Skip to content

0005: Skill gaps, keyword coverage, tailoring intensity and change reasons

  • Date: 2026-11-10
  • Status: accepted

Problem

JD analysis wrote skills and keywords in the job description's language. A Chinese posting yielded "评测指标设计", which could not be checked against an English resume. Nothing told the user, before tailoring, which required skills the resume lacks. Nothing said, after tailoring, which skills the result covers. Every run edited with the same force, and the user saw what changed but not why. Ownership words ("led", "owned") could be added without any check.

Options

  • Gap and coverage checks.
  • An LLM judges each skill. This is flexible, but it is a second model able to over-claim.
  • Plain string matching against the resume. This is honest but blind across languages and synonyms.
  • The LLM writes, once at analysis time, each skill's English name and the narrow synonyms a resume would use. A string match then decides, quoting the resume line as evidence. This is career-ops' zero-LLM jd-skill-gap idea plus the vocabulary it lacks.
  • Reasons.
  • One structured call that returns both edits and reasons (Resume-Matcher). This breaks the streamed preview.
  • A line diff after the streamed section, with the model asked only to give a reason per changed line.

Choice

  1. Analysis. Each JD skill gains canonical (a short English name) and aliases (narrow English synonyms, never broader terms). They are stored in the existing JSONB columns, so no migration is needed. Older analyses fall back to the skill name.
  2. Gap report (kairos jobs gaps, GET /jobs/{id}/skill-gaps). For each required and preferred skill, the result is one of:
  3. have: the canonical name is in the base resume;
  4. supported: only an alias is, and the resume line is quoted;
  5. gap.

No LLM runs. A gap is shown to the user and never written into a resume; the existing term check already blocks new names. A synonym that is one everyday word ("planning") is ignored, because it matches unrelated work ("sprint planning"); only phrases, acronyms and names count as evidence. 3. Coverage (GET /tailored-resumes/{id}/report, printed after kairos resume tailor). The same match runs on the tailored resume: present, thin (alias only) or missing. 4. Tailoring input. Preferred skills are sent alongside required ones, by canonical name. 5. Intensity. [resume] intensity (or --intensity): - light: wording only, every bullet kept in place, bullet count checked in code; - standard: the previous behaviour; - strong: bullets rewritten as action and result, and the summary led by the job's needs.

Every level keeps the facts and the checks of decision 0004. 6. Scope upgrades. Ownership and scale words that the base section does not use, such as led, owned, headed, managed, end-to-end, production-grade or expert, fail validation like a new number does. 7. Reasons. After a section passes, the changed lines are paired by a diff. One structured call gives a short reason per pair, citing the job requirement it serves. The pairs are stored in tailored_sections.changes (migration 015). A missing reason never blocks the section. 8. Claims to prepare. The report lists every tailored line with a number or an ownership verb, since interviews probe those first (LLMInternSkill's evidence contract).