Skip to main content

cmd-code-what

Catch the user up on what the agent has been doing in the current session with 3-5 ultra-tight bullet points. Use when the user invokes /cmd-what, asks "what are you doing", "what did you do", "catch me up", "status", "where are we", or otherwise needs a fast situational summary of in-flight agent work.

Informações da origem

Repositório
Olshansk/agent-skills
Última atividade na origem
3 de maio de 2026 às 20:47
Idioma detectado do SKILL.md
inglês
Estrelas
9
Forks
0

Opções de instalação

Por padrão, está selecionado o prompt que primeiro revisa a origem. Você pode mudar para um comando direto ou baixar uma cópia local.

Revise os arquivos de origem

Leia o SKILL.md e os arquivos complementares exibidos pelo SkillsMP antes de decidir se vai instalar.

Exibindo SKILL.md

SKILL.md
Instruções da origem · Visualização somente leitura
name
cmd-code-what
description
Catch the user up on what the agent has been doing in the current session with 3-5 ultra-tight bullet points. Use when the user invokes /cmd-what, asks "what are you doing", "what did you do", "catch me up", "status", "where are we", or otherwise needs a fast situational summary of in-flight agent work.
disable-model-invocation
true
# What <!-- omit in toc --> Give the user a high-signal, low-noise snapshot of what's happened in this session so they can re-enter the loop without scrolling back. ## Output format Reply with **3-5 bullet points only**. No preamble, no closing summary, no headers, no follow-up questions. Each bullet must be: ``` - **1-3 word label**: explanation in less than 120 characters ``` Hard rules: - **3-5 bullets, no more, no less in spirit** (4 is the sweet spot) - **Label**: 1-3 words, bolded, title-case or kebab-case, names a concrete thing (file, action, decision, blocker) - **Explanation**: a single clause, **< 120 characters including the label**, present tense or past tense — never future tense ("will", "going to") - No nested bullets, no code fences, no links, no emoji unless one is genuinely load-bearing - Skip anything the user already knows from the last 1-2 turns — focus on what they'd miss if they just walked back to the terminal ## What to include (in priority order) 1. **Current focus** — what file/task is actively being worked on right now 2. **Most recent concrete change** — what just got written, edited, run, or decided 3. **Blocker or open question** — anything waiting on the user, or that's stuck 4. **Next step** — the very next action the agent intends to take, if obvious 5. **Notable side effects** — files created, branches touched, processes started, only if relevant ## What to leave out - Restating the original user request verbatim - Tool-call narration ("I ran ls, then grep, then…") - Anything that happened more than ~10 turns ago unless it's still load-bearing - Hedging, qualifiers, or meta-commentary about the summary itself ## Example User: `/cmd-what` Good response: ``` - **Editing**: refactoring `auth/session.py` to drop the legacy cookie path - **Tests**: 14/16 passing, 2 failing on token refresh edge case - **Blocked**: need to know if expired refresh tokens should silently re-auth or 401 - **Next**: wire the chosen behavior into `refresh_token()` once you decide ``` Bad response (too verbose, narrates tool calls, future-tense, missing labels): ``` - I started by reading the auth module and then I grepped for cookie usage - I'm planning to refactor the session handling but first I want to check with you - The tests will probably need updates too ```
Ver no GitHub