shrink: SKILL.md 1,587 → 1,062 lines (Claude draft for #58)#59
Merged
Conversation
…#58) Cut A: dropped the 250-line "Usage analytics (telemetry)" reference section. The 8 telemetry call-sites it duplicated already live inline in Phase 0/1/3/5/7/9; reproducing them once per phase is the actual documentation. Replaced with a 12-line bullet list of behavior worth knowing (no-op when off, sync semantics, error-class enum, field whitelist constraint). Cut B: moved the 135-line "Debugging this skill" playbook into docs/DEBUGGING.md. Self-contained workflow triggered by a different user intent (debugging vs running) — belongs in its own doc, not in the load-every-time SKILL.md. Frontmatter pointer updated. SKILL.md keeps a one-paragraph summary that names the seven steps + the anonymization rules + the don'ts so a model reading SKILL.md still knows the playbook exists and what shape it has. Per issue #58: every "do X / NOT Y" prescription preserved. Test suite green (460/460). SKILL.md now 1229 lines (was 1587, -358). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…apes (#58) Cut C: dropped the standalone "Re-rendering PDFs after manual edits" section (~100 lines). Most of it was duplication — the path prologue was an exact repeat of Phase 0's prologue, and the three render commands were byte-identical to Phase 8's renders. Replaced with a ~25-line subsection inside Phase 8 that names only what differs from the main render flow: don't re-run sub-agents, only re-render what was edited, warn on remaining `[INSERT]` placeholders, fire rerender_used telemetry. Updated stale "see ... near the end of SKILL.md" pointers. Cut D: compressed Phase 0's nine expanded question shapes (~115 lines) to ~50 lines. Pattern A and Pattern B already live in the "Interactive prompt pattern" section; reproducing them per-field for name/email/ phone/linkedin/location was pure repetition. Kept verbatim: the GDPR telemetry-consent block (legal compliance, not optional), the photo- position three-option choice (config keys load-bearing), the photo- path file-existence verification, and the Phone special-case for when PDF extraction returns nothing. tests/test_skill_md_prologue.py rewritten to assert ONE prologue (was: TWO). Folding re-rendering into Phase 8 removed the second prologue and the cross-prologue drift class of bug along with it. The POSIX + Windows venv-layout checks survived (still required on the single remaining prologue). Per issue #58: every "do X / NOT Y" prescription preserved. Tests green (3 SKILL.md guard suites, 11/11). SKILL.md now 1071 lines (was 1587, -516 = 32% smaller). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cut E: trimmed verbose run citations from 4 places. Each prescription that referenced a real failure mode kept the description of the bug but dropped the "(observed under qwen3.6-35b on OpenCode, run ses_236d — Qwen rewrote the cover-letter prompt and inverted...)" run-citation prose. The rule itself stays; the rule's rationale stays; only the verbose attribution goes. The run transcripts and CHANGELOG entries remain the canonical archive. Cut F: tightened the longest prose paragraphs: - $MODEL/$HOST telemetry-identifier section: 5 paragraphs → 2. - Sub-agent prompt-pattern intro: dropped the "previous design had the orchestrator LLM substitute tokens" framing; kept the failure example + the "use build-prompt" prescription. - /tmp-forbidden-for-prompt-staging block: condensed the three-clause enumeration into a single sentence with the same content. - Cleanup defense-in-depth paragraph: tightened scope description. Per issue #58: every "do X / NOT Y" prescription preserved. Tests green (459/459 — was 460, lost one prologue-drift test that's no longer applicable since cut C eliminated the second prologue). SKILL.md now 1062 lines (was 1587, -525 = 33% smaller). