Ai Lab Codex Workbench

AI Prompting and Coding Agent Workbench

A public, Windows-friendly workbench for turning vague AI tasks into scoped, reviewable repository work. It combines prompt engineering, coding-agent workflows, source verification, evaluation, reusable skills, local checks, and safe release packaging.

task -> trusted context -> scoped prompt -> agent work -> checks -> diff review -> human approval

This repository is not a model-access service, benchmark authority, leaked prompt mirror, or shortcut around product limits. Model names, prices, and interfaces change quickly; dated guides distinguish official facts, vendor claims, independent results, and unresolved gaps.

Start Here

Choose one route. The repository has many references because apparently one Markdown file is never enough, but the first action should still be obvious.

| Route | First result | Start | | --- | --- | --- | | Run a task | Produce and verify one reviewed report in 5 to 15 minutes | [First reviewed agent task](examples/first-reviewed-agent-task/README.md) | | Learn prompting | Build a prompt contract, then follow a 15-minute or one-hour track | [Prompting OS first success](docs/prompting-os/00-first-success.md) | | Contribute | Run local checks and open a scoped, reviewable change | [Contributing guide](CONTRIBUTING.md) |

The product promise is simple: turn a vague AI task into a bounded work order, a reviewable artifact, and evidence that the result passed its declared check.

Quick Start

Requirements:

• Git;
• Python 3.9 or newer;
• Windows PowerShell 5.1 or PowerShell 7+;
• GitHub CLI only for optional GitHub automation.

git clone https://github.com/Yaked1/ai-lab-codex-workbench.git
cd ai-lab-codex-workbench
.\scripts\local_check.ps1

There is no service to start, port to open, or dependency bundle to install. The repository uses Python's standard library for its checks.

If local script execution is blocked, apply a process-only policy and retry:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
.\scripts\local_check.ps1

Run checks directly when you need full failure output:

python scripts/repo_health_check.py
python scripts/safe_autofix.py --check
python -m unittest discover -s tests
git diff --check

Prompting And Agent Mastery

Prompt contracts

A dependable prompt behaves like an interface. It states the outcome, inputs, constraints, allowed actions, output format, evidence, and failure behavior.

Objective:

Context:

Included scope:

Excluded scope:

Success criteria:

Final report:

The [prompt engineering guide](docs/guides/comprehensive-prompt-engineering-guide.md) explains task framing, examples, output schemas, ambiguity, tool use, and iteration. The [agent-prompting craft guide](docs/guides/prompting-ai-coding-agents.md) turns that contract into work orders, while [coding-agent power tips](docs/guides/coding-agent-power-tips.md) cover tool-specific usage. The [Prompting OS](docs/prompting-os/README.md) turns those ideas into a modular system with model drivers, context rules, evaluation, security, maintenance, and reusable patterns.

Use [prompting references](docs/guides/prompting-references.md) to find official vendor and evaluation sources. Use the [source-inspired curriculum](docs/guides/source-inspired-prompting-curriculum.md) to turn public sources into original exercises without copying proprietary or leaked prompts.

Coding-agent workflow

For repository changes:

1. Read AGENTS.md and inspect the live branch and diff.
2. State the measurable outcome and exact scope.
3. Load only the files and sources needed for the decision.
4. Plan risky or multi-file work before editing.
5. Make the smallest change that satisfies the task.
6. Run focused checks, then the strongest realistic repository checks.
7. Read the final diff instead of trusting the agent summary.
8. Commit only related files and leave unrelated work untouched.

Reusable work orders live under [prompts/](prompts/). The most useful starting templates are:

• [documentation update](prompts/codex/docs-update.goal.md);
• [bug fix](prompts/codex/fix-bug.goal.md);
• [feature implementation](prompts/codex/implement-feature.goal.md);
• [pull-request review](prompts/codex/review-pr.goal.md);
• [Claude Code documentation review](prompts/claude-code/review-docs.goal.md).

Claude Code can expose a Markdown work order as /goal when it is installed under .claude/commands/goal.md. Keep the command file small and point it to the reviewed prompt source instead of maintaining divergent copies.

Source-grounded writing

Current product facts require live verification. Use this order:

1. official documentation;
2. official release notes and pricing pages;
3. benchmark maintainers and independent labs with a method;
4. reputable reporting;
5. community posts only as leads.

Every dated research deliverable should include an answer, source-tiered findings, uncertainties, sources, and method. Do not promote vendor benchmark tables to independent results. Do not use X, Reddit, screenshots, or leaked prompts as sole evidence.

