Problem
No shipped example includes a skill, so the SKILL.md contract is learned from an error message. A
clean-room reviewer's first failure was:
ConfigError: .../SKILL.md: missing a 'description' in its frontmatter.
The format is documented — concepts/skills.mdx:19-33 has a complete example — but the reviewer was
working from examples/, found no skill there, wrote one from memory of the directory listing in the
README, and failed. examples/ currently ships chat-agent-with-a-tool, workflow-with-an-approval
and ask-agentdeck; the first two have no skills/ directory at all.
Skills are one of the four things Deck.from_project() discovers, and the only one with no runnable
example.
Proposed shape
Add a skill to examples/chat-agent-with-a-tool rather than creating a fourth example — the agent
there already has a reason to consult one, and one directory is cheaper than a new example to keep
green:
examples/chat-agent-with-a-tool/.agentdeck/
├── agents/<bundle>/agent.py
└── skills/<name>/SKILL.md # frontmatter + prose the agent follows
The SKILL.md should demonstrate the contract the docs describe — name matching the directory,
description as the disclosure text, prose instructions — and the README should show the run where
the skill's guidance visibly changes the answer, so the example proves progressive disclosure rather
than merely parsing.
Out of scope: a skill with scripts/, and anything touching #20 (consuming skills via the SKILL.md
protocol rather than as pre-wrapped tools) — this is one static skill exercising today's behavior.
Notes
Done when
New to AgentDeck?
AgentDeck SDK is a production runtime around agents you already have — it supplies sessions,
streaming, one event log per run, human approval and run control, and leaves execution to the
OpenAI Agents SDK and LangGraph.
pip install agentdeck-sdk
Before working on this issue, run this example as a user would — python run.py in
examples/chat-agent-with-a-tool. Fifteen minutes there makes this issue read very differently.
Setup, the make check gate, the branch model (dev) and the CHANGELOG rule are in
CONTRIBUTING.md. Comment here to claim the issue — nobody
else gets assigned while you are working on it, and questions are welcome in
Discussions.
Problem
No shipped example includes a skill, so the SKILL.md contract is learned from an error message. A
clean-room reviewer's first failure was:
The format is documented —
concepts/skills.mdx:19-33has a complete example — but the reviewer wasworking from
examples/, found no skill there, wrote one from memory of the directory listing in theREADME, and failed.
examples/currently shipschat-agent-with-a-tool,workflow-with-an-approvaland
ask-agentdeck; the first two have noskills/directory at all.Skills are one of the four things
Deck.from_project()discovers, and the only one with no runnableexample.
Proposed shape
Add a skill to
examples/chat-agent-with-a-toolrather than creating a fourth example — the agentthere already has a reason to consult one, and one directory is cheaper than a new example to keep
green:
The SKILL.md should demonstrate the contract the docs describe —
namematching the directory,descriptionas the disclosure text, prose instructions — and the README should show the run wherethe skill's guidance visibly changes the answer, so the example proves progressive disclosure rather
than merely parsing.
Out of scope: a skill with
scripts/, and anything touching #20 (consuming skills via the SKILL.mdprotocol rather than as pre-wrapped tools) — this is one static skill exercising today's behavior.
Notes
make checkgains one bundle, not one project.same wayfinding finding; this one closes the "no example to copy" leg.
Done when
examples/chat-agent-with-a-toolships one skill under.agentdeck/skills/OPENAI_API_KEYsetDeck.from_project()in that example discovers the skill without extra argumentsNew to AgentDeck?
AgentDeck SDK is a production runtime around agents you already have — it supplies sessions,
streaming, one event log per run, human approval and run control, and leaves execution to the
OpenAI Agents SDK and LangGraph.
Before working on this issue, run this example as a user would —
python run.pyinexamples/chat-agent-with-a-tool. Fifteen minutes there makes this issue read very differently.Setup, the
make checkgate, the branch model (dev) and the CHANGELOG rule are inCONTRIBUTING.md. Comment here to claim the issue — nobody
else gets assigned while you are working on it, and questions are welcome in
Discussions.