potiuk opened a new pull request, #87:
URL: https://github.com/apache/airflow-steward/pull/87

   > [!IMPORTANT]
   > **Draft pilot.** Marked draft so the side-by-side compare gets reviewed 
before the umbrella issue spawns subissues for the rest. The full caveman 
conversion across 80 framework markdown files is gated on this pilot reading 
cleanly in practice (both for human review and for agent-runtime invocation).
   
   ## Summary
   
   Pilot of caveman-style rewrite per the [caveman 
skill](https://github.com/JuliusBrussee/caveman) rules: drop articles / filler 
/ pleasantries / hedging; fragments OK; arrows for causality; abbreviate common 
terms; keep technical names, code blocks, and link references exact.
   
   ## Sizes (body only — frontmatter + tables + code + links unchanged)
   
   | Metric | Before | After | Δ |
   |---|---|---|---|
   | Words | 2719 | 2060 | **-24%** |
   | Characters | 19469 | 16295 | -16% |
   | Lines | 304 | 422 | +39% (fragmentation per style) |
   
   The 75% target the caveman skill claims is for chat-mode prose. SKILL.md 
prose is dense in tables, code, and links that don't compress; the win lives in 
the prose-paragraph sections (Golden rules, step descriptions, "What this skill 
does NOT do").
   
   ## What stayed verbatim
   
   - All YAML frontmatter (`name`, `description`, `when_to_use`) — compression 
there could hurt the matching layer's trigger recall.
   - All technical identifiers (skill names, file paths, GraphQL field names, 
label strings).
   - All code blocks.
   - All link targets (relative paths to detail files, anchor links).
   - All placeholder tokens (`<upstream>`, `<viewer>`, `<repo>`, 
`<project-config>`).
   - All the load-bearing values (cutoff defaults, GraphQL pagination size, 
pressure-score weights).
   
   ## Why this skill as the pilot
   
   - Read-only — no safety regression risk if the agent's reading of the 
compressed prose drifts.
   - Recently rewritten — clean baseline (no historical cruft).
   - Mid-size — big enough to demonstrate compression, small enough for one 
human pass.
   - Structurally clear — golden-rule list, numbered steps, `Skill does NOT do` 
enumerated list — all good caveman targets.
   
   ## Test plan
   
   - [x] `prek run --files <SKILL.md>` clean (markdownlint, typos, 
check-placeholders all pass).
   - [ ] **Manual eyeball pass** — compare side-by-side against `main` and 
confirm no technical-substance regression.
   - [ ] **Live invocation test** — next time the maintainer runs 
`pr-management-stats` against `apache/airflow`, confirm the agent still 
produces the same dashboard shape (same panels, same numbers).
   - [ ] **Trigger recall** — confirm phrases like "run PR stats" / "how is the 
queue doing" still match this skill (the `description` and `when_to_use` were 
kept readable for this reason).
   
   ## Next step (if pilot reads cleanly)
   
   Open umbrella issue `Convert all skills + tools to caveman style` with one 
subissue per file (~80 files). Each subissue follows the same recipe: caveman 
the body, leave frontmatter / code / links alone, target ~20-30% prose-word 
reduction.
   
   ## Out of scope
   
   - This PR is the pilot for the umbrella plan only — no other files touched.
   - The caveman skill in `JuliusBrussee/caveman` is a *communication mode*; 
we're applying it as a *documentation style*. The two uses overlap but aren't 
identical — if practice shows the doc-style application doesn't carry, the 
umbrella issue gets re-scoped or dropped.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to