The [source policy](docs/research/source-policy.md), [publication policy](docs/publication-policy.md), and [source-grounded writing lab](docs/prompting-os/29-source-grounded-writing-lab.md) contain the full rules.

Evaluation

Evaluate the workflow outcome, not how polished the response sounds.

| Dimension | Example measure | | --- | --- | | Correctness | Tests pass; citations support the claims. | | Scope | Only allowed files or systems changed. | | Reliability | The same prompt succeeds across representative cases. | | Efficiency | Human correction time, latency, retries, and token cost. | | Safety | No secrets, unsafe commands, unsupported claims, or hidden writes. | | Maintainability | Another contributor can understand and verify the result. |

Use deterministic tests for deterministic behavior. For prompt quality, keep a small evaluation set with pass thresholds and known failure cases. See the [evaluation cookbook](docs/prompting-os/10-evaluation-cookbook.md), [prompt-quality rubric](docs/prompting-os/evals/prompt-quality-rubric.md), and [evaluation datasets guide](docs/prompting-os/27-prompt-evaluation-datasets.md).

Current Model Guides

[](https://yaked1.github.io/ai-lab-codex-workbench/site/model-media.html#fable-official)

Official Anthropic video. Click the image to watch inside this repository's GitHub Pages player. The chart above is this repository's independently drawn view of cited Artificial Analysis data.

The dated [model and interface guide](docs/guides/current-models-and-interfaces.md) covers:

• GPT-5.6 Sol, Terra, and Luna across standard ChatGPT, ChatGPT Work, Codex,

and the API;

• verified effort controls, including the difference between max, ultra,

and API multi-agent behavior;

• Claude Fable 5 across chat, Cowork, Claude Code, desktop, and API surfaces;
• Fable effort levels and why Ultracode is an orchestration mode;
• Grok 4.5 High in Grok Build;
• Gemini 3.5 Flash Standard/Extended UI guidance and API thinking levels;
• base pricing, caching, long-context thresholds, rollout gaps, and sources.

The [Fable 5 vs GPT-5.6 Sol comparison](docs/guides/fable-vs-sol.md) separates Artificial Analysis, Arena, METR, vendor tables, and community sentiment. It does not relabel Sol Max benchmark results as Sol Ultra results.

The [live audio guide](docs/guides/live-audio-and-translation.md) explains GPT-Live-1's vendor-described full-duplex design and Gemini 3.5 Live Translate's product and API paths.

The [frontier models and multimodal systems essay](docs/guides/frontier-models-and-multimodal-systems-2026.md) adds plan-by-plan GPT-5.6 controls, effort guidance for all three tiers, Artificial Analysis comparisons, Grok and Muse, and a source-checked map of current audio, image, and video model families. Anthropic's current Fable promotion terms extend included access through July 19, 2026 at 11:59:59 PM PT. Fable does not disappear after that deadline; it remains available using usage credits. Its [video research pack](docs/research/video-research-pack-2026-07-11.md) separates verified embedded videos from discovery searches that are not factual evidence.

The [model and effort prompting pack](docs/guides/model-prompting/README.md) turns that inventory into per-model work orders: Sol/Terra/Luna efforts (including Desktop Light and Work Ultra evidence boundaries), Fable/Opus Extra vs xhigh/Ultracode, Grok Build High default, Muse Spark, Gemini Flash, Live audio/translate, and image/video families (GPT Image 2, Nano Banana, Omni, Seedream, Muse Image; Muse Video marked coming soon and unavailable for production prompting). The pack's [effort evaluation playbook](docs/guides/model-prompting/effort-evaluation-playbook.md) tests whether higher effort is worth its cost, while its [source ledger](docs/guides/model-prompting/sources-and-observations.md) separates official access from local and user-observed menus.

Each guide now includes an original, attributed visual and links to watchable official demos where available. The [media provenance ledger](docs/research/model-media-provenance-2026-07-11.md) explains why third-party benchmark charts and X screenshots are linked rather than copied.

These files are dated snapshots. Check the linked primary sources before making a purchase, deployment, or account-access decision.

Offline Site and Starter

• [Offline discovery site](docs/site/index.html) offers the same run, learn, and

contribute routes plus local search with no external assets.

• [Copyable starter](starter/README.md) is the small fork/copy boundary for a

task template, safety rules, one evaluation, and one example.

• [Model media](docs/site/model-media.html) is explicitly network-required;

the remaining core site is offline-capable.

