Internship Application Engine

An AI hiring partner that runs your internship hunt end to end.

It audits your profiles like a recruiter would, fixes them, finds roles across job boards and company career pages, writes a truthful one-page resume and cover letter for every job, fills in the applications up to the final screen, drafts recruiter outreach, and tracks everything in a dashboard. You review each application and click Submit.

It runs as a Claude Code skill plus a handful of small scripts. Nothing here is a hosted service; it all runs on your machine with your own logins.

Contents

1. What it does
2. How it works (architecture)
3. Requirements
4. Installation, step by step
5. First run: setting up your profile
6. The seven phases in detail
7. Daily workflow
8. Folder structure
9. Data formats
10. Scripts reference
11. Safety rules and platform limits
12. Troubleshooting
13. Sharing this setup and privacy
14. FAQ

1. What it does

Guarantees

• Nothing is invented. Tailored resumes only select, reorder and reword facts from your master resume. A fact-checker flags any number or skill that doesn't trace back to it.
• You click Submit. Claude fills the form and stops at the final screen.
• Claude asks instead of guessing on work authorization, salary or CTC, demographics, legal declarations, passwords and addresses.

2. How it works (architecture)

                         ┌──────────────────────────────┐
           you  ───────▶ │  Claude Code + this skill     │  (skill/SKILL.md = the playbook)
                         └──────────────┬───────────────┘
        ┌──────────────────┬────────────┼─────────────────┬──────────────────┐
        ▼                  ▼            ▼                 ▼                  ▼
  gh CLI (GitHub)   ego-lite browser   web search /   parallel sub-agents   local scripts
  audit, CI, pins   (your logins:      public APIs    (discovery, tailoring) (render, score,
  repo cleanup      LinkedIn, Gmail,   Greenhouse,                           fact-check, PDF,
                    Unstop, Workday…)  Lever, Ashby,                         tracker)
                                       LinkedIn guest
        └──────────────────┴────────────┴─────────────────┴──────────────────┘
                                        ▼
            resumes/   outreach/   tracker/   profile-audit/   memory/   DECISIONS.md

The resume pipeline

resumes/master/master.json ──(sub-agent picks and rewords true bullets for one JD)──▶ resumes/tailored/<Co>_<Role>_<date>.json
      │                                                                                  │
      │                                    scripts/score.py  ◀── tracker/jds/<id>.txt ────┤  keyword match: under 70 → improve, under 55 → skip
      │                                    scripts/fact_check.py ◀───────────────────────┤  nothing untraceable to the master
      ▼                                                                                  ▼
scripts/render_resume.py ──▶ .html ──▶ scripts/html2pdf.mjs (browser print, auto-fit to one A4 page) ──▶ .pdf

The application pipeline

1. Discovery agents write tracker/discovered_*.json and tracker/jds/<id>.txt.
2. build_tracker.py merges them into tracker/applications.json, removing duplicates. Every new row starts as Found.
3. Tailoring agents run 4–6 at a time in parallel, 5–6 jobs each. They write the tailored JSON and HTML, the cover letter, the answers, and tracker/results/<id>.json. Those rows become Tailored.
4. Claude renders the PDFs, fact-checks them, and fills each application in the browser. Those rows become Ready.
5. You review and click Submit, and Claude marks the row Applied with a 7-day follow-up date.
6. Outreach drafts go out after the application, once you approve them. Rows move to Messaged, then Replied, then Interview.

3. Requirements

Accounts

Software

Why ego-lite? It's a Chromium browser built for agents. Claude works inside your real logged-in sessions, you can watch every action, and you can take control at any time. Claude stops when you do. It's also how PDFs are rendered: Chrome's Page.printToPDF gives a clean, ATS-parseable text layer.

Optional

• gh auth refresh -h github.com -s user lets Claude edit your GitHub bio and location from the CLI. Pins still need the browser, signed in to GitHub.
• A password manager. Workday makes you create a separate account for every company.

4. Installation, step by step

4.1 Get the folder

mkdir -p ~/projects && cd ~/projects
git clone https://github.com/imajij/do_waqt_ki_roti.git internship-engine
cd internship-engine
cp memory/preferences.example.md memory/preferences.md          # fill in, or let Claude do it on first run
cp resumes/master/master.example.json resumes/master/master.json # Claude builds the real one from your resume

The .gitignore keeps your personal campaign data (resumes, tracker, outreach, audits, preferences) out of git, so pulling template updates never conflicts with your data and you can't push it by accident.

4.2 Install the skill into Claude Code

Symlink it so edits to skill/SKILL.md take effect immediately:

mkdir -p ~/.claude/skills
ln -sfn "$PWD/skill" ~/.claude/skills/internship-engine
ls ~/.claude/skills/internship-engine/SKILL.md   # should exist

For a single-project install, use <project>/.claude/skills/internship-engine instead.

