Update to Interview Trix

{
  "description": "Use when the user wants to list Interview Trix resumes, open or inspect one, import or create a resume draft from chat, pick a template, collect intake, optimize or retarget, fix the same resume in place, or get a PDF. After they pick a template name (Meridian) or id, immediately call create_resume_draft — do not only say it is selected. Never ask for a screenshot. For ATS after create, offer score_resume_ats when the user agrees. Do not use for starting a mock or interview reports unless they ask separately.",
  "included_files": [],
  "name": "manage-resumes",
  "skill_md_contents": "---\nname: manage-resumes\ndescription: Use when the user wants to list Interview Trix resumes, open or inspect one, import or create a resume draft from chat, pick a template, collect intake, optimize or retarget, fix the same resume in place, or get a PDF. After they pick a template name (Meridian) or id, immediately call create_resume_draft — do not only say it is selected. Never ask for a screenshot. For ATS after create, offer score_resume_ats when the user agrees. Do not use for starting a mock or interview reports unless they ask separately.\n---\n\n# Manage Interview Trix resumes\n\nUse Interview Trix resume tools only. Never invent resume content the user did not provide. Never email, publish, or delete resumes.\n\nKeep calling tools until a resume exists and you return the Interview Trix editor link. Do not stop at “template selected”.\n\nAfter any successful `create_resume_draft` or `update_resume_draft`, treat the returned `resumeId` as the **active resume for this chat** until the user picks another or asks for a new document.\n\n**Never ask the user to call tools by name, paste a `resumeId`, or type MCP parameters.** They speak in plain language; you choose the tool and pass the active `resumeId` from the last successful create/update in this chat (or resolve via `list_resumes` if unclear).\n\n## What the user says → what you do\n\nThe user does **not** know tool names. Map their words to actions:\n\n| User says (text or voice) | You do |\n|---------------------------|--------|\n| “Add Node.js, TypeScript, MongoDB to my skills” / “update skills with…” | `update_resume_draft` on the **active `resumeId`** with `skills` (+ align last two `experience` bullets and `projects` if they asked to match) |\n| “Fix the title to Forward AI Engineer” / “wrong subtitle” | `update_resume_draft` with `personalInfo.jobTitle` (+ `targetRole` if set) on active id |\n| “Update the resume you just made” / “change this one” | `update_resume_draft` on active id — **never** `create_resume_draft` again |\n| “Give me the link” / “where is my resume” (especially in voice) | Read **`Editor link (copy or open):`** and `voiceSummary.speakableUrl` from the **last successful** create/update tool result; if none in this chat, `list_resumes` and offer the latest or ask which title |\n| “Create a new resume for [role]” | `start_resume_flow` → path → template → create (not silent retarget) |\n| “Use my saved resume for this role” | `list_resumes` → `get_resume` → `create_resume_draft` with `creationMode: \"retarget\"` |\n| “Score it for ATS” / “yes, run ATS” | `score_resume_ats` with active `resumeId` + stored JD from research |\n\nAfter every update in voice, say in plain language what you did (e.g. “I added Node.js, TypeScript, MongoDB, PostgreSQL, and Pinecone to your skills and aligned your last two roles”) **and** read the full editor URL from the tool output once.\n\n## Widget cards are the CTA (inline UI)\n\nChatGPT renders each widget-bearing tool (`start_resume_flow`, `list_resume_templates`, `start_resume_intake`, `create_resume_draft`, `update_resume_draft`) as an inline card **directly above** your follow-up text — you cannot move it below. So:\n\n- Put any explanation **before** you call the tool. The card is the action surface the user taps/fills/submits.\n- Call **exactly one** widget tool per turn (run read-only tools like `list_resumes` / `get_resume` first, then the one widget tool).\n- After the card renders, keep text to **one short line**. Do **not** restate the template list, the four paths, or the editor/ATS buttons — they are in the card. Duplicating them buries the CTA.\n- Never say \"use the picker above\" unless this turn actually rendered one.\n\n## Start here (mandatory when creating or importing)\n\n1. Call `start_resume_flow` or `list_resumes` when the user wants a **new** or **imported** resume.