IntentDriven

Composable agent skills for Intent-Driven Development (IDD): humans own product intent, ambiguous decisions, and acceptance; agents own planning, implementation, verification, review, evidence, and documentation reconciliation.

Core invariant:

Implementation must not silently redefine intent.

The plugin turns a feature request into a gated chain:

Intent -> Spec -> Domain Knowledge -> Decisions -> Implementation -> Evidence -> Human Acceptance -> Ownership

For one bounded, local, reversible milestone, /intentdriven automatically uses a small-work route with an embedded intent contract. For clear intent and non-trivial agent-led coding, it uses collaborative-implementation. The human approves consequential technical choices. The agent owns routine engineering, milestone commits, verification, and the concise mental-model handoff.

IntentDriven uses DDD as a language and context discipline. Specifications stay concise, assign stable IDs to normative items, and use one central map from each ID to implementation and evidence.

Skills

Group	Skills	Purpose
Intent	intentdriven, map-authority, grill-feature, write-spec, prototype-ui	Discover which local or external artifact owns each truth type, clarify product intent, write business contract, branch into UI prototypes when interaction choices matter.
Decisions / Bridge	reconcile-domain, record-decision, plan-implementation	Keep ubiquitous language, domain context, ADRs, and technical plans aligned with accepted intent.
Collaborative Delivery	collaborative-implementation	Compare real approaches, obtain one risk-based approval, implement autonomously through verified milestone commits, and pause only for consequential deviations.
Execution / Evidence	implement-feature, verify-feature, review-change, prepare-acceptance	Execute the approved plan, map stable specification IDs to evidence, review the result, and package acceptance evidence.
Consistency	detect-drift, resolve-conflict, reconcile-docs	Detect semantic or traceability drift, surface human decision packets, and update docs only after confirmed intent.
Human Ownership	own, own-report, own-visual	Build a mental model for explaining, defending, debugging, modifying, and extending completed AI-assisted work.
Plugin Operations	intentdriven-install-skills	Reinstall this local plugin into Pi, OMP, Claude Code, and Codex with separate evidence for each harness.

Install

skills.sh

Install the whole skill pack:

npx skills add dmltdev/intentdriven

Install only the root workflow skill:

npx skills add dmltdev/intentdriven --skill intentdriven

Replace intentdriven with any skill from the table above to install a narrower phase skill.

Claude Code

Install from GitHub as a Claude Code plugin marketplace:

/plugin marketplace add dmltdev/intentdriven
/plugin install intentdriven@intentdriven

For local development, pass the path to your local clone:

/plugin marketplace add /path/to/intentdriven
/plugin install intentdriven@intentdriven

pi

Install from GitHub with pi's package installer:

pi install git:github.com/dmltdev/intentdriven

Install project-locally, writing the package entry to .pi/settings.json:

pi install -l git:github.com/dmltdev/intentdriven

Run a one-off pi session with the package loaded:

pi -e git:github.com/dmltdev/intentdriven

omp

Install from GitHub as an omp plugin marketplace. See https://omp.sh/docs/plugins for omp's plugin marketplace workflow.

omp plugin marketplace add dmltdev/intentdriven
omp plugin install intentdriven@intentdriven

For local development, pass the path to your local clone:

omp plugin marketplace add /path/to/intentdriven
omp plugin install intentdriven@intentdriven --force

Codex

Install from GitHub as a Codex plugin marketplace:

codex plugin marketplace add dmltdev/intentdriven --ref main
codex plugin add intentdriven@intentdriven

For local development, pass the path to your local clone:

codex plugin marketplace add /path/to/intentdriven
codex plugin add intentdriven@intentdriven

Use

Invoke /intentdriven for automatic route selection. Small work uses one plan with an embedded intent contract. Clear, non-trivial work uses collaborative-implementation. Ambiguous or consequential work uses the full lifecycle. Invoke a narrower phase skill when entering a known phase.

Every completed implementation includes a concise mental-model handoff. Use own, own-report, or own-visual when deeper ownership transfer is useful.

The detailed feature, task, bugfix, and refactor workflow is described in skills/intentdriven/references/development-workflow.md. The shared state and central traceability format are described in skills/intentdriven/references/feature-state.md.

Boundary

• Platform-neutral by default. Linear, Jira, Confluence, Notion, local Markdown, ADRs, tests, schemas, and code can own different kinds of truth.
• Local project docs or agent instructions define each authority: truth type -> authority -> location/link -> write policy -> conflict policy.
• When domain or vocabulary authority is missing and local docs are viable, the plugin recommends a replaceable DDD-lite fallback with a glossary, domain context, and context map.
• Feature specifications stay separate from technical plans. Stable normative IDs connect the specification to implementation and evidence through one central traceability map.
• Drift between documented intent and implemented behavior is information. Agents classify and surface it; they do not silently choose a winner.