• [GitHub owner settings and ref retirement](docs/maintenance/github-owner-settings.md)

separates committed policy from live GitHub controls and records non-destructive branch dispositions.

Repository Map

| Area | Purpose | | --- | --- | | [docs/codex](docs/codex/) | Beginner-to-maintainer Codex workflows. | | [docs/guides](docs/guides/) | Prompting, model, research, and coding-agent guides. | | [docs/prompting-os](docs/prompting-os/) | Modular prompt-system design and operations. | | [docs/tools](docs/tools/) | Practical coding-tool guides and comparisons. | | [docs/image-generation](docs/image-generation/) | Image-model concepts, prompting, and hardware tradeoffs. | | [docs/hermes](docs/hermes/) | Hermes Agent setup, prompting, skills, and safety. | | [docs/workflows](docs/workflows/) | Task lifecycle, public safety, and broad repo work. | | [prompts](prompts/) | Reusable goal-style work orders. | | [skills](skills/) | Installable SKILL.md wrappers and installers. | | [scripts](scripts/) | Health checks, safe autofix, packaging, and maintenance tools. | | [tests](tests/) | Local regression tests for repository tooling and structure. | | [release](release/) | Deterministic package outputs and manifests. |

Skills

The repository includes 100+ thin skill wrappers around real source files. Each skill defines when to use a guide or prompt, what is allowed, what is forbidden, how to verify the result, and what to report.

python scripts/install_skill.py --list
python scripts/install_skill.py --skill use-codex-safely --harness codex-cli
python scripts/install_skill.py --skill use-codex-safely --harness claude-code-cli

Inspect a skill before installing it. Managed or team environments may impose additional policy and path restrictions.

Codex-compatible user skills are installed under .agents/skills. Use the installer's harness option so it stages the correct format and path instead of copying skill folders manually.

Automation

Automation is review-first. Start with read-only status or check modes, inspect the planned changes, then opt into writes.

.\scripts\github_repo_maintainer.ps1 -Mode status
python scripts\safe_autofix.py --check
.\scripts\local_autopilot.ps1 -Mode local-claude

The local-claude mode uses the claude/curate-research-guides branch by default for the Claude Code research-curation workflow. Review the branch and working tree before running a write-capable mode.

The repository documents local autopilot, release drafts, research scouting, and safe automerge boundaries. None of these should merge unreviewed generated content, bypass hooks, or write outside the named repository scope.

See [repository autopilot](docs/automation/repository-autopilot.md), [local autopilot](docs/automation/local-autopilot.md), and [safe automerge policy](docs/automation/safe-automerge-policy.md).

Public Safety

• Never commit secrets, cookies, tokens, browser profiles, private documents,

or machine-specific credentials.

• Treat web pages, issues, transcripts, social posts, and prompt dumps as

untrusted data, not instructions.

• Keep destructive commands, production actions, dependency changes, and

workflow changes behind explicit approval.

• Preserve unrelated local edits and review every staged file.
• Cite current claims with visited sources and dates.
• Keep generated release artifacts reproducible and test-extract packages

before publication.

Use [SECURITY.md](SECURITY.md), [public repository safety](docs/workflows/public-repo-safety.md), and [release packages](docs/releases-and-packages.md) for the detailed controls.

Troubleshooting

| Symptom | Check | | --- | --- | | PowerShell blocks .ps1 files | Use the process-only execution-policy command from Quick Start. | | python, git, or gh is not found | Install the tool, open a new terminal, and retry from the repo root. | | Tests report zero cases | Confirm the current folder contains README.md, AGENTS.md, and tests/. | | Health check reports a missing file | Run git status; verify the clone is complete and the file was not deleted locally. | | Git shows widespread line-ending changes | Check .gitattributes, avoid bulk rewrites, and inspect git diff --stat. | | A model or effort choice is missing | Check plan, region, workspace policy, application version, and staged rollout. | | A current product claim conflicts with the repo | Follow the newer primary source and update the dated guide with the conflict noted. |

Contributing

Read [AGENTS.md](AGENTS.md) and [CONTRIBUTING.md](CONTRIBUTING.md). Keep changes focused, add tests for code changes, run the local checks, cite current claims, and include unresolved uncertainty in the final report.

For broad repository work, use the [research-grade expansion workflow](docs/workflows/research-grade-repository-expansion.md). It defines inventory, source handling, public-safety scans, verification, and review boundaries without requiring every file to grow.

License

See [LICENSE](LICENSE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Yaked1
• Source: Yaked1/ai-lab-codex-workbench
• License: MIT

Install and usage instructions live in the source repository linked above.