AI prompt-engineering. Job Search System

What This Prompt Does & Why It's Powerful

Demo

Demo version 3.0 Demo version 3.0

Demo version 4.0 Demo version 4.0

Overview

This is not a simple job search query. It is a structured, multi-source intelligence system designed to surface every active Senior Angular / Front-End Developer opportunity in the UK market — across commercial, government, and remote EU sectors, with a focus on verified, real job links and zero fabricated results.

When executed, it runs 14 sequential search steps across 13 distinct job platforms, applies strict quality control to every result, and outputs a fully formatted, horizontally scrollable HTML report with colour-coded badges, live apply links, and deduplication across sources.

The Problem It Solves

Most job searches suffer from three critical failure modes:

1. Too narrow — searching one or two boards misses roles posted exclusively elsewhere (e.g. TotalJobs surfaced an investment bank contract at £nnn/day not found on Reed or CWJobs)
2. Too noisy — no filtering logic means junior roles, pipeline ads, and speculative CV grabs pollute results
3. Unverifiable — AI-assisted searches often return plausible-sounding but fabricated or dead links

This prompt eliminates all three by enforcing URL verification before any role is included, applying hard exclusion rules for seniority, and separating results into clearly defined sectors and location categories.

🏗️ Angular Job Search — System Design Spec

Version 1.0 | February 2026 search-job-prompt4.1.md Implements: Evaluation Loops (Principle 3) · Latency & Cost Discipline (Principle 5)

Overview

The prompt alone handles what the model searches, validates, and renders. This spec defines everything around the model: how runs are triggered, how results are compared across versions, how drift is detected, how costs are controlled, and how outputs are cached and served.

Without this layer, the prompt is a good one-shot tool. With it, it becomes a measurable, maintainable production system.

Architecture Diagram

┌─────────────────────────────────────────────────────────────┐
│                      TRIGGER LAYER                          │
│         Scheduler · Manual UI · Webhook · CLI               │
└─────────────────────┬───────────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────────┐
│                   ORCHESTRATION LAYER                        │
│    Step Prioritiser · Token Budget Manager · Retry Logic     │
└──────┬──────────────┬──────────────────────┬────────────────┘
       │              │                      │
┌──────▼──────┐ ┌─────▼──────┐ ┌────────────▼───────────────┐
│ CACHE LAYER │ │ MODEL CALL │ │      FALLBACK ROUTER        │
│ Static URLs │ │ Claude API │ │ 403 → flag · 404 → exclude  │
│ Section 4   │ │ Prompt 4.1 │ │ Timeout → retry (max 2)     │
└──────┬──────┘ └─────┬──────┘ └────────────┬───────────────┘
       │              │                      │
┌──────▼──────────────▼──────────────────────▼───────────────┐
│                    OUTPUT LAYER                             │
│     Schema Parser · QC Validator · HTML Renderer            │
└──────────────────────┬──────────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────────┐
│                   EVAL & STORAGE LAYER                       │
│   Run Logger · Version Comparator · Drift Detector · DB      │
└─────────────────────────────────────────────────────────────┘

Principle 3 — Evaluation Loops

3.1 Run Storage Schema

Every prompt execution produces a run record. Store in a lightweight database (SQLite locally, Postgres in production, or even a flat JSON file for solo use):

{
  "run_id": "uuid",
  "prompt_version": "4.1",
  "executed_at": "ISO8601 timestamp",
  "steps_attempted": 14,
  "steps_successful": 10,
  "steps_blocked": 3,
  "steps_empty": 1,
  "roles_found_raw": 42,
  "roles_excluded": 11,
  "roles_output": 31,
  "exclusions": {
    "junior_title": 4,
    "no_angular": 2,
    "unverified_url": 3,
    "duplicate": 2,
    "pipeline_ad": 0,
    "other": 0
  },
  "confidence_distribution": {
    "high": 18,
    "medium": 9,
    "low": 4
  },
  "field_completeness": {
    "salary_rate_confirmed": 24,
    "ir35_confirmed": 19,
    "date_confirmed": 27,
    "url_verified": 28
  },
  "tokens_used": 18420,
  "cost_usd": 0.23,
  "output_file": "Angular_Jobs_2026-02-18.html",
  "known_gaps": ["EU remote section empty", "Hiring.cafe blocked"]
}

3.2 Automated Evaluation Checks

Run these checks automatically after each execution and flag regressions:

def evaluate_run(current_run, previous_run):
    checks = []

    # Role count regression
    delta = current_run.roles_output - previous_run.roles_output
    if delta < -5:
        checks.append(("WARN", f"Role count dropped by {abs(delta)} vs previous run"))