4.3 Install and connect the browser

1. Install Ego Lite and open it once.
2. Check the CLI from a terminal:
   copy

   ego-browser nodejs -e 'const t = await taskSpace("smoke test"); await t.page("p1").goto("https://example.com"); console.log(await t.page("p1").title()); await t.finish({keep:[]});'
   

   It should print Example Domain.
3. In the Ego Lite window, log in to LinkedIn, Gmail and GitHub. Add any other boards you plan to use (Unstop, workatastartup.com).

4.4 GitHub CLI

gh auth login                               # pick SSH or HTTPS
gh auth refresh -h github.com -s user       # optional: lets Claude set your bio and location
gh auth status

4.5 Verify the scripts

python3 scripts/render_resume.py --help 2>/dev/null || head -25 scripts/render_resume.py
python3 scripts/score.py 2>&1 | head -3      # prints usage/traceback = OK, script runs

4.6 Folder skeleton (if starting empty)

mkdir -p profile-audit resumes/{master,tailored} outreach/{cover_letters/pdf,answers,drafts,posts} tracker/{jds,results} memory skill scripts
echo '[]' > tracker/applications.json
echo '[]' > outreach/sent_log.json

5. First run: setting up your profile

Start Claude Code inside the folder:

cd ~/projects/internship-engine && claude

Then give it everything in one message:

Run the internship engine for me.
Resume: ~/Downloads/My_Resume.pdf
Portfolio: https://mysite.dev   LinkedIn: linkedin.com/in/me   GitHub: myuser (orgs: X, Y)
Target roles: Backend, Full-Stack, DevOps intern
Locations: Remote, Bengaluru, Pune
Minimum stipend: ₹40,000/month (brand/learning can override)
PPO target: ≥ ₹20 LPA
Graduation: Aug 2028   Start: immediately   Duration: 2–6 months
Dream companies: …   Avoid: …
Work authorization: Indian citizen, no sponsorship needed for India
I will click Submit myself.

Claude will:

1. Copy your resume into resumes/master/ and write memory/preferences.md.
2. Run Phase 1 (the audit) and show you the report in profile-audit/<date>_audit.md.
3. Wait for your approval before changing anything. You can approve all of it ("fix everything"), or pick by ID ("approve G1, L2–L5; skip P12").

Open questions go into DECISIONS.md, numbered. Answer them in chat by number ("1 – my title was X; 12 – approved").

6. The seven phases in detail

Phase 1: Hiring-partner audit

What gets checked:

• Resume: ATS parse (text layer, single column), graduation date, vague bullets versus measurable ones, metrics you can defend, skills backed by bullets, broken links (for example private repos that 404), length, and special characters.
• LinkedIn, read in the browser: headline keywords, About, banner, Featured, a description on every role, skills (top 5 plus ~30), Open to Work titles and locations, and consistency with your resume on titles, dates and education.
• GitHub, via gh: pins, bio, profile README, secrets committed to public repos (paths only, values never shown), resume claims versus the actual code, CI and tests on flagship repos, and forks or class work cluttering your profile.
• Portfolio: time until real content appears, what a visitor learns in 5 seconds, mobile layout, dead links, a stale resume PDF, OG/SEO tags.

Every issue is written as Issue | Why it gets you rejected | Severity | Exact fix, with a "Fix first" list at the top.

Phase 2: Optimize

• Master resume: resumes/master/master.json holds every true fact, tagged by role family. Unconfirmed metrics go under verify and never reach a tailored resume.
• LinkedIn, edited in the browser: intro, About, experience descriptions, skills linked to roles, education, and Open to Work. "Share with network" prompts are always dismissed, so nothing gets posted. Every change is logged in profile-audit/linkedin-changes.md.
• GitHub:
  • Repos with leaked credentials and class work are made private.
  • Untouched forks, and repos serving GitHub Pages or linked from your portfolio, are archived instead, so they stay public and no links break. Nothing is deleted.
  • Flagship repos get CI, tests and benchmarks.
  • Topics and descriptions are added, and the profile README is rewritten.
  • Everything is logged in profile-audit/github-changes.md.

Phase 3: Find roles

Sources, highest-yield first:

1. Public ATS APIs. These are free, reliable and include the full JD:
   • Greenhouse boards-api.greenhouse.io/v1/boards/<slug>/jobs?content=true
   • Lever api.lever.co/v0/postings/<slug>?mode=json
   • Ashby api.ashbyhq.com/posting-api/job-board/<slug>
2. LinkedIn guest search (no login, at most 1 request per second): linkedin.com/jobs-guest/jobs/api/seeMoreJobPostings/search?...
3. Unstop, YC (Work at a Startup), HN "Who is hiring", Wellfound, Instahyre.
4. Naukri, Indeed and Glassdoor usually block automated access.

