prompt·architect

prompt·architect

An open-source prompt-engineering for tool-calling AI agents.

what it does

The operator lens · how it works

You describe the task. It compiles the prompt.

A task goes in. Prompt Architect interviews you briefly, picks the right structure, applies the field's documented best practices section by section, audits the draft, and hands back a finished prompt — with a test that proves it works. Works with Claude Code, Gemini CLI, or any API runtime.

A prompt, compiled as an operator stack — hover a section to apply the chain up to that point
Order is structural: Â B̂ ≠ B̂ Â — swap two sections and you get a different answer.

The core idea

A prompt doesn't find the right answer hidden in the model. Each section changes what the model will say — and order, combination, and strength are the real design decisions.

Concretely: the role you assign projects the model into one way of working; constraints narrow it; examples teach patterns no rule states; the output schema fixes the form. The framework calls each section an operator and gives it a measurable profile — what it amplifies, what it suppresses, how strongly, and where in the order it belongs. Every operator is drawn from documented prompt- and context-engineering practice, not invented rules.

How it works

What happens when you use it.

Two modes: build a prompt from scratch, or hand it one you already have.

01

You ask

“Help me design a prompt for [task] — use the prompt-architect skill.” Works in any tool-calling runtime.

02

It interviews

A few targeted questions. The structure is inferred from your request — you're never quizzed about things you already said.

03

It drafts & audits

Every section written and placed deliberately, then checked — structure, section interactions, honest scoring.

04

You receive

The finished prompt, paired artifacts when the shape needs them, usage notes for your runtime, and a runnable verifier.

Mode 1 · default

From scratch

You describe a task. The framework infers the shape, runs a brief operator-aware interview, drafts, audits, scores itself honestly, and delivers.

  1. Shape inference — the structure is read from your request, not asked
  2. Adaptive interview — capped, co-authoring on hard decisions
  3. Drafting — every section placed deliberately
  4. Audit — structure and section-interaction checks
  5. Honest self-evaluation — three axes, named gaps
  6. Delivery — prompt, verifier, usage notes
Mode 2 · existing prompts

Refactor

You paste a prompt you already have. The framework recovers its intent, then does exactly the level of work you ask for — never a silent rewrite.

  • AUDIT — diagnostic report: findings, severity, fixes. No edits.
  • TARGETED — surgical edits for a named symptom, via a symptom-to-cause table.
  • REWRITE — full redraft with explicit diff and tradeoff analysis; the original is preserved.

Anatomy of a generated prompt

A finished prompt is a stack of sections, each doing one job. Inspect two real ones.

Pick a situation. The framework assembles a different set of sections for each shape — and treats every section as an operator: something that changes what the model will do, with a strength, a profile, and a correct position in the order. Select a section to inspect it.

Worked example Agentic loop · seed brief
seed-brief.md Shape 1 · agentic loop

Seven shapes

Every prompt has a structural archetype. The framework infers it, then builds to its spine.

Shape is orthogonal to domain. The same seven shapes apply whether the work is engineering, research, writing, or ops — domain changes the content of sections, not which sections appear. The framework reads your task and infers the shape, so the interview stays short.

S1 · AGENTIC LOOP

Long-horizon work

Multi-checkpoint projects with a memory file and human-in-the-loop handoffs.

Inspect a real one ↗
S2 · ONE-SHOT

Single complex call

Classification, extraction, transformation, generation — done in one pass. Cognitive tools attach here when the task is genuinely hard-reasoning.

Inspect a real one ↗
S3 · WORKFLOW / CHAIN

Staged pipeline

Coordinated calls with gates between stages and aligned inter-stage schemas.

S4 · SUB-AGENT / TOOL

Consumed by an agent

Output designed for another model to parse, not a human to read.

S5 · SYSTEM PERSONA

Stable assistant

Identity that holds across unknowable inputs, with layered capability lockdown.

S6 · LLM-AS-JUDGE

Graders & scorers

Calibrated rubrics with bias controls and reasoning forced before the score.

S7 · AGENT TEAM

Multi-agent systems

Four coordination topologies — orchestrator-workers, parallelization (sectioning & voting), evaluator-optimizer — composed from the six single-agent shapes above, with verified interface contracts between roles.

What it applies — and where it's from

Every technique traces to published work.

The framework absorbs a documented technique only when it carries measurable consequences — vocabulary without measurable effect is what its audit rejects. Here is the core catalog, each practice paired with the work it comes from. The complete annotated set, including what was evaluated and rejected and why, is in CITATIONS.md.

  • Context engineeringBudget every token; load critical context early and persist it across turns.
    SourceEffective context engineering for AI agentsAnthropic
  • Few-shot · abstention · prefillingCurated examples; explicit permission to say “I don't know”; prefilling to lock the output's form.
    SourceBest practices for prompt engineeringAnthropic
  • Cognitive toolsModular reasoning steps shipped as real tool definitions, not prose.
    SourceEliciting Reasoning in Language Models with Cognitive ToolsEbouky, Bartezzaghi, Rigotti · IBM Research Zurich · arXiv:2506.12115central
  • Multi-agent topologiesOrchestrator–workers, parallelization (sectioning & voting), evaluator–optimizer.
    SourceBuilding Effective AI AgentsAnthropic — four coordination topologies
  • Parallel agent teamsCoordinating a team of parallel agents against one codebase.
    SourceBuilding a C compiler with a team of parallel ClaudesCarlini · Anthropic Engineering
  • Memory files & lockdownA CLAUDE.md that outlives the context window; layered capability lockdown; “never delegate understanding.”
    SourceClaude Code internal architectureAgentTool, Explore/Plan agents, lockdown
  • The operator lensThe organizing metaphor: sections as operators on the model's semantic state.
    SourceA quantum semantic framework for natural language processingAgostino, Le Thien, Apsel, Pak, Lesyk, Majumdar · arXiv:2506.10077organizing metaphor
  • Mechanistic basisWhy operators compose the way they do — the symbolic mechanisms behind abstract reasoning.
    SourceEmergent Symbolic Mechanisms Support Abstract Reasoning in LLMsYang, Campbell, Huang, Wang, Cohen, Webb · ICML 2025 · arXiv:2502.20332mechanistic basis