GhostCheck

A Claude Code skill pack that audits why senior job applications get silence instead of rejection.

Existing CV checkers score keywords and ATS compatibility. The real silence drivers at senior levels are invisible to them — bucket mismatch (CV reads one tier below target), IT-services discount, channel mismatch (Easy Apply vs referral), theater postings, missing online surface, stale postings.

GhostCheck simulates the entire screening funnel using 11 narrow, specialised agents running in blind fan-out — every agent judges one dimension of silence risk in its own isolated context window. A deterministic weighted-logistic aggregator (no second LLM call) combines the verdicts into a callback probability with ranked silence drivers and evidence-cited fix hints.

What's interesting

• Pattern B subagent isolation — true blind fan-out per Claude Code's documented subagent runtime
• Strict AgentVerdict schemas with fail-closed coercion at every boundary
• Deterministic aggregator — refuses to interpret, only does math
• Evidence citation requirement on every verdict — no opinion without grounding
• The harness is the moat — ports cleanly across runtimes (Claude Code today, Python service or Gemini/Ollama wrapper tomorrow, same files)

Why this architecture

The thesis: any individual agent prompt is replicable; what's hard to replicate is the harness — shared schema contracts, fail-closed coercion, capability declarations in frontmatter, the deterministic combiner. See docs/HARNESS_ENGINEERING.md for the full reasoning.

---

Quick Start — from zero to your first audit

1. Prerequisites

Required for every audit:

• Claude Code installed (desktop, web, or a supported IDE extension).
• An active Claude subscription or API key.
• git installed locally.

Optional — only needed if your CV or JD is in DOCX, PPTX, or HTML format:

• MarkItDown installed with the relevant extras package:
  • pip install 'markitdown[all]' — supports every input format (recommended).
  • pip install 'markitdown[pptx]' — PowerPoint only.
  • pip install 'markitdown[docx]' — Word only.
• The single-quotes around the package name are required on macOS because the default shell (zsh) interprets unquoted square brackets as a glob pattern and the install will fail with zsh: no matches found. On bash either form works; quoting is harmless.
• Note that the bare pip install markitdown (without extras) installs the CLI but cannot actually read PPTX, DOCX, or HTML — the install command needs the extras.

PDF and Markdown files work out of the box — Claude Code's native file reading handles them without any install. The MarkItDown CLI with extras is only needed when you have a Word document, PowerPoint deck, or HTML page that needs to be converted to markdown before the agents can read it. If you have a DOCX or PPTX you can also convert it to PDF first (most operating systems have built-in export to PDF) and avoid the install entirely.

2. Get the repo

Fork it on GitHub (recommended — keeps your audits private in your own fork), then clone your fork:

git clone https://github.com/<your-username>/Ghost_Check.git
cd Ghost_Check

Or clone directly if you just want to try it out:

git clone https://github.com/santosha86/Ghost_Check.git
cd Ghost_Check

3. Open it in Claude Code

Open the Ghost_Check folder as a Claude Code workspace. This lets Claude Code see the .claude/skills/ folder and register the /ghostcheck command.

4. Drop in your CV

Copy the example CV and replace it with yours:

cp profile/cv.example.md profile/cv.md

Then open profile/cv.md and paste your CV. If your CV is a PDF or DOCX, put the file in profile/ and GhostCheck will parse it for you when you run the audit.

5. Set your target profile

Copy the example profile config and fill in your target titles, locations, and seniority level:

cp config/profile.example.yml config/profile.yml

This takes one minute — it tells the agents what "target role" actually means for you.

6. Add the job description you want audited

Save the JD as a markdown or text file anywhere in the repo — for example jobs/acme-senior-ai.md. Paste the full JD text in the body.

Optional but recommended — add a short YAML header at the top of the JD file so the audit can use company context without a web lookup:

---
company: Acme Fintech
size: 250
hq: Dubai, UAE
industry: fintech
funding_stage: Series B
posted_date: 2026-03-15
source_url: https://acme.com/careers/senior-ai-architect
---

# Senior AI Architect

<rest of the JD as plain markdown>

Every field is optional. If the header is missing or partial, GhostCheck fills gaps via web search (only the agents that declare web_search in their capabilities can do this). If both sources fail, the agents that need company context return UNKNOWN instead of guessing — fail-closed.

7. Run the audit

In your Claude Code session:

/ghostcheck audit --cv profile/cv.md --jd jobs/acme-senior-ai.md

8. Read your result

Your audit lands at applications/YYYY-MM-DD_acme-senior-ai/audit.md. Open it.

What you get back

A markdown report with a 0–100% callback probability score, the top silence drivers for this specific CV+JD pair (ranked by severity — usually two to five), and an evidence-cited fix suggestion for each. Every finding links back to the exact line of your CV, the JD, or the external context that triggered it.

Treat the findings as prioritised hypotheses to investigate, not guarantees. They are diagnostic signals from pattern-matching agents, grounded in cited evidence — not verdicts from an oracle. Your judgement is still the final step.

