cover-letter · diff
git:20260929.e0789b7 to git:20260929.2038fcc
42 added, 57 removed. Audit A to A.
---
name: cover-letter
- description: Write a tailored, concise cover letter (150-250 words) from a job description and the user's resume, humanized for natural tone.
+ description: Write a short cover letter (150-250 words) for a job from the user's resume, in a plain human voice that doesn't repeat the user's recent letters.
argument-hint: "<job_description>"
---
- # Cover Letter Generator
+ # Cover Letter
- Write a short, tailored cover letter connecting the candidate's resume to a specific role.
+ Write the letter a good candidate would write in fifteen minutes: short, specific, plain. The reader should learn one or two real things about the candidate and want to talk to them.
## Setup
- Follow `../_shared/setup.md` to load profile and resume. That structured resume is what you write from - experience, projects, skills, education, publications, awards, certifications and any custom sections are all in it. Don't also `Read` the source PDF: it's the same content again at several times the size, and inside an apply flow that cost is paid per job.
+ Follow `../_shared/setup.md`, then load the structured resume (the caller's `resumeId`, else `user.primaryResumeId`):
- ## Step 1: Analyze the JD
+ ```bash
+ jobpilot-api GET "/api/resumes/$RESUME_ID"
+ ```
- From the argument, identify: company + what they do, role title and level, key responsibilities, required/preferred qualifications, tech stack and domain, culture cues.
+ Write from its `content`. Don't `Read` the source PDF: it's the same content at several times the size, and inside an apply loop that cost is paid per job.
- ## Step 2: Read Your Last 5 Letters
+ ## Step 1: Read the job
- Do not skip. Letters that are individually fine and collectively identical are what reads as AI - not word choice.
+ From the argument, note the company and what it does, the role and level, the two or three things the job most needs, the stack and domain, and the tone (startup casual or enterprise formal).
+ ## Step 2: Read your last five letters
+
+ Letters that are each fine but all the same shape are what reads as AI. Don't skip this.
+
```bash
- jobpilot-api GET /api/cover-letters --query page=1 --query pageSize=5
+ jobpilot-api GET /api/cover-letters --query page=1 --query limit=5
```
- The list is metadata only, so `GET /api/cover-letters/<id>` for each `.items[].id` to get the body. No history (first letter, or the call fails) → Step 3.
+ The list is metadata only, so `GET /api/cover-letters/<id>` for each `.items[].id`. From each, keep four things and drop the rest: what it led with, its first sentence, its last sentence, and how it phrased its numbers. No history → skip to Step 3.
- Take three things from each and drop the rest: which experience it **led** with, its metric phrasings, and its closing sentence. Those are what rules 14-16 compare against; keeping five whole letters in context buys nothing beyond them.
+ ## Step 3: Choose what to say
- ## Step 3: Select Relevant Experience
+ Pick one lead, the strongest for this job:
- From the resume, pick the most relevant: 2–3 work experiences, 2–3 projects, research (if AI/ML/CV), education (if relevant to level).
+ - the result closest to their main need (for senior roles: scope, architecture, what the team shipped)
+ - direct experience with their stack, domain, or product (for research roles: publications)
+ - a real, specific reason for this company, only if the resume or JD gives you one
- **Apply rules 14-16 here**, against Step 2's letters. Rotating the lead is a selection decision - make it before drafting, not by patching a draft that already opened with the usual paragraph.
+ Don't lead with what either of the last two letters led with, unless the JD leaves no honest alternative (a healthcare role wants the healthcare project).
+ Then pick one or two proof points that fit this job, not your two best overall: a named project or role with one concrete detail each.
+
## Step 4: Write
- **Header** - values strictly from `user.*`, never from the resume file (resumes carry stale addresses). Omit any line whose fields are empty:
+ **Header.** Values from `user.*` only, never the resume (resumes carry stale addresses). Drop any line whose fields are empty:
```
[Full Name]
[City, State] | [Phone] | [Email]
[LinkedIn] | [GitHub] | [Website]
```
- **Body - 150–250 words, 3–4 paragraphs of uneven length.**
-
- Pick ONE lead angle - the strongest for this JD, not a stack of all three:
-
- - a result that matches their core need (senior/lead roles: scope, architecture, team output)
- - direct experience with their exact stack, domain, or product (AI/ML/research roles: publications)
- - a genuine, specific reason for this company - only if you actually have one
-
- Then write:
+ **Body.** 150-250 words in three or four paragraphs of different lengths:
- - **Lead (1–3 sentences):** the role + your angle. **No** "I'm excited to apply" / "I'm writing to express my interest".
- - **Proof (1–2 short paragraphs):** 1–2 proof points from the resume - a named project or outcome with one concrete detail each. Pick the points that fit *this* JD, not your two best overall. Let each stand on its own; don't explain why it's relevant - if you picked right, it's obvious.
