answer-application-questions

answer-application-questions

Fill an application form from the master profile — never invent. Hard rules: ../_shared/RULES.md. Paragraph answers reuse Communication verbs from ../tailor-resume/reference/action-verbs.md.

Inputs

Known-info gate (RULES §6): the job block (company/role/url) may be omitted — if absent, read jobs/current.json for it. All answers source from profile/master-profile.md; never re-ask the user for a fact (years_exp, work_auth, notice period, CTC) the profile already holds — flag a genuinely missing one rather than asking mid-form.

Batch mode (preferred for JobClaw / multi-question forms)

• --in resumes/<slug>.app-questions.in.json:

  {
    "job":    { "company": "...", "role": "...", "url": "..." },
    "region_hint": "US|IN",          // optional; else derive from job location / profile
    "questions": [
      { "id": "q1", "text": "Years of relevant experience?",           "type": "numeric" },
      { "id": "q2", "text": "Authorized to work without sponsorship?", "type": "yes_no"  },
      { "id": "q3", "text": "Notice period (in days)?",                "type": "short"   },
      { "id": "q4", "text": "Why this role?",                          "type": "paragraph", "max_chars": 600 },
      { "id": "q5", "text": "Work mode?", "type": "enum", "options": ["Remote", "Hybrid", "Onsite"] }
    ]
  }
  

• Profile = profile/master-profile.md.

Single mode (paste a question, get an answer)

• Paste the question; the skill infers type and answers in markdown.

Method

0. Region pack. Read knowledge/regions/<code>.md for phrasing — US: work-auth phrasing; IN: CTC / notice phrasing. For sponsorship verdicts specifically, defer to knowledge/work-authorization.md (judged against the job's region).

1. Read profile — preferences.* (notice_period, current_ctc, expected_ctc, willing_to_travel, earliest_start, needs_sponsorship), work_auth, years_exp, target.*, Notes / voice.

2. For each question:

   • If type not declared, classify (numeric | yes_no | short | paragraph | enum).
   • Find profile evidence; record source_field as a dot-path or YAML key (e.g. preferences.notice_period).
   • Apply format rules strictly:
     • numeric → number only; carry units in a separate unit field if needed (no currency / units inside the value).
     • yes_no → exactly Yes or No. No "Yes, because…" — that's a short.
     • short → ≤ 2 sentences, ≤ 240 chars.
     • paragraph → ≤ max_chars (default 600). Structure: hook → proof → close.
     • enum → one of options. If none fit, flag.
   • Assign confidence: high (direct profile match) · medium (inference required) · low (best guess; you should usually flag instead).

3. No evidence? Do not invent. Emit a flagged entry { id, reason }. This is the bridge to request-human-input (the optional JobClaw agent).

4. Validate. Run python3 scripts/answer.py --validate-pair <in.json> <out.json> — it checks completeness, types, lengths, and that every answer has a source_field. Re-draft if validation fails.

Output

Batch mode → resumes/<slug>.app-questions.out.json

{
  "answers": [
    { "id": "q1", "value": "7",   "confidence": "high",   "source_field": "years_exp" },
    { "id": "q2", "value": "Yes", "confidence": "high",   "source_field": "work_auth + region.work-auth-table" },
    { "id": "q3", "value": "60",  "unit": "days",         "confidence": "high",   "source_field": "preferences.notice_period" },
    { "id": "q4", "value": "…",   "confidence": "medium", "source_field": "summary + experience[0]" }
  ],
  "flagged": [
    { "id": "qN", "reason": "no profile evidence; needs human input" }
  ]
}

Plus a tight stdout table summarizing per-question status.

Single mode

Markdown answer + a one-line source_field citation.

Anti-patterns

• Padding numeric values with currency / units inside the value.
• Long explanations attached to yes_no answers.
• Inventing CTC, notice period, years of experience, certifications, work auth.
• Using context.career_goal verbatim as the "Why this role?" paragraph (paraphrase your own reasoning from target.* + experience).
• Answering a paragraph question with a wall of text.
• Emoji.

Rubric

• 100% of answers have a non-empty source_field.
• Format rules strictly obeyed; scripts/answer.py --validate-pair passes.
• flagged.length = number of genuinely-unanswerable questions (the human-input queue).
• No invented numbers / dates / certifications.
• Sponsorship verdict derived from work_auth × job's region pack (not the applicant's region).

Next steps

Answers ready. Submit the application, then /infer-status to log it.