output.
JakeSelby/model-citizen · ★ 21 · AI & Automation · score 74
Install: claude install-skill JakeSelby/model-citizen
# Transcript hygiene The operative lines live in the resident rule `primitives/rules/transcript-hygiene.md`. This skill carries the reasoning behind them, the numbers in context, and the examples. The user reads the transcript, not just the final message. **Thinking summaries are wanted and stay** — they are collapsed and opened deliberately. Everything else in the scroll is cost they did not ask for, and the worst offenders are tool output and relayed subagent reports, not prose. ## Read narrowly - **Never `cat` a whole file to answer a narrow question.** `sed -n '40,80p'`, `head -40`, or `grep -n` with the pattern. Read the whole file only when you will actually use the whole file. - **Never chain `cat A && cat B && cat C`** to orient. Orient with `grep -n '^#'` for headings, `wc -l` for size, then read the part that matters. - **Filter every `find` and `ls`.** A repo with a `renders/`, `data/` or `target/` directory will print hundreds of generated filenames. Scope the glob or pipe to `head`. - **A command printing more than ~100 lines needs a reason** you could state out loud. ## Bound what a subagent hands back `delegation.md` says to cap the return. These are the numbers. - **Gathering agent: 400 words.** Research digest: **600**. Adversarial review: findings only, no restatement of what it read. - **Detail goes to a file, not into the return.** Brief it to write the long version into the scratchpad and return the verdict, the findings that change a dec