Roles are filtered against memory/preferences.md (role, city, stipend, eligibility such as graduation batch, work authorization, final-year-only rules), deduplicated, and ranked by fit and recency.

Phase 4: Tailor and apply

For each job:

1. Analyze the JD: required and preferred skills, responsibilities, stack, the exact ATS phrases, and eligibility blockers.
2. Tailor: a summary for this company; the 3–4 most relevant experiences and 1–3 projects; bullets reworded into the JD's vocabulary with the facts unchanged; required skills listed first. Education goes right after the summary, because graduation year is a screening filter.
3. Score with scripts/score.py. Under 70 → add truthful missing keywords. Under 55, or core requirements you don't have → skip it as a poor fit.
4. Fact-check with scripts/fact_check.py.
5. Render to PDF, auto-fit to one A4 page.
6. Cover letter (150–220 words, specific to the company) and answers to likely form questions. Salary, notice period, work authorization and demographics are marked ASK CANDIDATE.
7. Apply: Claude opens the form in ego-lite, uploads the tailored PDF, fills every field it can verify, and stops at Submit.

Review mode: the first 5 applications need your explicit approval. After that, Claude prepares applications in batches and you spot-check each one before you submit. Flag a problem and it goes back to full approval mode.

Phase 5: Outreach

• For each company you've applied to, Claude finds 1–3 people: a recruiter or university recruiter, the hiring manager, and an engineer, preferably alumni of your school.
• Each message references one specific thing about the person or team, names the role, and gives one proof point. Connection notes are 300 characters or fewer.
• Drafts go to outreach/drafts/ for approval. Once you approve a style, similar messages need less review.
• Limits: at most 15–20 messages per day, spaced out, and only after the application is submitted. One follow-up after about 7 days, never more.

Phase 6: Tracker

tracker/applications.json, viewed through tracker/dashboard.html. Statuses: Found → Tailored → Ready → Applied → Messaged → Replied → Interview → Offer / Rejected / Skipped

Phase 7: Memory and learning

• memory/preferences.md: roles, cities, stipend and PPO, timing, confirmed facts, and approved default answers.
• memory/feedback.md: wording you like or dislike, message tone, and what got replies or interviews.
• The "Lessons" section of skill/SKILL.md is updated when a pattern or failure shows up.

7. Daily workflow

cd ~/projects/internship-engine && claude

Then say "Run today's internship batch." A typical day:

1. python3 scripts/build_tracker.py --summary to see what needs you.
2. Answer anything new in DECISIONS.md.
3. Claude finds new roles, tailors the top ones, and fills the forms in your browser.
4. You submit the batch, then reply "submitted".
5. Claude marks them Applied, then drafts or sends approved outreach within the daily limit.
6. Follow-ups that are due get a one-time nudge.

Open the dashboard any time: open tracker/dashboard.html

8. Folder structure

internship-engine/
├── README.md                  ← this guide
├── DECISIONS.md               ← numbered questions Claude won't guess; answer by number
├── .gitignore                 ← blocks .env / credential files
├── skill/
│   └── SKILL.md               ← the playbook Claude follows (symlinked into ~/.claude/skills)
├── memory/
│   ├── preferences.md         ← roles, cities, stipend, timing, confirmed facts, approved defaults
│   └── feedback.md            ← what you like / what converts
├── profile-audit/
│   ├── <date>_audit.md        ← Phase-1 report
│   ├── linkedin-changes.md    ← every LinkedIn edit, before → after
│   └── github-changes.md      ← every repo made private / archived
├── resumes/
│   ├── master/                ← master.json (source of truth), master.html/.pdf, your original
│   └── tailored/              ← <Company>_<Role>_<YYYY-MM-DD>.json/.html/.pdf, one per job
├── outreach/
│   ├── templates.md           ← connection note / DM / follow-up / cold-email templates
│   ├── cover_letters/         ← <stem>.md  (+ pdf/<stem>_CoverLetter.pdf)
│   ├── answers/               ← <stem>.md  drafted application answers
│   ├── drafts/                ← outreach batches awaiting approval
│   ├── posts/                 ← LinkedIn post drafts for you to publish
│   └── sent_log.json          ← every message sent
├── tracker/
│   ├── applications.json      ← one row per opportunity (the database)
│   ├── dashboard.html         ← generated view (filters, sort, links to PDFs)
│   ├── discovered_*.json      ← raw discovery output
│   ├── jds/<id>.txt           ← full job descriptions (needed for tailoring)
│   └── results/<id>.json      ← tailoring results merged into the tracker
└── scripts/
    ├── render_resume.py       ← JSON → ATS-friendly HTML
    ├── html2pdf.mjs           ← HTML → one-page A4 PDF via the browser
    ├── score.py               ← JD ↔ resume keyword match
    ├── fact_check.py          ← flags anything not traceable to master.json
    ├── build_tracker.py       ← merge discoveries/results, build dashboard, daily summary
    ├── mark.py                ← update a tracker row's status/notes
    └── TAILOR_AGENT_PROMPT.md ← instructions given to each tailoring sub-agent

