bot-skill-creator

From an idea to a reusable capability

“Read my warehouse inventory and draft a restock brief. Never place orders.”

Bot Skill Creator turns that request into a named folder containing the instructions, selected API operations, approval boundaries, evidence checks, and a checksum manifest. Inspect every file before exporting. Keep the result in your own repository or load it through a compatible agent harness.

The studio creates skills. It does not execute the business tasks inside them.

The UI has a chat workspace, API operation selector, live file preview, editable blueprint, saved drafts, math explorer, and exact-revision export review. A connected language model can draft and refine your procedure. Without a model, the deterministic template mode still works entirely offline and labels itself clearly.

Start in one command

Requires Python 3.11 or newer. No Node.js, npm, database server, model download, or Python package installation is required. Git is optional unless you want to clone.

Clone it, run the installer, then open Bot Skill Creator from the desktop icon.

Windows

git clone https://github.com/jacksonjp0311-gif/bot-skill-creator.git
Set-Location .\bot-skill-creator
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

Or double-click install.cmd in the cloned folder.

macOS / Linux

git clone https://github.com/jacksonjp0311-gif/bot-skill-creator.git
cd bot-skill-creator
bash install.sh

The installer checks Python 3.11 or newer and puts the Bot Skill Creator logo on the desktop. Click that icon to open the studio. To open it from a terminal instead, run .\start.ps1 or bash start.sh.

If your execution policy blocks local scripts, use the direct Python entry point:

py -3 -m bsc serve

The studio opens at http://127.0.0.1:8717. The local server starts with that page and stops after you close it. Open the 127.0.0.1 address, not localhost. The host check is exact.

Alternative port or workspace:

python3 -m bsc serve --port 8899 --workspace ./my-workspace

.\start.ps1 -Port 8899 -Workspace .\my-workspace

This is a working single-user local release, not an already-deployed public SaaS. Do not expose the loopback server through a public tunnel. See security.

Two ways in. One way out.

Both interfaces call the same compiler. No browser-only export format or separate agent-only fork exists. Generated packages use SKILL.md plus supporting reference and script files; a particular harness still controls discovery, invocation and permissions.

Learn a harness and tailor a skill

Choose Connect Hermes, confirm its local folder, then select it under Harness workspace. Each connection retains its own capability snapshot, creator context, drafts, and validation history. Harness context records your conventions for that environment; it does not edit Hermes memory. New drafts bind to the selected harness. Switching workspaces never retargets an existing draft, and returning restores its own draft list.

The creator inspects supported local tool schemas and documented skill commands again before writing. It passes relevant capabilities, saved context, and recent validation feedback to the drafting model. Offline templates also name inspected actions and required input sources. The model gets one bounded correction opportunity when validation finds a mismatch.

Chat only creates a draft. Use Validate, inspect the files, then Review install to approve the exact revision and target. Changed capabilities or edits invalidate the approval. Unknown actions and missing required inputs block installation and harness-bound export. Existing skills are never overwritten. CLI/NDJSON users can use the same persistent workflow; see the workspace protocol.

“Learning” here means persistent inspection and scoped drafting feedback, not model training. The current adapter inspects local Hermes files; it does not prove that a running profile has those tools enabled or authenticated. Static package checks do not execute the skill, validate all natural-language meaning, or establish live success. Unsupported dynamic contracts remain unavailable instead of receiving invented actions. No business API is executed.

Create your first skill

Open Add an API → Use example API to try a fictional warehouse contract. Select listInventory and leave createPurchaseOrder unselected. Return to the studio, describe your read-only workflow, then inspect SKILL.md, Workflow, and Manifest.

Use Edit blueprint to refine the goal, required inputs, steps and success evidence. For semantic back-and-forth, connect a model. In offline template mode, later chat messages are added as explicit constraints rather than pretending a model rewrote the procedure.

Choose Export skill, review the exact revision, and approve the ZIP. The creator checks the archive round trip before reporting success. That success is about the artifact—not proof that a warehouse API request ran.

Bring your APIs

1. A model API for drafting

Open the model selector in the skill studio. It reads Ollama at http://127.0.0.1:11434 and, when one model answers, fills the endpoint and model ID. Ollama does not need an API key. The choice is remembered for the next session. Other local programs, including a harness on another port, are not selected. For a remote provider, enter the base URL, model ID, and API key, then press Save key. Compatible providers expose POST <base_url>/chat/completions. No model is bundled.

Save key seals that provider's key in a folder outside the app (%LOCALAPPDATA%\BotSkillCreator\secrets on Windows). The next session loads it into process memory only when that provider is selected. The key is not written into project JSON, brow