Exécutez n'importe quel Skill dans Manus
en un clic

Exécutez n'importe quel Skill dans Manus en un clic

compact-smart

Étoiles4

Forks2

Mis à jour20 avril 2026 à 18:09

Write a session-specific /compact instruction that preserves irreconstructible context — tickets, PRs, async state, rejected alternatives, the why — instead of the default generic summary. Use when the user asks to compact, says "compacta", "compact smart", "/compact-smart", or when a long session is about to lose its thread.

Installation

Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.

Exécuter dans Manus

Source

walis85300

walis85300/marketplace

Ouvrir le dépôt GitHub Voir les dépôts du créateur

Téléchargement

Exécuter dans Manus

Métiers associésSOC

Basé sur la classification professionnelle SOC

Commis à la correspondanceBureau et soutien administratif·SOC 43-4021

SKILL.md

readonly

name	compact-smart
description	Write a session-specific /compact instruction that preserves irreconstructible context — tickets, PRs, async state, rejected alternatives, the why — instead of the default generic summary. Use when the user asks to compact, says "compacta", "compact smart", "/compact-smart", or when a long session is about to lose its thread.

Smart compaction

Plain /compact keeps what happened and drops why it mattered. This skill writes a better instruction before invoking it.

Principle

One question filters everything: if someone took over this repo in two minutes without the conversation, what would they need that isn't in the code or git log?

Five things usually qualify — the why (problem, origin, success criterion), external identifiers (ticket IDs, PR numbers, branch names, URLs), decisions and rejected alternatives (so the future-self doesn't re-litigate them), async state (subagents, background tasks, CI, deploys — each paired with a follow-up verb, not just a mention), and session-specific user corrections. Everything else is reconstructible and should be dropped.

Procedure

Scan the session for the five things. Be specific — real IDs, real names.
Ask the user via AskUserQuestion before drafting. In a long session the user has forgotten the first half, so they need to be reminded and asked, not handed a draft to approve.
- Always ask the anchor question: "¿Cuál es el hilo principal al que vas a volver después de compactar?" Options are the 2–3 main threads you detected in the scan (use their real names), plus an "otro, te digo" escape.
- If you detected meaningful side threads, ask a second question: "Además del hilo principal pasaron estas cosas. ¿Qué hago con ellas?" Options: "preserva todas", "solo las relevantes al anchor", "suéltalas".
- Stop at two questions. Do not interrogate.
Draft the briefing for an LLM consumer, not a human. The user trusts the anchor from step 2 and will not re-read the output. Optimize for token density: labeled segments separated by periods, no narrative, no filler articles, no "we did X then Y" storytelling. Dense > readable.
Output the briefing inside a fenced code block prefixed with /compact so the user can triple-click and paste. Example final message:
```
/compact ANCHOR: ... IDS: ... DECISION: ... ASYNC: ... NEXT: ...
```
Claude Code does not allow skills to invoke built-in /compact programmatically (not exposed to Skill or SlashCommand tool), so the user runs it themselves. Do not attempt to execute it via Bash or any other tool — that spawns a new session, not compacts the current one.

If the session is mid-way through an irreversible operation — deploy in progress, staged destructive command, pending force push, running migration — warn the user before starting.

Shape

LLM-optimized format. Labeled segments, no prose flow. Example:

ANCHOR: fix service X, root cause symptom→cause. IDS: LEAD-1234, PR#567, branch fix/.... DECISION: chose C; rejected A (reason), B (reason). ASYNC: subagent auditing [topic] → on finish [action]. CORRECTIONS: [list]. UNCOMMITTED: [files]. NEXT: [action]. DROP: tool output, resolved exploration, closed debug threads.

Omit any label whose content is empty. Keep content telegraphic — drop articles, drop verbs where a colon works, use arrows for causality and sequence.

Plus depuis ce dépôt

même dépôt

dynamodb-access-patterns

walis85300/marketplace

Interactive skill for eliciting, formalizing, and persisting DynamoDB access patterns. Use when the user wants to start designing a DynamoDB table, define entities, or document how their application will read and write data. This is Step 1 of a 3-step pipeline: access patterns -> table design -> query interfaces. The output is a structured .md file that feeds into the dynamodb-table-design skill.

