Token Optimizer

10 skills inclus

• 01

  Context Engineering

  Fait gérer à l'agent ce qu'il charge en contexte : fichiers de règles, specs utiles seulement, historique élagué, passations de session reprenables.+

  sf-context-engineering/SKILL.md, de addyosmani/agent-skills, licence MIT.

  1---

  2name: sf-context-engineering

  3description: Curates what a coding agent sees so it stays accurate over long sessions. Sets up rules files (C…

  4---

  5

  6# Context Engineering

  7

  8Give the agent the right information at the right time. Too little context and it invents APIs; too much and…

  9

  10## When to use

  11

  12- At the start of a session, a feature, or a switch to another area of the codebase.

  13- When output drifts: wrong patterns, hallucinated imports, re-implemented utilities, ignored constraints.

  14- When a project has no rules file, or the rules file is outdated.

  15- Before compacting, clearing, or handing work to a fresh session or another agent.

  16
• 02

  Prompt Engineering

  Fait concevoir, déboguer et optimiser par l'agent les prompts d'une app LLM : consignes, exemples, sortie JSON validée, cache, évaluation.+

  sf-prompt-engineering/SKILL.md, de wshobson/agents, licence MIT.

  1---

  2name: sf-prompt-engineering

  3description: Designs, debugs and optimizes prompts for LLM applications in production. Covers clear instructi…

  4---

  5

  6# Prompt Engineering

  7

  8Turn a vague or flaky prompt into one that produces the right output reliably, at a known cost, and prove it…

  9

  10## When to use

  11

  12- Writing or reviewing a prompt or system prompt for an LLM feature.

  13- Outputs are inconsistent, wrongly formatted, too long, or ignore instructions.

  14- Code needs machine-readable output (JSON, labels, extracted fields).

  15- A prompt is expensive or slow and needs trimming or caching.

  16
• 03

  Planning and Task Breakdown

  Fait découper par l'agent une spec en petites tâches ordonnées et vérifiables, avec critères d'acceptation, dépendances et points de contrôle.+

  sf-task-breakdown/SKILL.md, de addyosmani/agent-skills, licence MIT.

  1---

  2name: sf-task-breakdown

  3description: Breaks a spec or feature request into small, ordered, verifiable tasks with acceptance criteria,…

  4---

  5

  6# Task Breakdown

  7

  8Decompose work into small tasks that can each be implemented, tested and verified in one focused session, ord…

  9

  10## When to use

  11

  12- A spec, design or clear request exists and needs to become implementable units.

  13- The work is too large or fuzzy to start, or the build order is not obvious.

  14- Work will be spread over several sessions or agents, or scope must be explained to a human.

  15

  16**Do not use** for a single-file change with obvious scope, or when the spec already contains well-defined ta…
• 04

  Design Brainstorming

  Fait transformer par l'agent une idée floue en design validé avant tout code : une question à la fois, 2 ou 3 options, design écrit approuvé.+

  sf-design-brainstorm/SKILL.md, de obra/superpowers, licence MIT.

  1---

  2name: sf-design-brainstorm

  3description: Turns a new idea or an underspecified request into an approved design before anything is built.…

  4---

  5

  6# Design Brainstorm

  7

  8You are good at building. You build what you believe the user wants, and that belief is usually thinner than…

  9

  10## First: what do you actually know?

  11

  12Before replying, ask yourself what you know about what they want and why.

  13

  14- **Everything that matters is in the request** ("make the icon cornflower blue"): do it, and say what you di…

  15- **Something non-trivial is missing**: start the conversation below.

  16
• 05

  Writing Plans

  Fait rédiger par l'agent un plan d'implémentation test-first détaillé, lisible par quelqu'un qui ignore le code : fichiers, signatures, tests.+

  sf-plan-writing/SKILL.md, de obra/superpowers, licence MIT.

  1---

  2name: sf-plan-writing

  3description: Writes a detailed, test-first implementation plan from an approved spec or design, for an engine…

  4---

  5

  6# Plan Writing

  7

  8Write the plan for an implementer who knows the language but not this codebase or this spec. They will write…

  9

  10## When to use

  11

  12- A spec or design has been approved and the change spans several steps or files.

  13- The work will be executed later, by you after compaction, by another session, or by subagents.

  14

  15**Do not use** when requirements are still unclear (brainstorm or ask first), or for a one-step change where…

  16
• 06

  Executing Plans

  Fait exécuter par l'agent un plan écrit tâche par tâche : brief lu à chaque fois, tests d'abord, journal de progression, revue finale de branche.+

  sf-plan-execution/SKILL.md, de obra/superpowers, licence MIT.

