Add handler-quiet-output to the built-in skills

Agents narrate far more than anyone reads: the transcript is not the
deliverable, and prose there is spent tokens burying information where
no one looks. The new skill routes each kind of output to its store —
work happens through tool calls; a minimized NOTES.md ledger (one
bullet per action, committed with the work) records what happened and
how; problems and causes go to memory; status goes to the final
checkpoint-sized message the Stop hook captures onto the checkmark; and
questions go through the question tool, which reaches the operator as a
push notification and an answer prompt in the web and mobile apps
instead of stalling silently in the transcript.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01731mKtVzsfeT4Vi3TvkR48

CHANGELOG.md

@@ -38,10 +38,15 @@ since its last release:
 
 ### Added — built-in operator skills, pre-installed on every deployment
 
-Seven skills now ship inside Handler (`handler.builtin_skills`) and are seeded into
+Eight skills now ship inside Handler (`handler.builtin_skills`) and are seeded into
 the managed skill store on API startup, so every fresh install — and every existing
 deployment on upgrade — starts with the judgment layer the hard gates can't enforce:
 
+- `handler-quiet-output` — work through tool calls, not prose: the transcript is not
+  the deliverable. A minimized `NOTES.md` ledger records what happened and how,
+  problems go to memory, status goes to the checkpoint, and questions go through the
+  question tool (push notification + answer prompt in the web/mobile apps) — never
+  typed into the transcript.
 - `handler-gate-recovery` — respond to a blocked completion/push gate by fixing the
   real failure; never delete/skip tests, weaken the mise `test` task, or `--no-verify`.
 - `handler-testing` — every behavior change lands with a test that fails without it;

README.md

@@ -335,7 +335,8 @@ What the dashboard can now do (all state-changing actions require `ADMIN_TOKEN`)
   choices a human would be asked — always user scope, the instructions' defaults — and
   reports them in the command result for after-the-fact review.
 - **Built-in operator skills** ship with Handler and are seeded into the managed store
-  on API startup (`handler.builtin_skills`): gate recovery, testing standard,
+  on API startup (`handler.builtin_skills`): quiet output (tool calls + a minimized
+  `NOTES.md` ledger instead of transcript prose), gate recovery, testing standard,
   checkpoint quality, memory discipline, mise-task rules, scheduled-run continuity,
   and secrets hygiene — the judgment layer the hard gates can't enforce. Seeding is
   idempotent by name, so operator edits and disables survive upgrades; deleting one

src/handler/builtin_skills.py

@@ -24,6 +24,49 @@ from .db import repository as repo
 # minus front-matter (claude_gen adds name/description at sync time). Kept as plain
 # data so the content is easy to review and diff, exactly like skills_gen._SKILLS.
 BUILTIN_SKILLS: list[tuple[str, str, str]] = [
+    (
+        "handler-quiet-output",
+        "Work through tool calls, not prose: the transcript is not the deliverable. "
+        "Keep a minimized NOTES.md ledger instead. Always applies.",
+        """# Quiet output: tools and notes, not prose
+
+Nobody watches your transcript live, and nobody reads it afterwards — the operator
+reads NOTES.md, the checkmark, and memory. Narration is spent tokens that bury the
+information somewhere no one will look for it.
+
+## Work through tool calls
+
+- Don't announce what you're about to do, recap what you just did, or restate file
+  contents — the tool calls are the record.
+- No essays or running commentary in the transcript. A plan worth keeping goes in
+  NOTES.md; a plan not worth keeping isn't worth typing.
+- The one message that matters is your **final** one: the Stop hook captures it onto
+  the checkmark. Keep it checkpoint-sized — a few lines of status, not a report.
+
+## NOTES.md is the output
+
+Maintain `NOTES.md` at the repo root (create it if missing) as a minimized running
+ledger, committed with your work:
+
+- One bullet per meaningful action: what changed, where, and why in one line —
+  `- fixed retry loop in poller.py (cause: timeout treated as success)`.
+- Facts, not narration. Bullets, not paragraphs. No restated diffs — the commits
+  hold the code.
+- Append under a dated/session heading; don't rewrite earlier sessions' entries.
+- Scheduled runs: your prompt's state file serves this role — keep one file, not two.
+
+## Route everything to its store
+
+- **What happened, how**: NOTES.md.
+- **Problems, surprises, causes**: memory (gotcha/decision notes) — durable and
+  shared, per the memory skill.
+- **Current status + next step**: the final checkpoint message.
+- **Questions**: the question tool (AskUserQuestion; `ask_operator` on the pi
+  harness), never prose. A deferred question reaches the operator as a push
+  notification and an answer prompt in the web and mobile apps; a question typed
+  into the transcript reaches no one and stalls the run.
+""",
+    ),
     (
         "handler-gate-recovery",
         "What to do when the completion gate or push gate blocks you. Use whenever "