2026-02-254

dynamodb-query-interfaces

walis85300/marketplace

Generates language-agnostic repository interfaces (method signatures, input/output types, and entity models) from a DynamoDB table design document. Use when the user has a table design .md file (from dynamodb-table-design skill) and wants to generate the code skeleton for their data access layer. This is Step 3 of a 3-step pipeline: access patterns -> table design -> query interfaces.

2026-02-254

dynamodb-single-table

walis85300/marketplace

Orchestrator for the DynamoDB single-table design pipeline. Use when the user wants to design a DynamoDB table end-to-end or is unsure which step to start with. Routes to the appropriate specialized skill: dynamodb-access-patterns, dynamodb-table-design, or dynamodb-query-interfaces.

2026-02-254

dynamodb-table-design

walis85300/marketplace

Designs a DynamoDB single-table structure from a documented set of access patterns. Use when the user has an access patterns .md file (from dynamodb-access-patterns skill) and wants to design partition keys, sort keys, GSIs, composite keys, and map every access pattern to a concrete DynamoDB query operation. This is Step 2 of a 3-step pipeline: access patterns -> table design -> query interfaces.

2026-02-254

review

walis85300/marketplace

Reviews code for bugs, security vulnerabilities, performance issues, and style. Use when the user asks for a code review or wants feedback on their changes.

2026-02-124

explain-code

walis85300/marketplace

Explains code in plain language with architecture diagrams. Use when the user asks to understand how code works, needs onboarding help, or wants a high-level explanation of a module or system.

2026-02-124

name	compact-smart
description	Write a session-specific /compact instruction that preserves irreconstructible context — tickets, PRs, async state, rejected alternatives, the why — instead of the default generic summary. Use when the user asks to compact, says "compacta", "compact smart", "/compact-smart", or when a long session is about to lose its thread.

Smart compaction

Plain /compact keeps what happened and drops why it mattered. This skill writes a better instruction before invoking it.

Principle

One question filters everything: if someone took over this repo in two minutes without the conversation, what would they need that isn't in the code or git log?

Procedure

Scan the session for the five things. Be specific — real IDs, real names.
Ask the user via AskUserQuestion before drafting. In a long session the user has forgotten the first half, so they need to be reminded and asked, not handed a draft to approve.
- Always ask the anchor question: "¿Cuál es el hilo principal al que vas a volver después de compactar?" Options are the 2–3 main threads you detected in the scan (use their real names), plus an "otro, te digo" escape.
- If you detected meaningful side threads, ask a second question: "Además del hilo principal pasaron estas cosas. ¿Qué hago con ellas?" Options: "preserva todas", "solo las relevantes al anchor", "suéltalas".
- Stop at two questions. Do not interrogate.
Draft the briefing for an LLM consumer, not a human. The user trusts the anchor from step 2 and will not re-read the output. Optimize for token density: labeled segments separated by periods, no narrative, no filler articles, no "we did X then Y" storytelling. Dense > readable.
Output the briefing inside a fenced code block prefixed with /compact so the user can triple-click and paste. Example final message:
```
/compact ANCHOR: ... IDS: ... DECISION: ... ASYNC: ... NEXT: ...
```
Claude Code does not allow skills to invoke built-in /compact programmatically (not exposed to Skill or SlashCommand tool), so the user runs it themselves. Do not attempt to execute it via Bash or any other tool — that spawns a new session, not compacts the current one.

If the session is mid-way through an irreversible operation — deploy in progress, staged destructive command, pending force push, running migration — warn the user before starting.

Shape

LLM-optimized format. Labeled segments, no prose flow. Example:

ANCHOR: fix service X, root cause symptom→cause. IDS: LEAD-1234, PR#567, branch fix/.... DECISION: chose C; rejected A (reason), B (reason). ASYNC: subagent auditing [topic] → on finish [action]. CORRECTIONS: [list]. UNCOMMITTED: [files]. NEXT: [action]. DROP: tool output, resolved exploration, closed debug threads.

Omit any label whose content is empty. Keep content telegraphic — drop articles, drop verbs where a colon works, use arrows for causality and sequence.