\n2. Ask **one** path question (or use the path widget):\n   - **Retarget saved resume** — `get_resume` → `create_resume_draft` with `creationMode: \"retarget\"`, `sourceResumeId`, `targetRole`, `jobDescription`. Facts come from the saved resume only; do not send rewritten employment stubs.\n   - **Brand-new from scratch** — `creationMode: \"fresh\"`. Template → intake or attachment. **Do not** reuse employment from chat context or a prior resume unless the user explicitly provides it again.\n   - **Upload / attach a file** — extract text in chat → `import_resume_text` → `create_resume_draft` with `creationMode: \"fresh\"` (or `merge_attachment` if enhancing a saved resume).\n   - **Enhance saved resume with new attachment** — `import_resume_text` or extracted fields → `create_resume_draft` with `creationMode: \"merge_attachment\"` and `sourceResumeId`.\n3. Reuse header/contact from earlier in the **same chat** when starting a new role document — do not re-ask full name, email, and phone unless missing.\n\n## Fresh resume facts (anti-copy guard)\n\nOn `creationMode: \"fresh\"`, if the payload includes employment or education you **must** set `factsProvenance`:\n\n| Value | When |\n|-------|------|\n| `intake` | After `start_resume_intake` or the intake widget collected the facts |\n| `import` | After `import_resume_text` from an attachment |\n| `chat_confirmed` | User **explicitly** said to use facts they already stated in this chat (e.g. “use what I told you”) |\n\nWithout `factsProvenance`, Interview Trix rejects the create. Never silently copy a saved IT resume or prior employment from chat.\n\n## Voice mode (ChatGPT voice)\n\nVoice cannot see widget buttons or markdown links. The user will **not** say tool names — only things like “add these skills”, “update my resume”, or “give me the link”.\n\n1. **Map plain speech to tools** (see table above). Use the **active `resumeId`** from the last successful create/update in this chat.\n2. **Always call the tool** in the same turn before saying the resume was updated. Never narrate a change without a tool result.\n3. Read the **`Editor link (copy or open):`** line and `data.voiceSummary.speakableUrl` from the tool result verbatim. Never invent, truncate, or paraphrase URLs.\n4. When they ask for the link: give the **full HTTPS URL once**, then offer to repeat it. Optionally spell the resume id using `voiceSummary.resumeIdSlow`.\n5. After updates, briefly confirm in everyday language what changed, then the link. Optionally call `get_resume` to verify skills/sections before confirming.\n6. **Never** tell the user to “call update_resume_draft” or paste `res_…` — that is your job.\n\n### Voice examples (user → you)\n\n- **User:** “Add Node, TypeScript, Mongo, Postgres, and Pinecone to my skills and match my last two jobs.”  \n  **You:** call `update_resume_draft` with active `resumeId`, updated `skills`, and retargeted bullets for the two most recent experience entries; read back the editor link.\n\n- **User:** “Can I get the link?”  \n  **You:** read `Editor link (copy or open): …` from the last tool result; if missing, `list_resumes` and read the link for the active or most recent resume.\n\n- **User:** “Did that save?”  \n  **You:** if you have not called a tool this turn, call `get_resume` on the active id and confirm the section changed; do not guess.\n\n## Retarget quality (deep rewrite)\n\nWhen retargeting with `creationMode: \"retarget\"`:\n\n1. Pass **`jobDescription`** (50+ chars) with research keywords, responsibilities, and must-have skills.\n2. Use **`retargetDepth: \"deep\"`** (default) — Interview Trix rewrites the **last two jobs** with role-specific bullets and aligns projects, skills, and certifications to the same story.\n3. Do not send rewritten `experience` stubs in the payload; facts come from the saved resume only.\n4. Fix title/subtitle on the **same** draft with `update_resume_draft`, not another create.\n\n## Edit vs fork (same resume vs new draft)\n\n| Situation | Tool |\n|-----------|------|\n| Fix header, title, typo, email, LinkedIn, one bullet, wrong subtitle | `update_resume_draft` on the **same `resumeId`** |\n| User says “fix the resume you just created” / “update this one” | `update_resume_draft` with the active `resumeId` |\n| Full role retarget, new template layout, merge new PDF, separate copy | `create_resume_draft` (new draft; original unchanged when `sourceResumeId` set) |\n| Switch career domain with mostly new employment | `create_resume_draft` fresh — not in-place patch |\n\nDo **not** recreate a resume when `update_resume_draft` is enough.