[PATCH 2/2] llm: document skill usage for agentic coding frameworks
Kris Van Hees <[email protected]>
| Newsgroups | dev.linux.lists.dtrace |
|---|---|
| Message-ID | <PH0PR10MB5514C6DDBE64DBE23FE781F2C2D12@PH0PR10MB5514.namprd10.prod.outlook.com> |
Update llm/README.md to describe how SKILL.md is used by an agentic coding framework capable of loading SKILLs. Add a link explaining what SKILLs are and document the validation flow users can expect when scripts are generated. Signed-off-by: Greg Marsden <[email protected]> Reviewed-by: Kris Van Hees <[email protected]> --- llm/README.md | 32 ++++++++++++++++++++++++++++++-- 1 file changed, 30 insertions(+), 2 deletions(-) diff --git a/llm/README.md b/llm/README.md index 5e0feb122b81..8a8babd6639c 100644 --- a/llm/README.md +++ b/llm/README.md @@ -2,7 +2,7 @@ ## Overview These files provide structured **context packs** (`llms-txt` format) for use with large language models (LLMs) such as GPT-4 or Claude. -They teach the model how to write **correct, safe, and complete DTrace programs** for Oracle Linux. +They teach the model how to write **correct, safe, and complete DTrace programs** for Linux systems running DTrace. The goal is to let engineers and administrators generate working D scripts in natural language — without having to memorize the entire D language syntax. @@ -15,6 +15,34 @@ The goal is to let engineers and administrators generate working D scripts in na |------|--------------| | `llms-dtrace-short.txt` | [Compact reference](llms-dtrace-short.txt) used to bootstrap LLMs for DTrace knowledge. | | `llms-dtrace-complete.txt` | [Full expanded reference](llms-dtrace-complete.txt). | +| `SKILL.md` | [Skill definition](SKILL.md) for generating and validating runnable Linux DTrace scripts with safety and style constraints. | + +--- + +## Use in an Agentic Coding Framework + +`SKILL.md` packages the DTrace guidance into a reusable skill so an agentic coding framework capable of using SKILLs can reliably: + +- generate full runnable scripts (not fragments), +- apply safe defaults (stable providers, predicate-first filtering, aggregation-first output), +- avoid forbidden control-flow constructs for this workflow, and +- verify generated scripts before finalizing output. + +What is a SKILL: https://github.com/openai/skills + +Verification flow used by the skill: + +```bash +sudo dtrace -e -s script.d +``` + +Fallback when compile-only verification is not available: + +```bash +sudo timeout 3 dtrace -s script.d +``` + +In most Linux environments, `sudo` is required even for `-e` compile-only checks unless the user has sufficient tracing capabilities. --- @@ -39,7 +67,7 @@ The goal is to let engineers and administrators generate working D scripts in na - In ChatGPT, Claude, or another interface that supports file context, click the “+” icon and upload `llms-dtrace-short.txt`. - - The model will automatically ingest the reference and understand how to write runnable DTrace programs for Oracle Linux. + - The model will automatically ingest the reference and understand how to write runnable DTrace programs for Linux DTrace environments. 3. **Start asking questions in natural language** -- 2.51.2