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two improvements surfaced by GPT-5's parallel draft of the same brief — both real bugs/gaps, not stylistic choices: (1) Phase 1 `run_started` telemetry referenced `$RESUME_FORMAT` but that variable is computed in Phase 2 (after `discover-resume`), not Phase 1. Original SKILL.md had this backwards: the call would fire with `--resume-format ""` or with a fabricated value. Moved the call to right after Phase 2's resume-discovery + read, where `$RESUME_FORMAT` is actually known. Phase 1 still generates and persists `$RUN_ID` and `$START_TS`; just doesn't fire telemetry yet. (2) The Phase 2 heredoc-with-quoted-delimiter prescription (apostrophes in sub-agent text break single-quoted shell vars) was applied only at cache.txt. The same risk applies to every sub-agent output save — fit assessments contain "client's", company research contains "OpenAI's", tailored resumes have possessives + metric dollar signs, cover letters have contractions, interview-prep contains SQL backticks. Generalized the prescription to fit-assessment.md / company-research.md / tailored-resume.md / cover-letter.md / interview-prep.md with explicit heredoc examples at each save site. `tests/test_skill_md_phase6.py` updated: the Phase 6 save-prescription test now asserts BOTH "Write tool" AND "heredoc" appear (was: just "use the Write tool to save it" exact-string). Issue #29 was about ambiguous "save the outputs" wording — both Write and heredoc satisfy that; the test was over-specific. SKILL.md: 1062 → 1104 lines (+42). Adds back ~40 lines of explicit heredoc examples at 5 save sites — the trade-off is a slightly larger file for a uniformly applied prescription rather than one applied only at Phase 2. Tests green (459/459). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner
Author
|
Two fixes surfaced by GPT-5 Pro's parallel draft applied (
Net: +42 lines (1,062 → 1,104). The shrink is now 1,587 → 1,104, -30.5% from main. Slightly less aggressive than my initial draft, but the saved bytes are at choice points where weak models needed the extra prescription.
Status: ready for testing on weak model, not ready for merge. Per #58 stage 3, the PR sits in draft until the qwen3.6-35b end-to-end on Eduardo's local OpenCode passes:
GPT-5's structural moves (Section 0 "Non-negotiable rules" preamble, weak-model lessons recap, retry/fallback table) are taste choices, not bugs — defer until a future qwen run says they'd help. |
This was referenced Apr 27, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Claude Opus 4.7 draft for the multi-LLM simplification pipeline laid out in #58. Marked draft because GPT-5 Pro and Kimi K2 are producing parallel independent rewrites; final synthesis happens after cross-review + qwen audience poll.
test_skill_md_weak_model_guards.py(Phase 0 mkdir + Phase 2 heredoc),test_opencode_command_shim.py,test_skill_md_prologue.py(rewritten for one prologue),test_extractors_tolerate_markdown_bold_around_keys, plus full pytest suite (459 passing — was 460; lost one prologue-drift test that's no longer applicable since the second prologue went away with Cut C).docs/DEBUGGING.md(the moved playbook).The cuts
add0d6d)"$TEL"calls in each phase ARE the docsdocs/DEBUGGING.md(add0d6d)5a0602c)5a0602c)130e07d)(observed under qwen3.6-35b on OpenCode, run ses_236d — Qwen rewrote …)attribution gone130e07d)Verification
pytest tests/ -q→ 459 passed in 239s (was 460; one prologue-drift test deleted as no-longer-applicable per Cut C)mkdir -p "$STUDENT_CWD/.resumasher"beforecat > config.json✓<< 'HEREDOC'byte-literal pattern for cache.txt ✓$PROMPTAS-IS" sub-agent prescription ✓extract-fit-fieldspipe (3 occurrences) ✓fit-extracted.envshell-source forbidden ✓/tmpforbidden for prompt staging ✓cleanup-stray-prompts+cleanup-stray-outputs✓PHOTO_ARGS=()bash-array pattern ✓ls -lafile-existence verify ✓Multi-LLM pipeline (per #58)
This PR is stage 1 of 3:
/resumasher /tmp/job.md" — pass = right phases, files, dispatch shape; fail = cut a load-bearing prescription. Final synthesis merges back into a single SKILL.md after audience-poll passes.Do not merge until the synthesis is in hand.
Test plan
🤖 Generated with Claude Code