- - **Close (1–2 sentences):** you'd like to talk, thanks. No recap of fit.
+ - **Open** with the role and your lead in one to three sentences.
+ - **Proof** in one or two short paragraphs. Say what you did and what happened. Don't explain why it's relevant; if you picked well, the reader sees it.
+ - **Close** in one or two sentences: you'd like to talk. A thank-you is fine. No recap of fit.
- If the draft came out as opening → experience → tech depth → why-company → closing, that's the template every AI letter follows - merge or cut a section and vary where the company mention lands.
+ Quote at most one short phrase from the JD. Write in first person as the candidate, only from resume facts, in the company's tone.
**Sign-off:**
```
Best regards,
[Full Name]
```
- ## Step 5: Apply Humanizer
+ ## Step 5: Humanize and compare
- Invoke the `humanizer` skill on the full text in **embedded mode** - final rewrite only, no draft or audit bullets.
+ Invoke the `humanizer` skill on the full letter in embedded mode.
- Then re-check against Step 2's letters: no shared sentence, different lead, different closing. The humanizer sees one document at a time, so this check is yours.
+ Then compare against the Step 2 notes; the humanizer sees one letter and can't do this. Fix anything that matches:
- ## Step 6: Save to History
+ - no sentence reused, including how a metric is phrased (reuse the number, not the sentence)
+ - a different first and last sentence
+ - the company mentioned in a different place
+ - a different overall shape: if the last letters went intro, experience, tech, why-company, close, merge or drop a section
- Persist the final letter so it's reviewable in the web app. Best-effort - if the call fails, continue. Write `{"content":"<final letter text>","jobUrl":"<job url>","jobTitle":"<role title>","company":"<company>","source":"<source>"}` to `$JOBPILOT_TEMP/cover-letter.json` (omit `jobUrl`/`jobTitle`/`company` when empty), then:
+ ## Step 6: Save to history
+ Best effort; continue if it fails. Write `{"content":"<final letter>","jobUrl":"<job url>","jobTitle":"<title>","company":"<company>","source":"<source>"}` to `$JOBPILOT_TEMP/cover-letter.json`, omitting empty `jobUrl`/`jobTitle`/`company`:
+
```bash
jobpilot-api POST /api/cover-letters --data @"$JOBPILOT_TEMP/cover-letter.json"
```
- `jobUrl`/`jobTitle`/`company` come from the JD argument (`$DIGEST` fields when present). `source` is the invoking context - `apply`, `auto_apply`, or `manual` (default `manual` when the caller didn't specify).
-
- ## Rules
-
- 1. **150–250 words for the body.** Short reads as confident.
- 2. **No fluff** - drop "passionate", "dedicated", "committed", "excited", "thrilled", "leverage", "utilize", "innovative", "cutting-edge", "eager", "dynamic".
- 3. **No generic openings.**
- 4. **Be specific.** Real project names, metrics, technologies from the resume.
- 5. **Tailor by selection, not narration.** Prove fit by choosing the right material. Quote back at most ONE short JD phrase in the whole letter. Never end a paragraph by mapping yourself onto the JD - no "which is exactly what your team needs", "the same kind of X that Y requires", "which is basically the job description".
- 6. **End paragraphs on the fact** - what you did or why you're interested - not a tie-back or a flourish.
- 7. **One dash clause max** per letter (em dash or " - "); restructure the rest into commas or separate sentences.
- 8. **One contrast max** per letter - "X rather than Y" / "not X, but Y" is a crutch; usually delete the negative half.
- 9. **No rule-of-three.** Don't default to three-item lists; name one thing, or two, or four.
- 10. **Show, don't tell.** No "I'm a strong communicator" - demonstrate it through the writing.
- 11. **Match tone.** Startup → conversational; enterprise/gov → formal.
- 12. **No fabrication.** Only reference projects/skills from the resume.
- 13. **First person** as the candidate.
-
- ### Anti-repetition (against Step 2's letters)
-
- 14. **Rotate the lead.** Don't open with the experience either of the last two letters opened with, unless the JD makes it the only honest choice (a healthcare role genuinely wants the healthcare project). A recruiter comparing two of your letters is comparing openings.
- 15. **No sentence reused verbatim**, metric phrasings included. A reused number is fine; a reused *sentence about* it is the tell. Rephrase the result or lead with a different one.
- 16. **Vary the closing and the shape.** No repeated closing, and never "how this maps to/onto <Company>". Move the company mention too - same skeleton with different nouns is still a template.
+ Take `jobUrl`/`jobTitle`/`company` from the JD argument (`$DIGEST` fields when present). `source` is the caller: `apply`, `auto_apply`, or `manual` (the default).
## Output
- Plain text with header and sign-off, ready to paste or convert to PDF.
+ Plain text with the header and sign-off, ready to paste or render to PDF.