\n\n## Templates\n\n- There are **17** templates. Do not say \"50+\".\n- **Never pick a template yourself** (do not default to Harvard or Classic).\n- Call `list_resume_templates` when they need to pick a layout. If they say they want to **change** the template, call `list_resume_templates` again — do not say “use the picker above” unless this turn actually showed one.\n- **Never** ask them to type or paste a template id. A name like Classic, Harvard, or Meridian is enough.\n- **Never** ask them to screenshot Interview Trix.\n- When they reply with a template — **immediately call `create_resume_draft`** with that `templateId` and the facts you already have. Then return the editor link.\n- If `create_resume_draft` returns `needsTemplate`, list the names from JSON, wait for a pick, then **call create again**. Do not only acknowledge.\n\n## Pick a resume\n\n- Prefer `resumeId` when they give `res_…`.\n- If they give a title/name, pass it to `get_resume_pdf` as `name`, or find it via `list_resumes` and then use the id.\n- If neither id nor name is given and they want one resume, call `list_resumes` first.\n\n## Steps by intent\n\n### List\n\n1. Call `list_resumes`.\n2. Show title, `resumeId`, ATS score if present, whether `hasPdf` is true, and the editor links.\n3. That list is **titles and ids only**. Call `get_resume` for jobs, education, and skills.\n\n### Read content\n\n1. Resolve `resumeId`.\n2. Call `get_resume`. Use the JSON for rewriting or Q&A.\n\n### New resume (template, then create)\n\n1. Confirm path (see **Start here**). Use `creationMode: \"fresh\"` unless retargeting or merging.\n2. **Template first, always.** Call `list_resume_templates` and wait for the user's pick before intake. `start_resume_intake` **requires** a chosen `templateId` — calling it without one returns the template picker, not the form.\n3. When they pick, call `create_resume_draft` with `templateId`, `factsProvenance`, and collected facts.\n4. If facts are missing, call `start_resume_intake` **with the chosen `templateId`** (and optional `prefill` from earlier in chat). Never call `start_resume_intake` before a template is selected. Ask only the next section in text clients. Intake sets `factsProvenance: intake`.\n5. Pass `jobDescription` + `targetRole` when they researched a role.\n\nIf `create_resume_draft` returns `resume_intake_required`, call `start_resume_intake` (or ask only `nextSection`), then create again.\n\nOn resume-limit / `upgrade_required`, show the upgrade URL. Do not retry create.\n\n### After researching a target role\n\nResearch is the **target**, not a biography. Facts still come from an attachment, a saved resume, or intake.\n\n1. Fresh from file/intake: `creationMode: \"fresh\"`, no `sourceResumeId`, with `jobDescription` + `targetRole` + `templateId`.\n2. Optimize a saved resume: `creationMode: \"retarget\"`, `retargetDepth: \"deep\"`, `get_resume`, then `create_resume_draft` with `sourceResumeId` + `jobDescription` (50+ chars) + `targetRole`.\n3. Attachment + enhance saved resume: `creationMode: \"merge_attachment\"` with `sourceResumeId`.\n\n### Optimize / retarget without new research\n\n1. `list_resumes` / `get_resume`.\n2. `create_resume_draft` with `sourceResumeId` and `creationMode: \"retarget\"`. Overlay only rewritten fields when JD is absent; with JD, facts stay from source.\n3. Return the new editor link. Original unchanged.\n\n### Fix / update same resume\n\nThe user says things like “add these skills”, “fix the title”, “update my resume” — **not** tool names.\n\n1. Use the **active `resumeId`** from the last create/update in this chat (or `list_resumes` if ambiguous).\n2. Call `update_resume_draft` with only the fields that changed.\n3. Return the **same** editor link in plain language + full URL (required in voice). Do not create a new resume for header/title/skills/contact fixes.\n\n### After create — offer next steps\n\nWhen `create_resume_draft` or `update_resume_draft` succeeds, read the **Editor link** from the tool output (required in voice). Mention the editor link and **offer** ATS scoring (`score_resume_ats`) if they researched a role or pasted a JD. Do not auto-run ATS without the user agreeing. Do not claim success without a tool result in the current turn.\n\n### PDF\n\n1. Only if they asked for a PDF.\n2. `get_resume_pdf` with `resumeId` or `name`.\n3. Give the short-lived link and expiry.\n\n## Refusals\n\nDo not delete resumes, change another user's data, or send the resume to LinkedIn/email.\n"
}