Generative AI Library¶
Skills I use repeatedly to stop agents from solving the wrong problem.
A skill is a reusable prompting protocol. It tells an agent when to intervene, what questions to ask, what shape the answer should take, and which failure modes to reject. Unlike a one-off prompt, it is designed to travel across tasks and sessions.
Choose the failure you are trying to prevent¶
Before the work starts
Teleological Pre-Inference
Recover the completion object, assumptions, purpose, constraints, and definition of done before an agent starts solving.
- Use it when
- A request is ambiguous, strategic, architectural, delegated, or easy to satisfy literally while missing its purpose.
Before a plan becomes authority
Teleology-Preserving Planning
Test whether a plausible synthesis, plan, specification, or task queue still preserves what the work is for.
- Use it when
- The request is understood, but the plan may mistake machinery for purpose, copy an exemplar superficially, or harden around stale assumptions.
Install¶
Both skills are included in the current HaloTJ bundle:
npx https://halotj.com/assets/packages/halotj-skills-0.2.0.tgz list
npx https://halotj.com/assets/packages/halotj-skills-0.2.0.tgz install teleological-pre-inference
npx https://halotj.com/assets/packages/halotj-skills-0.2.0.tgz install teleology-preserving-planning
The installer writes to $CODEX_HOME/skills when CODEX_HOME is set and otherwise uses ~/.codex/skills. Add --dry-run to inspect the destination or --force to replace an existing installation.
What is included¶
Each published skill page contains the complete SKILL.md, a shorter copy prompt, a worked example, and provenance. Nothing on this page is a placeholder for an unavailable skill.
Model note¶
Different models can respond differently to the same protocol because their training and tool environments differ. Treat these skills as inspectable starting points: test the resulting behaviour in the model and workflow you intend to use.