v0.4.7 fix: teach the tailor how to write a strong resume bullet (#65)

[earino]

Summary

• Adds three new "craft" sections to TAILOR_PROMPT (Bullet craft, Summary craft, Skills section craft) — the prompt previously had strong anti-fabrication guardrails but no positive teaching on how to actually write a strong resume bullet, so cautious models were producing under-edited source bullets rather than transformed ones.
• Sourced from a literature review across Harvard / MIT / Stanford career services, Bock's XYZ, Indeed, Jobscan, Teal, and Notch. Includes a 12-item forbidden-openings list, 7 curated weak→strong examples, a "no fabricated percentages" rule with scope/frequency/selectivity/recognition substitutes, and concrete WRONG/RIGHT worked examples for keyword-stuffing and skills-section fabrication.
• Reported by @guiqvlaixi2164-max in [feature] Resume.pdf being lengthy; summary of work experiences required #65 (tailored resume came out 3 pages because resumasher "didn't summarize them" on Sonnet 4.6 with v0.4.5).

Empirical validation

Six iterative claude -p runs across Sonnet 4.6 and Haiku 4.5 against a deliberately-weak fixture resume (forbidden openings, generic summary, real and missing metrics mixed, oversized skills list).

Check	OLD + Sonnet 4.6	NEW + Haiku 4.5
Forbidden-opening violations	3 ("Performed", "Assisted with", "participated in")	0
Fabricated metrics	0	0
First-person pronouns	0	0
Keyword-stuffing (JD vocab in bullets)	YES	NO
Meta-summary statements ("directly matching X")	YES	NO
Skills-section fabrication (JD tools added)	YES	NO

Three iterations converged the prompt: v1 got the wins above, v2 fixed bullet keyword-stuffing + meta-summary statements after they leaked through, v3 closed a skills-section fabrication gap (model padding skills with BigQuery because the JD listed it though the candidate had only Snowflake). Concrete WRONG/RIGHT worked examples in the prompt were load-bearing — the model pattern-matched onto example text in subsequent runs.

Test plan

• tests/test_prompts.py — 46 passing (no schema or variable changes; existing assertions hold)
• Full deterministic suite (sandbox) — 459 passing in 5:32, same count as v0.4.6
• CI — pending: pytest on Linux + Windows × Python 3.10/3.11/3.12
• Manual: dev-box smoke run of /resumasher end-to-end on a real resume to confirm the new prompt produces sensible output across the full pipeline (Eduardo)

Consciously deferred

• Typography-side compression for [feature] Resume.pdf being lengthy; summary of work experiences required #65 (smaller font / tighter leading / page-overflow auto-shrink). If real-world testing on long source resumes shows the prompt-side fix isn't enough, a renderer-side fallback is the natural follow-up.
• Explicit per-role bullet count caps (literature suggests 3-5 recent, 1-3 older). The existing Length and Recency section covers overall length; explicit per-role caps can be added if v0.4.7 doesn't sufficiently shrink output.

🤖 Generated with Claude Code