| name | claude-response-format |
| description | Canonical spec for the assistant's chat reply format in this workspace: every substantive reply is organized into `## [emoji] Short Headline` sections (one emoji + a short readable headline, content beneath); turns that change files end with a git-derived file-change report and a CONDITIONAL 'which files to review' list. The repo CLAUDE.md points here to make it always-on. Trigger: response format, reply format, section headers, emoji headers, file-change report, 回复格式. |
| argument-hint | (reference spec — usually not invoked directly) |
| allowed-tools | Bash, Read |
| metadata | {"version":"0.1.2","last_updated":"2026-06-26","summary":"Emoji headline format for chat replies + end-of-run file-change & conditional review report."} |
Skill: claude-response-format (0_utils)
Canonical format for conversational replies in this workspace. CLAUDE.md Rule 5
points here; this file is the detailed spec + examples. The CLAUDE.md line is
what makes it always-on — this skill is the reference it cites.
Scope
- Applies to CHAT replies — what the assistant writes back to the user.
- Does NOT apply to file / doc contents. Those keep CLAUDE.md Rule 4 (ASCII
headers
=== / ---, no ##). Never let this format leak into .md/.tex
files the assistant authors.
The format
Every substantive reply is split into one or more sections. Each section is:
## [emoji] Short Headline
<content for that section>
[emoji] — ONE emoji that fits the section's role. Suggestive palette (not
fixed): 🧩 short answer · 🎯 recommendation · ⚠️ caveat/risk · 🛠️ how-to ·
📋 summary/list · 🔍 findings · ✅ done · 🙋 question for you · 📁 file changes ·
👀 files to review · 🧪 experiment · 💡 idea · 📊 results · 🚧 in progress.
Pick what fits.
Short Headline — a short, readable, natural-language headline, ~2-5 words,
title case (e.g. My Recommendation, What Actually Works, Next Step).
NOT kebab-case — write it like a headline a human would scan.
- Content sits beneath the header: prose, bullets, tables, code — all fine.
Guidance
- Structure any non-trivial reply into sections; order them most-important-first.
- A trivial one-line reply can be a single section, or skip the header entirely
if it would only add noise — judgment call, don't be robotic about it.
- Don't over-fragment: ~2-5 sections for a typical reply; each earns its header.
- Keep headlines honest to the content; don't pad to hit a count.
- One emoji per header, at the start, right after
## .
End-of-run file report (📁)
End ANY turn that changed files with a ## 📁 file-changes section. Derive the
list from git — never from memory:
git status --short
git -C <submodule> status --short # if a submodule (e.g. Tools/) was touched
Group the entries: code / scripts; generated artifacts (.ipynb, results, build
outputs); data side-effects (_WorkSpace/…, local/…). Flag anything NOT
git-ignored that must not be committed (e.g. data stores).
Which files to review (👀) — conditional
Add a ## 👀 which-files-to-review section ONLY when files were really changed
AND some warrant a human read. Skip it entirely for trivial / mechanical / no-op
turns — it is not mandatory.
When shown, name the human-judgment files (hand-written logic, prose / docs, the
largest diff, the highest transcription risk), ranked by what most needs a human
eye. Mark derived / generated / copied files "derived — skip" so the user knows
not to bother (e.g. an .ipynb converted from a .py).
Examples
## 🧩 Short Answer
Yes — and one cheap test settles it.
## ⚠️ The Catch
A skill alone can't make a behavior always-on; it only runs when invoked.
## 🎯 My Recommendation
Run the back-test first, because it resolves the direction before any build.
## 📁 File Changes
code: code-dev/1-PIPELINE/3-Case-WorkSpace/builder_x.py
derived: code/haifn/fn_case/x.py (rebuilt from builder — skip)
## 👀 Files to Review
builder_x.py — the hand-written CaseFn logic; the generated x.py is derived — skip.
## 🙋 What I Need From You
Pick the model: Bedrock (BAA-covered) vs a local in-VPC model.