OPEN SOURCE · LOCAL-FIRST · SOURCE-BACKED

Build prompts that fit the way you actually work.

openPrompting combines model-specific prompting guidance, coding-tool context and behavior, and task-specific structure into a prompt that fits your setup.

Choose the AI model you’re using, the tool you’re working in, and what you’re trying to do. openPrompting brings the relevant guidance together — without calling another AI.

No API key · No model calls · No backend

§ 01 — HOW IT WORKS

One prompt doesn’t fit every setup.

Different AI models respond differently to instructions. Coding tools manage context and project rules in different ways. And planning a feature needs a different structure from fixing a bug.

openPrompting combines all three into guidance for the setup you’re actually using.

Guidance for this setup

GPT-6 Astra · Codex · Planning

Model — GPT-6 Astra

Model guidance
  • State the target outcome, acceptance criteria, constraints, and relevant context.
  • Specify required output structure and genuine invariants.
  • Define when to ask for missing evidence, when a reasonable assumption is allowed, and when the task is complete.
Open GPT-6 Astra entry →

Harness — Codex

Harness behavior
  • Put durable repository instructions in a scoped AGENTS.md.
  • Keep project instructions concrete, current, and compatible with repository tooling.
  • Include actual validation commands and boundaries the agent must preserve.
Open Codex entry →

Task — Planning

Task structure
  • Inspect the current system, relevant files, configuration, tests, and constraints before prescribing a change.
  • Separate known requirements and observed facts from assumptions, decisions, and open questions.
  • Identify the smallest sound implementation path, its dependencies, owners or handoffs, and the points where the plan should be revisited.
Open Planning entry →
  • ModelHow this AI responds best to instructions.
  • HarnessHow your coding tool handles context and project instructions.
  • TaskWhat structure this kind of work needs.

Selections above are choices, not rankings. openPrompting does not rank models or imply that any combination is objectively superior. Profiles are user-defined.

§ 02 — Catalog

A browsable index of the current knowledge base.

Every entry is a versioned document in the repository with structured evidence and a verification date. New entries appear here automatically at build time.

§ 03 — Real entry

What a knowledge entry actually contains.

Pulled directly from knowledge/harnesses/cursor.md at build time — not written for this page.

Harness

Cursor

Cursor is an editor and CLI coding-agent harness. These notes cover its rules, context, and permission behavior rather than the underlying model.

  • ProviderCursor
  • Statusactive
  • Evidence4 claims · 3 sources
  • FreshnessCurrent · verified Sep 17, 2026
Instruction files
AGENTS.mdCLAUDE.md.cursorrules

Recommended

  • Put simple project-wide instructions in AGENTS.md; use .cursor/rules/*.mdc when the guidance needs file scoping, intelligent attachment, or manual invocation.
  • Use the .mdc extension and explicit rule frontmatter (description, globs, and alwaysApply) for project rules; keep rules focused and version-controlled.
  • Use nested AGENTS.md files or scoped rules for genuinely local conventions, and review the resulting instruction set when a task crosses directories.
  • Treat CLI shell and file-read permissions as part of the setup. Grant only the command and path access the workflow needs, especially before using non-interactive mode.

Evidence

  • officialCursor project rules are version-controlled .mdc files under .cursor/rules and can be applied always, by file pattern, intelligently, or manually.
    Rules ↗
  • officialCursor supports AGENTS.md project instructions, including more specific instructions in nested directories.
    Rules ↗
  • officialCursor CLI uses the IDE rules system and reads AGENTS.md and CLAUDE.md at the project root when present.
    Using Agent in CLI ↗
  • officialCursor CLI permissions can separately control shell commands and file reads through global or project configuration.
    Permissions ↗
§ 04 — Local-first

Normal operation touches nothing outside your machine.

openPrompting resolves guidance from a local knowledge base. After installation it needs no account, no API key, no model call, no hosted backend, and no network connection. Setup is the only command that writes project state; doctor is read-only.

  • No account
  • No API key
  • No model call
  • No hosted backend
  • No telemetry
  • No runtime network
§ 05 — Evidence

Every claim has a source and a class.

Each knowledge entry carries structured evidence metadata: what class of evidence supports a claim, which source it came from, and when that source was last checked. Every entry page surfaces that metadata.

official

Official

Supported directly by first-party provider or harness documentation.

tested

Tested

Reproduced by maintainers with a procedure recorded in the entry.

community

Community

Useful practice without first-party confirmation.

legacy

Legacy

Historical or migration advice, not treated as the current default.

§ 06 — CLI

Six commands. Nothing hidden.

The command line configures, resolves, generates, diagnoses, and compares. It never calls a model — it reads the same knowledge you are browsing here.

help
Show the command list and common examples. No config required.openprompting help
setup
Create or update .openprompting/config.yml; --check validates without writing.openprompting setup
guide
Display resolved model and harness guidance: recommendations, cautions, evidence, sources.openprompting guide --profile reviewer
new
Render a deterministic prompt skeleton for a task. No model call, no invented facts.openprompting new planning
doctor
Run read-only local diagnostics. Never executes project scripts.openprompting doctor
compare
Compare two profiles, setups, models, or harnesses from documented evidence. No winner inferred.openprompting compare builder reviewer

Full selectors, resolution rules, and exit behavior: Docs →

§ 07 — Start

From install to a real task prompt in four steps.

  1. 01
    Install
    npm install --global openprompting
  2. 02
    Configure
    openprompting setup
  3. 03
    Read guidance
    openprompting guide
  4. 04
    Start a task
    openprompting new planning
§ 08 — New here

Agentic coding, in one paragraph.

A model reasons and generates; a harness gives it repository access, tools, and instruction files. Together they form a setup, which you can name as a profile. Good prompting is rarely magic words — it is giving the agent goals, context, requirements, constraints, acceptance criteria, and validation.