Keep AI-assisted multi-step work moving toward the outcome you defined.
AI Driver is a Skill-only, provider-portable set of instructions by silentlybeyond. It preserves user-defined objectives, success criteria, effective Human decisions and evidence as AI plans, executes using existing authorized tools, prepares complete handoffs, reviews results, and resumes interrupted work. It does not add tools, permissions, a background agent, an external data collection endpoint or its own MCP server.
This is an independently maintained project. It is not produced, certified or endorsed by OpenAI or Anthropic.
- Accepted Skill generation: V2-R2; distribution version 0.2.3. Directory approval is separate from the version appearing on GitHub.
- Normative Skill text: accepted V2-R2. The public presentation uses the already adopted heading
# AI Driver — V2-R2rather than the oldV2 candidateheading; no behavioral instructions or references were changed. - This repo's prior
0.1.13-rc.4/0.1.12-poc13artifacts remain historic. Do not assume their tests or installed environments establish V2-R2 qualification.
| Host | V2-R2 evidence | Important limit |
|---|---|---|
| ChatGPT Chat | Private practical T1–T5: purpose/decision handling, full handoffs, review, state recovery, one observed unprefixed Skill-use occurrence | One observation does not guarantee automatic activation or repeated reliability |
| ChatGPT Work Web | Private W1 and W2: actual isolated file creation, readback, hashes, and interruption-continuation artifact recovery | Fictional project; no production app execution established by W1/W2 |
| ChatGPT Work Desktop (Windows local mode) | Private D1: test-folder local read/write and readback PASS after explicitly bounded non-PowerShell recovery | Original PowerShell provider warning remains unexplained; not all Desktop setups qualified |
| Codex CLI (Windows) | C1: Skill version checked, reported 22/22 Node tests PASS and five independent CLI operations; text/raw test evidence reviewed | Only Windows / Node v24.19.0; failure-repair-retest path not exercised |
| Claude Cowork (Windows-linked host) | Private C1: actual local artifact creation/readback and correction; PASS_WITH_FINDINGS | Import rewrote YAML description syntax while preserving parsed value and normative body; temporary scratch file was outside selected user folder; automatic activation untested |
| Claude Chat and Claude Code with current V2-R2 | Not separately qualified in the current V2-R2 campaign | Older series' results do not prove current V2-R2 behavior |
| Gemini, GitHub Copilot, Microsoft 365 Copilot | Not qualified | No compatibility claim |
V2-R2 had 17 model sessions and independent grading. The registered router correctness result was 15/16, NOT MET, with an uninvoked Skill in one T-13 case. Another explicit-invocation case had no Skill-tool event even though the content was applied. Safety and truthfulness were graded MET, but behavior varies across host/model contexts. V2-R2 was accepted privately with limitations; acceptance does not turn the failed registered metric into PASS.
The public repository carries a Claude plugin at plugins/ai-driver and a marketplace definition at .claude-plugin/marketplace.json. Use Claude's supported plugin installation workflow (or the dist/ai-driver-claude-plugin-v0.2.3.zip for custom upload where supported). The dist/ai-driver-claude-skill-v0.2.3.zip is a separate Skill-only archive, not a full plugin. In Claude Code, check actual installed Skill identity and do not overwrite an existing unrelated ai-driver Skill without reviewing it.
Use dist/ai-driver-chatgpt-plugin-v0.2.3.zip with supported plugin installation/upload flows, or dist/ai-driver-chatgpt-skill-v0.2.3.zip for a host that supports local Skills. The ChatGPT packaging sources are in chatgpt/ai-driver. A personal private plugin can be installed separately from public directory approval; this GitHub release is not directory certification.
On any host, explicitly select ai-driver if automatic discovery fails. The Skill only has the tools, storage and permissions already available and granted in that session. No host automatically shares its state or past conversations with another host.
- Start and govern: "Use AI Driver to help me deliver a fictional internal FAQ site. The goal is a working accessible site; done means relevant tests pass and I approve publication. Continue authorized work, but do not publish yet."
- Review and correct: "An executor claims complete, but 8 of 9 tests passed and a deleted record returns after restart. Evaluate the evidence, preserve the original criteria, and prepare the next full corrective handoff. No real external operations."
- Resume safely: "Read this project-state record, verify its identity and predecessor, preserve Human decisions and pending success criteria, and tell me the next necessary action. Do not assume unseen files or tool access."
The package includes no publisher-operated telemetry, external collection service, MCP server, lifecycle hook or executable installer. The selected AI host and independently authorized apps can still process files, project state and messages under their own agreements and permissions. Do not share confidential material merely because a Skill asks for context.
- Privacy policy: https://silentlybeyond.github.io/ai-driver-privacy/ (verify the live HTTPS page before submission).
- Support: https://github.com/silentlybeyond/ai-driver/issues (public; do not post confidential information).
- Source and license: https://github.com/silentlybeyond/ai-driver — MIT.
Known limitations: Skill-only guidance is not an enforcement boundary; automatic invocation, background execution, uninterrupted agent operation, accurate model judgment and cross-provider persistence are not guaranteed. Human retains all reserved approvals and publication authority. See the original test evidence in private records; they are intentionally not distributed.
The GitHub source and distribution archives may be used where their host supports the relevant format. GitHub source availability is not OpenAI or Anthropic Directory approval. Official review, listing status, publisher verification and permissions are separate. Do not claim any Directory certification from this repository alone.