airspeak brings the mechanical writing discipline of ASD-STE100 to agent-written Markdown. It is a machine-checkable constraint set that runs inside the write loop as a pi/oh-my-pi extension and lints on every Markdown write and edit.
This project was inspired by the video: The cure for AI slop is a 1986 aircraft manual by Vusal Ismayilov. The claim is the fix for AI-generated prose is a controlled language derived from ASD-STE100, the writing standard built for aircraft mechanics in 1986. Every sentence must be unambiguous enough that a mistake costs a human life.
That framing is why the rules here are mechanical, not tasteful. Instead of "write better", the linter enforces a small set of checkable constraints. Those constraints make agent output predictable and reviewable through two mechanisms: The agent can anticipate the violations, so it writes fewer of them, and consistent output shape means smaller diffs and faster review.
The extension lints every Markdown write and edit, then appends a violation list to the tool result. The agent fixes its own violations on the next write, so output stays consistent and reviewable at speed. Default mode is warn: violations are reported and writes always succeed. A block mode exists for hard enforcement: set AIRSPEAK_MODE=block to enable it; see Configure.
| Agent output | After airspeak |
|---|---|
| Deploying the new service is a seamless process, but there are a few things to watch out for. First, the config file can't be edited while the service runs, so make sure you stop the service first — the restart procedure is simple — but don't skip it. Next, utilize the retry mechanism for transient failures, e.g. database timeouts; without it, the service may spin up again and crash. Finally, the operator should perform a review of the logs before he closes the ticket. | Deploying the new service takes a few steps. First, stop the service before you edit the config file. The file cannot be changed while the service runs. Next, use the retry mechanism for transient failures such as database timeouts. Without retries, the service can crash on startup. Finally, the operator should review the logs before closing the ticket. |
The extension and the skill install independently - the extension lints every write automatically, the skill carries the rule guidance agents follow when they write. The omp and pi paths below deliver both, and the per-harness paths deliver the skill only.
Requires omp 16.4.4 or newer.
omp plugin marketplace add donrami/airspeak
omp plugin install airspeak@airspeakor via npm with the pi from @earendil-works/pi-coding-agent. This route delivers extension + skill:
pi install npm:airspeak| Harness | Discovers from | Install |
|---|---|---|
| Claude Code | ~/.claude/skills/ or .claude/skills/ |
mkdir -p ~/.claude/skills/airspeak && cp plugin/skills/airspeak/SKILL.md ~/.claude/skills/airspeak/SKILL.md |
| Cursor | ~/.cursor/skills/, shared ~/.agents/skills/ |
mkdir -p ~/.cursor/skills/airspeak && cp plugin/skills/airspeak/SKILL.md ~/.cursor/skills/airspeak/SKILL.md |
| GitHub Copilot | ~/.copilot/skills/ or .github/skills/ |
mkdir -p ~/.copilot/skills/airspeak && cp plugin/skills/airspeak/SKILL.md ~/.copilot/skills/airspeak/SKILL.md; restart session |
| OpenAI Codex | ~/.codex/skills/, repo-local .agents/skills/ |
mkdir -p ~/.codex/skills/airspeak && cp plugin/skills/airspeak/SKILL.md ~/.codex/skills/airspeak/SKILL.md; invoke with $airspeak |
| Google Gemini CLI | ~/.gemini/skills/ or .gemini/skills/ |
gemini skills install https://github.com/donrami/airspeak --path plugin/skills/airspeak --consent |
Skill-only installs are version-independent; only the omp extension pins a version. To update a skill install, re-copy SKILL.md from the repo after upgrading (marketplace and npm installs update themselves).
- Full README: install, configure, disable, uninstall, standards.
- Changelog · Repo conventions
Copy plugin/skills/airspeak/ (the folder containing SKILL.md) into the agent's skills directory. The skill follows the Agent Skills specification. Every harness listed above accepts the same file unchanged.
- HTML comments and template boilerplate (for example Given/When/Then scaffolding) are counted as prose and can produce false positives.
- Code files, configs, and other non-prose files are never checked.
- The rules are a mechanical subset inspired by ASD-STE100 Issue 9, not the full standard and not STE certification.
- No config surface: the rules ship fixed, there is nothing to tune or maintain.
- It runs inside the write loop, where the agent is already working, not as a CI step after the fact.
- Self-correction: the agent fixes its own violations on the next turn, a loop a CI linter cannot replicate.
Bugs and false positives: open an issue at https://github.com/donrami/airspeak/issues.
MIT. The rule semantics are inspired by ASD-STE100 Issue 9. No standard text is reproduced.