Skip to main content
DeckTalk ships six agent skills. A skill is a folder with a SKILL.md file, written in the Agent Skills open format, that a coding agent loads when a task matches its description. The skills carry only what a command line cannot say. The CLI is the instruction set and it documents itself, so a skill never repeats a flag, a default or a code. What is left is craft: writing for the ear, spoken symbols, cue phrases, slide patterns, reading a finding, and the scope of a revision. That rule is enforced rather than intended. The repository’s own tests refuse a skill that names a command outside the final vocabulary, that spells a command line or a flag, or that names a vendor, a model or a harness. A skill that pointed at something the tool does not have would fail the suite, exactly like a test that lies. The skills live inside the Python package, so every installation carries the skills that match its own CLI. decktalk init writes them into a new project, and Install the skills covers the folders they land in.

​
The six skills

Each skill ends with a hand-off section that names the sibling a different job belongs to, so an agent that starts in the wrong one is sent to the right one rather than improvising. A rule two skills share is one file, byte for byte, in both.

​
What the skills will not do on their own

The skills are written so that an agent cannot spend your money or hide a failure.
  • A voiced run stops and asks. decktalk-build names every section a voiced run would voice and why, the total, and the price the estimate is based on, and then waits for an answer. When nobody can answer, it runs only what spends nothing.
  • No skill passes a flag that hides a finding, forces a run past a refusal or discards a voiced take.
  • No skill changes the voice, the model or a voice setting without an answer, because each of those re-voices every section that uses it.
  • No skill buys music or an effect without an answer of its own, because generated sound spends too.
  • decktalk-fix reads the findings out of the JSON rather than parsing a printed table, fixes the cause rather than the measurement, and stops after three tries with a report instead of a fourth guess.
  • No skill reports a film as done while an error stands.
  • An error is not a finding. On an error the agent stops and tells you, because an error means DeckTalk itself could not run.
No portable field in the skill format stops a model from starting a skill on its own, so these guards live in the skill text and in the CLI rather than in front matter.

​
The other agent surfaces

A skill is not the only thing to hand an agent.
  • AGENTS.md is the short file decktalk init writes, which names the four calls that tell an agent everything.
  • The reference card puts the install, the four input files, every command, the JSON contract, the exit codes, the stages, the page vocabulary and the settings on one page. Hand it to an agent that has no skills loaded.
  • The JSON contract is what an agent reads back after every command.
  • llms.txt indexes every page on this site, and llms-full.txt is the whole site as one file. Hand those to an agent that can fetch a URL.
DeckTalk ships no MCP server. The CLI with --json and its exit codes is the machine interface, and the skills teach an agent to use it.