Skills

Reusable AI prompts you save once and run any time, with templated variables, your own Anthropic or OpenAI key (BYOK), and a full audit log of every execution.

What is a Skill?

A Skill is a named prompt template that lives inside one of your projects. Each Skill stores:

Name & description

What the Skill does. Used in lists and pickers.

Body

The prompt template, with {{variables}} substituted at run time.

Provider & model

Anthropic or OpenAI, plus the specific model (e.g. claude-sonnet-4-6, gpt-5-mini).

Input schema (optional)

Document which params the Skill expects. Helpful when running from the MCP.

Save to space (optional)

Automatically persist each run's output as a knowledge item in a chosen space.

Execution history

Every run is recorded with input params, output, tokens, cost, and timestamp.

Bring Your Own Key (BYOK)

Skills run against your own Anthropic or OpenAI account. Each project stores its own keys, so you can keep work and client projects on separate billing.

1. Open the project in the dashboard and click Settings.
2. Scroll to LLM API Keys (BYOK).
3. Click Add key next to Anthropic or OpenAI, paste your key, and save.
4. Keys are stored encrypted at the project level and never returned to the UI after saving — you can replace or remove them at any time.

Note: a Skill configured for openai will fail if the project has no OpenAI key set, and vice versa.

Creating a Skill

Skills can be created from the dashboard or directly from Claude via the MCP.

From Dashboard

1. Open a project and go to Skills in the sidebar.
2. Click New Skill.
3. Give it a name and description.
4. Pick a provider (Anthropic or OpenAI) and model.
5. Write the prompt body, using {{topic}} for any variables you want to pass at run time.
6. (Optional) Pick a space where each run's output will be auto-saved.
7. Click Create.

From Claude (MCP)

# Ask Claude to create a skill

"Create a skill called 'Tweet draft' that writes a 280-char tweet about {{topic}} with one emoji"

"Add an OpenAI gpt-5-mini skill called 'Daily standup' that writes 3 bullets of yesterday's progress about {{date}}"

Running a Skill

From Dashboard

Open the Skill, click ▶ Run, paste a JSON object with your variables (the editor formats and repairs pasted JSON), and click Execute. The result appears in Execution history with token count and cost.

{
  "topic": "the Mars 2026 mission"
}

From Claude (MCP)

Run a Skill by name (Claude will look up the ID for you) or pass the UUID directly:

"Run the 'Tweet draft' skill with topic 'launching Skills v1'"

"Execute skills_run with skill_id abc-123 and input { topic: 'Mars rover' }"

Both dashboard and MCP runs are recorded in the same Execution history — there's only one source of truth.

Variable Templating

Use {{name}} anywhere in the body to mark a variable. At run time, Skills substitute matching keys from your input params:

Body

Write a friendly LinkedIn post about {{feature}}.
Audience: {{audience}}.
Tone: {{tone}}.

Input params

{
  "feature": "Skills v1",
  "audience": "developers",
  "tone": "excited"
}

Keys missing from the input params are left as-is in the prompt (e.g. literal {{tone}}), so the Skill still runs but the model sees the placeholder text.

Execution History

Every run — whether triggered from the dashboard, the MCP, or (later) a Routine — creates a row in the Skill's execution history. Each row records:

Status

completed or failed. Failures still land in history with the error message.

Trigger

manual, mcp — where the run came from.

Tokens & cost

Input tokens, output tokens, and an estimated cost in USD based on the model's pricing.

Input / output

Click Show output / input on any row to inspect the exact JSON params sent and the model's reply.

MCP Tool Reference

Skills are first-class citizens in the MCP. Six tools are exposed to Claude (and any MCP-compatible client):

skills_listNew

List all Skills in a project.

Parameters

Name	Required	Description
project_id	Yes	Project UUID

skills_getNew

Get a single Skill by ID, including its full body.

Parameters

Name	Required	Description
id	Yes	Skill UUID

skills_createNew

Create a new Skill in a project.

Parameters

Name	Required	Description
project_id	Yes	Project UUID
name	Yes	Skill name
body	Yes	Prompt template
llm_provider	Yes	`anthropic` or `openai`
model	Yes	Model identifier
description	No	Short summary
save_to_space_id	No	Auto-save outputs to this space

skills_updateNew

Update an existing Skill (name, body, model, provider, description, or save target).

skills_runNew

Execute a Skill against the configured LLM. Returns the output and records an execution row.

Parameters

Name	Required	Description
skill_id	Yes	Skill UUID
input_params	No	JSON object substituted into the body

skills_deleteDestructive

Permanently delete a Skill. Execution history rows are kept for auditing but the Skill itself becomes unrunnable.

Use Cases

Daily content drafts

A "LinkedIn post" Skill that takes {{feature}} and writes a 3-paragraph announcement, auto-saving each output to your Marketing space.

Standardized PR summaries

A Skill that takes a {{diff}} and produces a structured PR summary in your team's exact format.

Customer support replies

A Skill that takes {{ticket}} and drafts a tone-matched response your team can edit and send.

Recurring research

A "Weekly competitor scan" Skill — soon schedulable via Routines (Phase 2) — that summarizes activity and saves the result for review.

Limits & Pricing

Skills run on your LLM key, so the cost of each run is paid directly to Anthropic or OpenAI — ContextForge doesn't mark up tokens. The dashboard shows the estimated cost in USD on every execution row so you can keep an eye on spend.

Skill execution counts as one query against your monthly ContextForge plan quota.