9. Data formats

resumes/master/master.json (source of truth)

{
  "name": "…", "contact": {"location","phone","email","portfolio","linkedin","github"},
  "headline_options": {"backend": "…", "fullstack": "…", "devops": "…", "fde": "…", "mobile": "…", "swe": "…"},
  "education": [{"school","degree","location","dates","details":[…]}],
  "experience": [{
     "id": "acme", "company": "…", "title": "…", "location": "…", "dates": "Jun 2026 – Present",
     "bullets": [{"text": "…", "tags": ["backend","devops"]}],
     "verify":  [{"text": "unconfirmed metric", "note": "DECISIONS #4"}]   // never used in tailored resumes
  }],
  "open_source": [{"id","name","stack","desc","link","bullets":[…]}],
  "projects":    [{"id","name","stack","desc","link","tags","bullets":[…]}],
  "skills": {"Languages": […], "Backend": […], …}
}

Tags: backend, fullstack, frontend, mobile, devops, fde, ai, rust, swe.

Tailored resume (resumes/tailored/*.json)

Same idea, flattened: summary, experience[].bullets (as strings), open_source, projects, skills, education, section_order. The full schema is documented at the top of scripts/render_resume.py.

Tracker row (tracker/applications.json)

{ "id": "stripe-software-engineer-intern", "company": "Stripe", "role": "Software Engineer, Intern",
  "source": "Greenhouse API", "jd_url": "…", "apply_url": "…", "ats": "greenhouse",
  "location": "Bengaluru", "remote": false, "stipend": null, "ppo": null, "posted_date": "2026-10-01",
  "role_category": "swe", "fit_score": 88, "match_score": 100, "flags": ["stipend unknown"],
  "status": "Applied", "date_found": "…", "date_applied": "…", "follow_up_due": "…",
  "tailored_resume": "resumes/tailored/Stripe_SWE-Intern_2026-10-07.json", "cover_letter": "…",
  "contact": "…", "messaged_on": null, "notes": "…" }

10. Scripts reference

Example of the full tailoring loop for one job:

J=tracker/jds/acme-backend-intern.txt; R=resumes/tailored/Acme_Backend-Intern_2026-10-08
python3 scripts/score.py $J $R.json --master resumes/master/master.json
python3 scripts/fact_check.py $R.json
python3 scripts/render_resume.py $R.json $R.html
echo "[{\"html\":\"$PWD/$R.html\",\"pdf\":\"$PWD/$R.pdf\"}]" > tracker/.pdf_jobs.json
ego-browser nodejs < scripts/html2pdf.mjs

11. Safety rules and platform limits

These come from real incidents during the first run.

Claude never: invents a fact, types a password, clicks a final Submit/Send/Quick Apply, posts to your feed, deletes repos, answers demographic, legal or work-authorization questions on its own, or messages anyone before you approve.

12. Troubleshooting

13. Sharing this setup and privacy

The engine contains no credentials: logins live in your browser and the gh keyring. It does contain personal data: resumes with your phone and email, your address if you saved it to preferences, job history, and contacts.

This repo is that clean template. The .gitignore excludes every personal folder. If you keep your own campaign under version control, use a private repo for it.

To build a clean template from a used folder by hand:

cp -R internship-engine internship-engine-template && cd internship-engine-template
rm -rf resumes/master/* resumes/tailored/* tracker/*.json tracker/jds/* tracker/results/* \
       outreach/cover_letters/* outreach/answers/* outreach/drafts/* outreach/posts/* \
       profile-audit/* memory/* DECISIONS.md
echo '[]' > tracker/applications.json; echo '[]' > outreach/sent_log.json
grep -rIl -e '@gmail.com' -e '+91' . || echo "clean"

Keep skill/, scripts/, outreach/templates.md, README.md and .gitignore.

14. FAQ

Will it apply without me? No. It prepares and fills applications; you press the final button. That keeps you in control and keeps your accounts within platform rules.

How many applications a day? 10–20 high-quality, tailored ones is a realistic, safe pace: about 10 LinkedIn Easy Apply, plus any number on Greenhouse, Lever or Workday.

Can it lie to beat the ATS? No. Keywords are only added when a true fact supports them. Gaps are listed (true_gaps) so you can close them for real.

What if my resume has a metric I can't back up? It goes in verify in the master and stays off tailored resumes until you confirm how it was measured.

Can I use it for full-time roles? Yes. Change the roles in memory/preferences.md and the headline_options in your master.

How do I teach it my style? Just say "I don't like X" or "this framing got replies". It's written to memory/feedback.md, and the defaults in skill/SKILL.md are updated.