---

Runtime

GhostCheck V1 runs end-to-end inside Claude Code, which is why the Quick Start above uses Claude Code commands. That is what the /ghostcheck router and the skill files expect today.

The architecture, however, is deliberately runtime-portable:

• Every agent is a plain markdown file with YAML frontmatter declaring its inputs, outputs, and capabilities.
• Data contracts in docs/SCHEMAS.md are Pydantic-ready.
• The aggregator is deterministic math, not another LLM call.

That means a future Python service, a different orchestrator, or a different LLM provider can drive the same 11 agents without rewriting any of them. The reasoning behind this design — and why it is the actual moat of the project — is documented in docs/HARNESS_ENGINEERING.md. The bring-your-own-LLM path is on the roadmap under V1.2.

---

The 11 agents at a glance

The full prompt, inputs, outputs, and severity model for each agent live in its own .md file under .claude/skills/agents/. Short version:

ID	Agent	What it checks
A1	google-test	What a recruiter sees about you when they Google you before your CV.
A2	posting-decoder	Whether the JD is a genuine opening or a theater (pre-filled) posting.
A3	bucket-classifier	Whether your CV reads at the target seniority or one level below.
A4	it-services-discount	Whether tenure at large IT-services firms triggers a silent downgrade.
A5	headline-filter	Whether your one-line identity passes the six-second recruiter filter.
B1	funnel-math	Whether your application-to-callback rate is normal or broken.
B2	channel-mix	Whether you are applying through the right channel for this seniority.
B3	stale-detector	Whether you applied too long after the JD was posted.
C1	ats-simulator	Keyword coverage, years match, title match (classic ATS score).
C2	recruiter-30sec	First-scan impression: companies, tenure, trajectory.
C3	hm-deep-read	Whether your bullets show decisions owned or activities performed.

Note: the A / B / C prefix marks the tier (and is the filename prefix — e.g. .claude/skills/agents/A3-bucket-classifier.md). The value stored in every AgentVerdict.agent_id is the second column — e.g. bucket-classifier, not A3.

Every agent returns the same shape (AgentVerdict) defined in docs/SCHEMAS.md. Agents run blind to each other — no agent sees another agent's output — which keeps the aggregator's input independent.

---

Example output (illustrative)

This example is illustrative only. The first real audit lands end of Day 2 of the build and will replace this block.

# Audit — Senior AI Architect @ Acme
Date: 2026-04-22
Callback probability: 34%

## Top silence drivers
1. HIGH — bucket-classifier: CV reads as Staff, JD targets Director.
   Evidence: no owned P&L line, no org-size metric in last 3 roles.
   Fix: re-scope two bullets to decision ownership + org-size metric.

2. HIGH — google-test: No discoverable surface for this title.
   Evidence: top 10 Google results for your name return internal-only posts.
   Fix: publish one 600-word post on the agentic architecture you led.

3. MEDIUM — channel-mix: 80% of your applications are Easy Apply.
   Evidence: applications/ log shows 12 of 15 via Easy Apply.
   Fix: route next 3 applications through referral or direct DM.

---

Architecture in one paragraph

GhostCheck is a skill pack, not a Python application. Every agent is a markdown file with YAML frontmatter, Claude Code handles orchestration, and the aggregator is deterministic math — not another LLM. Nothing calls out to a server, nothing holds state beyond the files in this repo, and every file is human-readable. The reason this matters, and why it is the real differentiator of the project, is covered in docs/HARNESS_ENGINEERING.md.

---

Trust and privacy

Your CV never leaves your machine unless you push it to a remote yourself. Agents cannot call each other, cannot write to your files, and declare their capabilities in frontmatter (for example, only the google-test agent is allowed to call web search). Any agent that cannot back a verdict with cited evidence returns UNKNOWN rather than inventing one — this is called fail-closed.

The full security posture lives in docs/ZERO_TRUST.md.

---

Roadmap

• V1 (now) — single-application audit. Eleven agents, four enrichment skills, deterministic weighted-logistic aggregator, markdown audit artefact. Functional and shipped.
• V1.1 — batch pattern analysis (the killer feature: your personal silence signature across multiple ghosted applications) plus DSPy+GEPA self-evolution of agent prompts using applications/ as execution traces. Sharable PNG cards and architecture slide deck land here.
• V1.2 — calibration loop with real callback outcomes (tunes weights and aggregator hyperparameters from data, not intuition); bring-your-own-LLM provider abstraction (Gemini, Ollama, etc.); the Karpathy LLM Wiki pattern for compounding cross-session knowledge; runtime Zero Trust policy enforcement.
• V2 — Python / FastAPI runtime, web frontend, Microsoft AI Observability framework alignment, hosted multi-user runtime for career coaches and teams.

Full detail (Hermes/GEPA references, calibration mechanism, V2 architecture, contribution guidelines) in docs/ROADMAP.md.

---

License

MIT. See LICENSE.

Built on top of MarkItDown for CV and JD parsing.