comeback

An interactive assistant that supports a laid-off professional through the whole arc of finding their next job — from the shock of the layoff to signing the next offer. One conversation, in Bulgarian, that covers benefits, CV, job-search strategy, and interviews — with a steadying, human layer running underneath all of it.

• Language (what you see): Bulgarian
• Who it's for: any laid-off professional — junior to senior, any industry
• Where it runs: Claude Code, or OpenCode for free (free tool + free built-in models — no paid AI plan needed) — it's an agent you talk to, not a website
• Sibling project: the uncoachable.work blog; the guide that grounds this agent is also published there

Why this exists

The market is full of single-purpose tools: CV builders, auto-appliers, interview drills. People stack two or three of them because none covers the whole journey — and every one optimizes a document or a performance while ignoring the person. There is almost nothing good in Bulgarian, and nothing that holds someone steadily across the entire process.

comeback does the opposite: one assistant for the entire comeback, in your language, that also handles the thing Bulgarian tools ignore — the money-and-deadlines side of losing a job — and never loses sight of the human going through it.

What it does

It's not a CV tool with extras. It covers six things, and it decides which one you need by talking to you:

• 📋 Benefits & the legal clock (Bulgaria-specific) — Are you owed unemployment benefit, how much, and by when must you act? It works out your actual Бюро по труда deadline (7 working days) and НОИ deadline (3 months) from your termination date, estimates your benefit, and orients you on severance and the electronic trudova kniga at НАП. The dates and amounts are computed deterministically and verified against official НОИ/НАП sources — never guessed.
• 📄 CV / resume — Build one from scratch or clean up what you have, tailor it to a specific posting, sanity-check it for ATS, and reframe the layoff honestly. You get a finished .docx and a .pdf.
• 🎯 Job-search strategy — A networking-first plan, a living application tracker the assistant maintains for you, and outreach drafts — so you're not just firing CVs into the void.
• 🗣️ Interview practice — Mock interviews with real feedback, the "why did you leave?" story worked out until it lands, and delivery.
• 🔎 Company research (optional) — For a specific posting: an honest fit read, concrete "why here" angles, and sharp questions to ask.
• 🫂 Emotional footing — An always-on steadying layer underneath everything. You never have to ask for it; it's the part competitors skip.

A note on trust: anything the assistant tells you about benefits, deadlines, or documents is checked by a separate verifier against the official source before you see it. If something can't be confirmed, it says so and points you to НОИ/Бюро по труда rather than handing you a confident wrong number.

How to use it

Before you start — pick how you'll run it

The assistant is the same either way; the only difference is the tool you talk to it in.

Option A — OpenCode (free, no paid plan). OpenCode is free and ships with free built-in models (OpenCode Zen), so you can run the whole assistant at no cost — no Claude subscription, no API key.

1. Install OpenCode and open this project folder in it.
2. Start a new session and pick the Comeback agent (the agent selector, bottom-left).
3. Say what happened — it opens with triage and takes it from there. The default free model works out of the box; the full benefits → CV → search → interview flow runs on it.

Option B — Claude Code. The tool this project is authored in.

1. Open this project in Claude Code with this folder as the working directory — that's what puts the subagents and the CV skill in scope. (On a plain Claude.ai project without them, only the conversation works — no file output.)
2. Type /comeback, or just say what happened.

Optional (either tool): install LibreOffice if you want CV PDFs rendered locally with correct Cyrillic. Without it you still get the .docx.

Just talk to it

Whichever tool you picked, say what happened — in Bulgarian or English:

„Съкратиха ме." · „Останах без работа." · „laid off"

It opens with a short, human check-in — what happened and where you are — and then routes to whatever you need. You don't pick a mode or memorize commands; you just talk to it. (In Claude Code you can also type /comeback; in OpenCode you pick the Comeback agent first, then talk.)

What you can ask for

You steer it in plain language. A few examples of what to say and what happens:

CV only? If all you want is a tailored CV, you can invoke /tailor-cv directly.

What you get, and where

Everything produced for you lives in a private, git-ignored Personal/<date>-you/ folder — so nothing about your situation is ever committed to the repo:

• a dossier — the assistant's running memory of your situation (so it never re-asks)
• your tailored CV (.docx + .pdf)
• a job-search tracker
• interview notes

How it works

One orchestrator is the single voice you talk to. It holds the relationship — the check-in, the steadying layer, and one consistent tone — and never hands that off. For concrete work it quietly delegates to specialized subagents (CV, interview, search, benefits, company research), each of which reads and writes your shared dossier. A separate freshness-checker independently verifies every benefits/legal claim before it reaches you. You never see the machinery — just one assistant.

agent/            orchestrator (the voice), triage, always-on support, dossier schema
.claude/agents/   the subagents (authored here): cv-builder, interview-coach, search-strategist,
                  bg-navigator, company-intel, and the freshness-checker verifier
.claude/skills/   /comeback (the front door) and /tailor-cv (the CV engine)
.opencode/agents/ the same agents, generated for OpenCode from .claude/ (tools/gen_agents.py)
knowledge-base/   the Bulgarian guide + BG legal/benefits facts + sources
scripts/          harness-independent CV .docx/.pdf + tracker .xlsx output

The assistant is authored once in .claude/; the other tools' trees are generated from it with python tools/gen_agents.py, so they run the same agent. Contributors edit .claude/ and regenerate — they never hand-edit the generated trees. Per-tool model choices live in tiers.json.

• OpenCode (.opencode/agents/) — generated and validated end-to-end (free to run).
• Codex (.codex/agents/) and Copilot (.github/agents/) — generated and structurally linted, but not yet behaviorally tested (those tools aren't set up here). Treat them as experimental; the agent bodies are the same, but each tool's entry point and web-search wiring still need a real run to confirm.

Full architecture and conventions are in CLAUDE.md.

Status

Shipped and stable (see VERSION). The full orchestrator → subagent → verification flow works end-to-end and is backed by an automated test suite (tests/), enforced before every push by a local test gate (see CONTRIBUTING.md). History is in CHANGELOG.md.

Contributing

See CONTRIBUTING.md, and CLAUDE.md for the full conventions.

License

Dual-licensed. The code and everything outside knowledge-base/ is under the MIT License. The written content in knowledge-base/ — the guide, bg-legal, and sources — is under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0): share and adapt with attribution, non-commercial only.