Template Library

This is the product layer of the Agent Workflow Blueprint: 100 specialist source modules and deterministic, copy-ready compositions for controlling coding-agent work with evidence, specialist review, and honest final status.

What is included

• 40 v3 scenario-specific prompts for execution, review, UI, accessibility, testing, security, documentation, and release control.
• 30 executable skills for repeatable agent workflows.
• 30 enforceable contracts for acceptance, rejection, and evidence gates.

Every asset references the concise shared GOVERNANCE-KERNEL.md and the normalized SPECIALIST-CONTROLS.md registry. Reused controls exist once and are included once during composition. Every prompt also includes concrete inputs, five to ten domain instructions, decision gates, evidence, recovery paths, rejection conditions, an output record, and a worked example. Skills and contracts retain their executable procedures and hard gates. The controlled final statuses remain verified, partially verified, not verified, blocked.

Use the system this way

1. Open an asset's single dialog, review its goal, applicability boundary, and research references, then copy only the agent-ready body with its compact copy control; or compose selected asset IDs with python3 -S scripts/compose_assets.py --asset <id>.
2. Choose the smallest prompt that matches the task.
3. Attach the matching skill when the work needs a repeatable procedure.
4. Attach the matching contract when the result needs acceptance or rejection language.
5. Run repository verification from requirements.toml where applicable.
6. Do not claim more than the evidence proves.

Fast paths

• Start with START-HERE.md for the shortest route through the library.
• Use STARTER-PACKS.md for bundled prompt + skill + contract combinations.
• Use QUALITY-RUBRIC.md to evaluate agent output.
• Use HUMAN-AI-QUALITY-STANDARD.md to check sociotechnical, human-AI, trust-calibration, and ML-evaluation quality.
• Use RESEARCH-BASIS.md to understand the standards and behavioural controls behind the library.
• Use CODEX-USAGE-GUIDE.md to apply the assets in Codex.
• Use EXAMPLES.md to see weak output rejected and stronger output accepted with limits.

Human-readable catalogue

Open CATALOGUE.md for all 100 assets grouped by type and category.

Source modules and copy-ready assets

The Markdown files under prompts/, skills/, and contracts/ are authoring modules. They contain specialist work directly and reference only genuinely shared controls. Do not provide a source module alone unless the recipient already has its kernel and registry dependencies. The website presents one dialog with applicability outside the code region, the complete executable prompt/skill/contract body as literal Markdown code, one compact copy control beside that body, and research plus implementation references outside the copied content. scripts/compose_assets.py produces the same body-only output by default; pass --with-references only when creating a review artefact rather than an agent paste.

Machine-readable catalogue

Open assets.json for site integration, version and risk metadata, required inputs, expected outputs, governance profiles, evaluation-case mappings, ordered specialist-control references, curated research labels, and automation. research-sources.json is the normalized citation registry used to resolve those labels to exact scholarly or standards sources.

Usage rule

Do not paste these assets blindly. Select the smallest asset that matches the failure mode, then combine it with AGENTS.md, docs/engineering/workflow.md, and the relevant engineering contract.

Research-informed control layer

This library includes a source-mapped control layer:

• RESEARCH-DNA.md — maps CoT, Tree-of-Thoughts, ReAct, Least-to-Most, Self-Consistency, Reflexion, System 2 forcing, and traceability into product behaviour.
• PROMPT-PATTERN-MATRIX.md — shows which research pattern to use for each task type.
• SCIENTIFIC-CONTROL-CHECKLIST.md — a pre-merge checklist for evidence-backed agent work.
• HUMAN-AI-QUALITY-STANDARD.md — a source-mapped quality bar for prompts, skills, contracts, rules, generated projects, and agent outputs.

The product does not rely on prompt length alone. It relies on observable controls: decomposition, branch comparison, environment observation, independent verification, correction loops, traceable acceptance, and failure disclosure.

Research-informed architecture layer

These documents map published techniques to inspectable repository controls without claiming that the repository independently validates those techniques:

• SCIENTIFIC-DNA.md — maps CoT-safe public reasoning, Tree-of-Thoughts-style branching, ReAct loops, least-to-most decomposition, self-consistency, self-refinement, process supervision, constitutional/contract critique, and traceability to repository behaviour.
• PROMPT-ARCHITECTURE.md — defines the v3 source contract, composition order, common handoff, and evidence architecture.
• TRACEABILITY-MATRIX.md — provides the requirement-to-evidence matrix used by contracts and release gates.
• SYSTEM-2-PROMPTING-GUIDE.md — explains how to use “System 2” as a deliberate-work metaphor without overclaiming human-like cognition.

The key rule is: ask agents for a public decision record and evidence trail, not hidden chain-of-thought.