Reusable prompt patterns for AI-assisted software development.

This repository is a prompt toolkit for software work.

Use it to:

• start from one entry point
• load only the guides that matter for the task
• keep shared rules stable across model families
• handle memory clearly when the host supports it
• reuse specialist agents for broader audit or evaluation work

If you are new to this repository, do not start by reading every guide.

Use this short path instead:

1. Start with this README when you are routing the guide set by hand. Use .github/prompts/prompt-bootstrap.prompt.md only when your host wants a ready-made routing prompt.
2. Pick one concrete task, such as debugging a Python service or running a release audit.
3. Load core/shared-contract.md plus one main workflow guide for that task.
4. Add model-adapter, memory, or agent files only if the task actually needs them.

Good first combinations:

• Debugging: core/shared-contract.md + development/debugging.md
• Code review: core/shared-contract.md + development/code-review.md
• Release audit: core/shared-contract.md + development/exhaustive-review.md + core/production-ready-check.md

Treat this repository as a routing system, not a pile of files.

Start with this README. Its job is to help you choose the smallest useful set of guides for the task.

Do not load everything by default. Use one small guide bundle per task.

This repository is designed for selective loading.

If you package it for an AI tool, search system, MCP server, or skill format:

• do not merge the whole repo into one default prompt payload
• start from the router and load only the files needed for the current task
• keep shared rules separate from task workflows, prompt templates, and specialist agents

The repo now includes lightweight metadata for that purpose:

• meta/asset-manifest.yaml: defines packaging groups and export rules
• meta/guide-index.yaml: describes what each guide is for and when to load it

For packaging or host integration, treat the entry point as README.md plus the meta/ folder.

Validate the metadata with:

python3 scripts/validate_metadata.py

This validator checks that the metadata files point to real repository files, that guide markdown files keep the required frontmatter, and that internal markdown links still resolve.

Use this repository in two simple steps:

1. Use this README or the bootstrap prompt to choose the smallest useful set of guides.
2. Run the real task with only those guides loaded.

If the consumer is a tool rather than a person, pair this README with the metadata files under meta/ before selecting guides.

Do the routing step once per task. Repeat it only if the task changes in a meaningful way or the earlier setup is no longer available.

Task: debug a regression in a Python service after a refactor.

Small bundle:

• core/shared-contract.md
• development/debugging.md
• development/context-management.md only if the session is large or memory matters
• one prompt wrapper only if the host needs it

Why this bundle:

• the shared contract keeps baseline rules stable
• the debugging guide gives the task method
• context guidance is optional support, not default baggage

What not to load:

• release audit guides
• ML bootstrap guides
• specialist agents unless the task expands into a broader workflow

The point is to end up with a task-sized bundle, not the whole repository.

If you are assembling context by hand, keep the order simple:

1. start with prompt-bootstrap.prompt.md or this README
2. load core/shared-contract.md
3. add memory or model-adapter guidance only if the task needs it
4. load one main workflow guide from development or setup
5. add a prompt template or specialist agent last

The full routing logic lives in prompt-bootstrap.prompt.md.

The repository is organized into layers. Use this section as a map.

core/shared-contract.md holds the rules that should stay the same across tasks and model families.

If the host supports memory, core/memory-contract.md explains what should be remembered and where it belongs.

Use development/model-adapters.md when prompt structure should change for a model family without changing the task itself.

Choose the task-specific guide from development or setup.

Examples:

• development/context-management.md for session context and memory decisions
• development/debugging.md for debugging
• development/code-review.md for review work
• development/test-generation.md for tests
• development/git-workflow.md for commits, PRs, and releases
• core/production-ready-check.md for release readiness

For host setup, hooks, commands, external tools, and other specialized workflows, see the rest of the files under development.

Use the files under .github/prompts when you want a ready prompt wrapper around the selected guide set.

• prompt-bootstrap.prompt.md is the main router
• gpt-task-template.prompt.md is the GPT task template
• claude-task-template.prompt.md is the Claude task template

Use the files under .github/agents when the job is broader than a single prompt or needs a dedicated workflow.

Examples:

• public-release-auditor.agent.md for high-recall release audits
• fix-and-recheck.agent.md for fixing findings and rechecking them
• prompt-evaluator.agent.md for prompt comparisons
• model-adapter-designer.agent.md for adapter design work

Use an agent when you need coordination across a broader workflow. Use a prompt when you need a focused task setup.

• core: shared rules and release gates
• development: task-specific workflows for coding, reviews, tests, debugging, context, memory, git, and audits
• setup: project setup, architecture, documentation, and reproducibility guidance
• .github/prompts: prompt entry files and model-family templates
• .github/agents: specialist agent definitions

This repository is licensed under CC BY 4.0.

You may use, adapt, and share the material, including commercially, as long as you provide attribution.

If you reuse or adapt this repository, credit:

• byteshiftlabs
• https://github.com/byteshiftlabs/ai-dev-prompts

If you made changes, say that clearly in your attribution.

Example:

Based on AI Development Prompts by byteshiftlabs
https://github.com/byteshiftlabs/ai-dev-prompts
Used under CC BY 4.0. Changes were made.