  1---

  2name: sf-plan-execution

  3description: Executes a written implementation plan inline, task by task, in the current session. Reads each…

  4---

  5

  6# Plan Execution

  7

  8Execute the plan yourself, task by task, in this session: no implementer subagent per task and no reviewer pe…

  9

  10Inline execution gives up a fresh context and a second pair of eyes on every task. It keeps what they bought…

  11

  12**Core principle:** the plan already did the thinking. Execute it exactly, prove each step with a test you wa…

  13

  14## When to use

  15

  16- A plan exists (for example from `sf-plan-writing`) and the user chose inline execution.
• 07

  Verification Before Completion

  Interdit à l'agent d'annoncer un travail fini sans preuve fraîche : commande lancée en entier, sortie et code de retour lus, statut réel rapporté.+

  sf-verify-before-done/SKILL.md, de obra/superpowers, licence MIT.

  1---

  2name: sf-verify-before-done

  3description: Forces fresh evidence before any claim that work is done, fixed, passing or ready. Identifies th…

  4---

  5

  6# Verify Before Done

  7

  8**Evidence before claims, always.** If you have not run the verification in this turn, you cannot say it pass…

  9

  10## When to use

  11

  12Before any statement that implies success or correctness:

  13

  14- "Done", "fixed", "works", "all green", "ready", or any paraphrase, including enthusiasm ("Great!", "Perfect…

  15- Committing, pushing, opening a PR, marking a task complete, moving to the next task.

  16- Reporting on work delegated to a subagent or another tool.
• 08

  Skill Creator

  Fait créer et améliorer des skills par l'agent avec une boucle mesurée : prompts de test, comparaison à une base, revue humaine, réglage de la description.+

  sf-skill-creator/SKILL.md, de anthropics/skills, licence Apache 2.0.

  1---

  2name: sf-skill-creator

  3description: Creates new agent skills (SKILL.md plus scripts and references) and improves existing ones throu…

  4---

  5

  6# Skill Creator

  7

  8Create a skill, test it on realistic prompts, review the results with the user, improve it, and repeat until…

  9

  10The loop:

  11

  121. Decide what the skill should do and roughly how.

  132. Write a draft.

  143. Run an agent with the skill on a few test prompts, plus a baseline without it.

  154. Evaluate with the user, qualitatively (viewer) and quantitatively (assertions, benchmark).

  165. Rewrite based on feedback and on flaws the benchmark reveals.
• 09

  Skill Finder

  Fait chercher, vérifier puis installer par l'agent des skills avec le CLI skills : réputation, installations, lecture du contenu, accord avant installation.+

  sf-skill-finder/SKILL.md, de vercel-labs/skills, licence MIT.

  1---

  2name: sf-skill-finder

  3description: "Finds, vets and installs agent skills from the open skills ecosystem with the skills CLI (npx s…

  4---

  5

  6# Skill Finder

  7

  8Help the user find a skill that already solves their task, check that it is trustworthy, and install it only…

  9

  10## When to use

  11

  12- "Find a skill for X", "is there a skill that...", "how do I do X" for a common, specialized task.

  13- The user wants to extend the agent in a domain, or install or update skills.

  14

  15If a suitable skill is already installed, use it instead of searching.

  16
• 10

  Cavemax

  Fait répondre l'agent en version compressée pour économiser des tokens, sans rien perdre du code, des chemins, des erreurs ni des avertissements.+

  sf-cavemax/SKILL.md, de Nixus-security/Cavemax-Skills, licence MIT.

  1---

  2name: sf-cavemax

  3description: "Compresses the agent's replies to cut tokens without losing accuracy: drops filler, articles, h…

  4---

  5

  6Respond hyper-terse. Maximum compression, zero accuracy loss. Every dropped token must be recoverable from co…

  7

  8## Persistence

  9

  10ACTIVE EVERY RESPONSE. No drift back to prose. Still active if unsure. Off only: "stop cavemax" / "normal mod…

  11

  12Default: **max**. Switch: `/sf-cavemax safe|max|brutal|mute|code`.

  13

  14## Core rules

  15

  16Drop: articles (a/an/the), filler (just/really/basically/actually/simply), pleasantries, hedging, subjects +…

Extraits. Les dossiers complets sont livrés après le paiement.

Installation