convez.ai
Skill
blueif16

agentic-prompt-design

blueif16/agentic-prompt-design

Ver no GitHub
InicialPluginScore · de 0 a 100

A Claude Code skill for authoring prompts that drive an agent/executor to produce thorough, bar-meeting output instead of thin, minimal work.

Instalação

$npx convez-ai blueif16/agentic-prompt-design

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

agentic-prompt-design

A Claude Code skill for authoring prompts that drive an agent/executor — a workflow node, an agent() call, a sub-agent, or a standalone generation/analysis/coding agent — to produce thorough, substantial, bar-meeting output instead of thin, literal, minimal work.

If you keep getting back the shortest plausible answer — the model places exactly the entities you named and stops, declares victory early, or summarizes what it would do — this skill is the fix. It treats thin output as a specification gap, not a model defect, and gives you the levers to close it.

Why it exists

Thin output is rarely solved by more doctrine ("be thorough!", "be comprehensive!"). Models optimize for the criteria you actually state. The cure is to make the bar explicit, observable, and enumerable so "do more" has a defined target the model can aim at and check itself against. This skill packages that doctrine into a reusable, citation-backed guide.

It is especially aimed at cheaper/weaker/non-frontier executors (MiniMax, 7B/SLM, Haiku-tier, DeepSeek Flash), which reward clear contracts rather than clever prompts — but the acceptance-bar techniques apply to any executor.

What's inside

FilePurpose
SKILL.mdThe craft guide — the node-prompt skeleton, thoroughness levers, in-prompt acceptance bars/rubrics, self-critique passes, hand-off contracts, weak-executor tactics, anti-patterns, a diagnose-and-fix loop, a pre-ship checklist, and a worked thin→substantial example.
references/research-and-sources.mdFull provenance — 35 external sources (official vendor docs, peer-reviewed research, practitioner playbooks) behind every technique.

The core ideas

  • The node-prompt skeleton — a prompt is a contract, not a wish: role · inputs · task · output_spec · the_bar · coverage · constraints · self_check · scope_fence. The load-bearing slots for thin→substantial are the_bar, coverage, and self_check.
  • Eliciting thoroughness — "go beyond the basics" modifiers, enumerate-then-build, coverage/quantity floors, anti-laziness anchors, and describing the artifact rather than naming a role.
  • Specifying the bar — acceptance criteria (Must / Must / Must not), small in-prompt rubrics anchored to observable evidence, an explicit good-vs-minimal contrast, and exemplars (the example is the spec).
  • Self-critique passes — a mandatory, named-criteria audit-and-revise step inside the node, with the known ceiling of self-review (use a separate critic node for a hard gate).
  • Node / hand-off contracts — read-this→write-that, structured output, scope fences, and explicit failure paths so pipelines don't drift.
  • Weak-executor tactics — one node = one job, inject the facts, schema-before-task, explicit decomposition as a generative scaffold, format pinning, and repair loops.
  • Anti-patterns & a diagnose-and-fix loop — what to delete, plus a step-by-step routine for when output is still thin.

Using it

This is a skill, so the cleanest path is to drop it where Claude Code discovers skills:

# Global (available in every project)
git clone https://github.com/blueif16/agentic-prompt-design.git \
  ~/.claude/skills/agentic-prompt-design

# …or project-scoped
git clone https://github.com/blueif16/agentic-prompt-design.git \
  .claude/skills/agentic-prompt-design

Once installed, Claude Code surfaces it automatically and you can invoke it with /agentic-prompt-design. It triggers whenever you're writing or rewriting a prompt that another model executes — a workflow/pipeline node, an agent()/sub-agent prompt, a system prompt for a coding/generation/analysis agent — especially when the executor is cheap/weak and output keeps coming back minimal, or when you need to put an acceptance bar / rubric / "definition of good" into the prompt.

You can also just read SKILL.md directly — it stands on its own as a checklist-driven playbook.

Provenance

Synthesized from vendor docs (Anthropic, OpenAI), peer-reviewed work (Self-Refine, SLM-scaffolding), and strong practitioner playbooks (research date 2026-06). Every technique traces to a cited source in references/research-and-sources.md. Inline tags: [O] official vendor · [R] research · [P] practitioner.

Ver o original no GitHub

Bom para

Relacionados