AI_ONBOARD View repository
@RSTHRIVES Portable Agent Workflow Framework

One request. The right context. A verified result.

Across Codex, Claude Code, and OpenCode, stop rebuilding your AI workflow for every tool. AI_ONBOARD connects durable project instructions, focused skills, and independent review into one legible path from intent to outcome.

Route a request

First-class support

  • Codexnative context + TOML
  • Claude Codeimport + settings
  • OpenCodenative context + JSON
68
portable skills
6
independent agents
3
composable layers

Context drift is an architecture problem

Every AI tool asks you to start over.

Tool-specific instruction files drift. Reusable procedures get pasted into every repository. Reviews happen in the same context that authored the work. The result is more configuration, less trust, and no clear route from a request to “done.”

  1. 01

    Instructions fragment

    Project facts split across AGENTS.md, CLAUDE.md, rules, prompts, and memory.

  2. 02

    Capabilities collide

    Large bundles load every procedure at once, even when one narrow skill would do.

  3. 03

    “Done” becomes a guess

    The author checks its own work, compile success substitutes for runtime evidence, and release state stays ambiguous.

Three layers, one shared workflow

Put knowledge where it earns its context.

01 / ALWAYS

Project context

AGENTS.md holds durable facts: architecture, commands, conventions, and safety boundaries.

Loaded every session
02 / ON DEMAND

Focused skills

SKILL.md packages one repeatable job and loads only when the request needs it.

Loaded by intent
03 / DELEGATED

Independent agents

Research, review, security, design, accessibility, and verification run in isolated context.

Loaded for separation

Portable does not mean disposable

Install deliberately.
Upgrade without surprises.

Select only the capability profiles a project needs. AI_ONBOARD records the desired state, locks the exact source, protects anything the user has changed, and can signal when fixes land.

  1. 01 / DECLARE

    Choose the surface

    ai-onboard.json

    Harnesses, capability profiles, agents, configs, and optional manual foundations.

  2. 02 / RESOLVE

    Lock what landed

    .ai-onboard.lock.json

    Revision, checksums, ownership, and managed configuration keys stay inspectable.

  3. 03 / PROTECT

    Preserve user work

    .ai-onboard/conflicts/

    Modified files are never overwritten or removed; incoming conflicts are staged for review.

  4. 04 / SIGNAL

    Know what changed

    .ai-onboard/update-status.json

    Opt-in checks classify security, fix, feature, and maintenance releases without auto-upgrading.

Install a focused surface python3 …/ai_onboard.py install --profile core --profile product
Check and cache release status python3 .ai-onboard/bin/ai_onboard.py upgrade --check --cache --json
Protect Git commit identity python3 .ai-onboard/bin/ai_onboard.py check-git
Invoke from an agent core → $check-ai-onboard-updates · --notifications → /ai-onboard-update · Codex copy → /prompts:ai-onboard-update
Prove what cleanup would do python3 .ai-onboard/bin/ai_onboard.py uninstall --dry-run
Smoke-test every first-class harness python3 scripts/test_deployments.py

Route work through the framework

What are you asking the agent to do?

Choose a request from a category. The route shows which orchestration skill owns the work, which specialists join only when relevant, and which independent lenses close the loop.

Create & design

Build & ship

Research & assess

Request

“Build a clear, conversion-oriented project website.”

  1. 01
    Context

    Read the project contract

    AGENTS.mdbrand evidencerelease boundary
  2. 02
    Orchestrator

    Own the strategy-to-launch loop

    design-and-build-website
  3. 03
    Specialists

    Add only the expertise the route needs

    define-brand-foundationdesign-product-contentaudit-accessibility
  4. 04
    Independent lenses

    Challenge the implementation

    design-reviewaccessibility-reviewverifier
  5. 05
    Evidence

    Prove the outcome in a real browser

    test-browser-workflowsresponsivekeyboardruntime
End result

A tested static site, independently reviewed and ready for an authorized publish.

The shared delivery loop

Specialists change. The standard of done does not.

01

Understand

GOAL and GRILL are original AI_ONBOARD skills. They are manual opt-ins: invoke either only when the user explicitly requests that workflow; otherwise skip them.

02

Compose

Load one orchestrator and the smallest set of bounded specialists required.

03

Build

Implement a coherent vertical slice and preserve accepted decisions across phases.

04

Challenge

Use independent design, accessibility, security, or correctness review where risk warrants it.

05

Verify

Run real checks and task scenarios. Failed evidence loops back to implementation.

06

Release

Cross the external-action gate, smoke-test the result, and report exactly what shipped.

Generated from the repository

68 skills. No mystery box.

Search the live catalog. Every entry links to its canonical SKILL.md, so the website and the agent-readable source stay inspectable.

Loading catalog…

The outcome

Less agent theater.
More evidence.

The final response is not “should work.” It is a compact handoff: what changed, which decisions governed it, which checks ran, what independent review found, what shipped, and what still needs authority.

Install the framework
verified-result.md ● ● ●
Outcome
Published project website
Context
AGENTS.md + repository conventions
Skills
design-and-build-website → audit-accessibility
Review
design-review · accessibility-review · verifier
Evidence
build ✓ browser ✓ keyboard ✓ responsive ✓
Release
GitHub Pages · smoke-tested