    # Confidence degradation
    prev_high_pct = previous_run.confidence_distribution["high"] / previous_run.roles_output
    curr_high_pct = current_run.confidence_distribution["high"] / current_run.roles_output
    if curr_high_pct < prev_high_pct - 0.15:
        checks.append(("WARN", "High-confidence role % dropped >15pp vs previous run"))

    # Exclusion rate spike
    prev_excl_rate = previous_run.roles_excluded / previous_run.roles_found_raw
    curr_excl_rate = current_run.roles_excluded / current_run.roles_found_raw
    if curr_excl_rate > prev_excl_rate + 0.20:
        checks.append(("WARN", f"Exclusion rate spiked: {curr_excl_rate:.0%} vs {prev_excl_rate:.0%}"))

    # Source regression
    if current_run.steps_blocked > previous_run.steps_blocked + 2:
        checks.append(("WARN", "More sources blocked than previous run — check board access"))

    # Cost spike
    if current_run.cost_usd > previous_run.cost_usd * 1.5:
        checks.append(("WARN", f"Cost spiked: ${current_run.cost_usd:.2f} vs ${previous_run.cost_usd:.2f}"))

    return checks

3.3 Test Dataset for Prompt Evaluation

Maintain a golden dataset of known roles used to measure prompt accuracy across versions. Update monthly.

{
  "test_dataset_version": "2026-02",
  "cases": [
    {
      "id": "TC-001",
      "description": "Should be INCLUDED — confirmed Senior Angular, remote UK, outside IR35",
      "input_snippet": "Senior Angular Developer | TypeScript | RxJS | Fully Remote UK | Outside IR35 | £450/day",
      "expected": { "excluded": false, "ir35": "outside", "work_mode": "remote_uk", "confidence": "high" }
    },
    {
      "id": "TC-002",
      "description": "Should be EXCLUDED — Junior in title",
      "input_snippet": "Junior Angular Developer | Remote | £30,000",
      "expected": { "excluded": true, "exclusion_reason": "junior_title" }
    },
    {
      "id": "TC-003",
      "description": "Should be EXCLUDED — Angular secondary",
      "input_snippet": "Senior React Developer with some Angular experience | London | £80,000",
      "expected": { "excluded": true, "exclusion_reason": "no_angular" }
    },
    {
      "id": "TC-004",
      "description": "IR35 must be TBC, not inferred — government role without explicit IR35 statement",
      "input_snippet": "Front End Developer | DWP | SC Cleared | London | £500/day",
      "expected": { "ir35": "tbc", "clearance": "sc_active" }
    },
    {
      "id": "TC-005",
      "description": "Salary must be null — listing says 'competitive salary'",
      "input_snippet": "Senior Angular Developer | FinTech | Competitive salary | Remote UK",
      "expected": { "salary_perm": null, "day_rate": null }
    },
    {
      "id": "TC-006",
      "description": "Should be low confidence — URL not fetchable",
      "input_snippet": "Senior Angular Developer | Company X | [403 on fetch]",
      "expected": { "source_url_verified": false, "confidence": "low" }
    },
    {
      "id": "TC-007",
      "description": "Should be EXCLUDED — EU role, no explicit UK acceptance",
      "input_snippet": "Senior Angular Developer | Amsterdam | €80,000",
      "expected": { "excluded": true, "exclusion_reason": "relocation_required" }
    },
    {
      "id": "TC-008",
      "description": "Should be INCLUDED as remote_eu — explicitly states UK accepted",
      "input_snippet": "Senior Angular Developer | Remote EU | UK candidates welcome | €90,000",
      "expected": { "excluded": false, "work_mode": "remote_eu", "sector": "eu_remote" }
    }
  ]
}

3.4 Version Comparison Report

Generate this diff report whenever the prompt version changes:

## Prompt Version Comparison
Previous: v3.1 → Current: v4.1
Runs compared: [run_id_old] vs [run_id_new]

| Metric                  | v3.1  | v4.1  | Delta  |
|-------------------------|-------|-------|--------|
| Roles output            | 23    | 31    | +8 ✅  |
| High confidence %       | 48%   | 58%   | +10pp ✅|
| Exclusion rate          | 18%   | 26%   | +8pp ⚠️ |
| URL verified %          | 74%   | 90%   | +16pp ✅|
| IR35 confirmed %        | 61%   | 71%   | +10pp ✅|
| Steps blocked           | 4     | 3     | -1 ✅  |
| Cost (USD)              | $0.18 | $0.23 | +$0.05 ⚠️|
| Token usage             | 14200 | 18420 | +4220 ⚠️|

Regressions flagged: Exclusion rate increased — investigate if new QC rules are too aggressive
Improvements: URL verification, IR35 completeness, overall role count

Principle 5 — Latency & Cost Discipline

5.1 Token Budget Manager

Before executing optional steps (Steps 9–14), estimate remaining token budget:

TOKEN_LIMITS = {
    "mandatory_steps": 8000,    # Steps 1–5: search + schema population
    "important_steps": 4000,    # Steps 6–8
    "optional_steps": 3000,     # Steps 9–14
    "render_budget": 5000,      # HTML generation
    "eval_block": 500,          # Self-evaluation block
    "buffer": 1000,             # Safety margin
    "total_target": 20000       # Stay under this per run
}

def should_run_optional_steps(tokens_used_so_far):
    remaining = TOKEN_LIMITS["total_target"] - tokens_used_so_far
    return remaining >= TOKEN_LIMITS["optional_steps"] + TOKEN_LIMITS["render_budget"] + TOKEN_LIMITS["buffer"]

5.2 Caching Strategy

5.3 Streaming Output Order

If running via API with streaming enabled, generate in this order to allow partial rendering in the UI:

1. Page header + source strip          → renders immediately (user sees activity)
2. Section 1 — Commercial              → highest value, longest section
3. Section 2 — Government              → second priority
4. Section 3 — LinkedIn Spotted        → lighter, faster
5. Section 4 — Live Search Links       → static, no model work required
6. Appendix — Excluded roles           → lowest priority, collapsed by default
7. Evaluation Report block             → last, also collapsed by default

5.4 Batching Strategy

For teams running this for multiple role types simultaneously:

# Instead of: run sequentially (4 × full token budget)
run_search("Senior Angular Developer")
run_search("Senior React Developer")
run_search("Front End Architect")
run_search("UI Engineer")

# Do: share mandatory steps, fork at schema population
shared_results = run_mandatory_steps()  # Steps 1–5 once

angular_roles = filter_and_schema(shared_results, role="angular")
react_roles   = filter_and_schema(shared_results, role="react")
# etc.
# Reduces mandatory step cost by ~75% for multi-role runs

5.5 Cost Monitoring Alerts

Set thresholds and alert if exceeded:

COST_ALERTS = {
    "per_run_warn":   0.30,   # USD — warn if single run exceeds this
    "per_run_stop":   0.50,   # USD — halt optional steps if approaching this
    "per_week_warn":  3.00,   # USD — warn if weekly total exceeds this
    "per_month_warn": 10.00,  # USD — flag for review
}

5.6 Retry & Timeout Policy

FETCH_POLICY = {
    "timeout_seconds": 8,        # Abort fetch after 8s — prevents blocking on slow boards
    "max_retries": 2,            # Retry once on timeout, then mark as blocked
    "retry_delay_seconds": 3,    # Wait before retry
    "blocked_codes": [403, 429], # Do not retry — mark as blocked immediately
    "exclude_codes": [404, 410], # Do not retry — mark as excluded
}

Drift Detection

Prompt and model behaviour can degrade over time without obvious cause. Run this check monthly or after any model version update:

Signals to Watch

Model Version Change Protocol

When Claude model version updates:

1. Run prompt unchanged against golden test dataset (Section 3.3)
2. Compare pass rate vs previous model
3. If pass rate drops >10% → review and update prompt before production use
4. Log model version in every run record

File Structure (Recommended)

angular-job-search/
├── prompts/
│   ├── Angular_Prompt_v4_1_PromptLayer.md      ← What goes to the model
│   └── changelog.md                             ← Prompt version history
├── system/
│   └── Angular_System_Design_Spec_v1.md         ← This file
├── evals/
│   ├── test_dataset_2026-02.json                ← Golden test cases
│   └── run_history.json                         ← All run records
├── cache/
│   └── section4_links.json                      ← Cached static search links
└── outputs/
    ├── Angular_Jobs_2026-02-18.html
    └── Angular_Jobs_2026-02-25.html

Quick Reference — What Lives Where

System Design Spec v1.0 · Pairs with Prompt v4.1 · February 2026

🔹 PRO TIPS FOR EXECUTION

✅ LinkedIn Post Search Timing: Agency posts peak Mon-Wed 8-11 AM UK time — schedule searches then ✅ Agency Name Boost: Add known UK tech agencies to Boolean if targeting specific recruiters: ("Robert Walters" OR "Hays" OR "Michael Page" OR "Understanding Recruitment" OR "Computer Futures") ✅ Clearance Variants: Always include: ("SC Cleared" OR "SC Clearable" OR "SC eligibility" OR "UKSV") ✅ IR35 Filtering: For contract roles, prioritize #OutsideIR35 posts to avoid inside-IR35 contracts ✅ Tool Requirement: This prompt requires AI with live browsing/search (Perplexity, Bing Chat, ChatGPT+Browsing). Standard LLMs cannot access live LinkedIn/job data.

⚖️ Legal Disclaimer & Fair Use Notice

#### What This Tool Is This is a personal, non-commercial AI-assisted job search tool. It instructs an AI model to conduct structured searches across publicly accessible job board URLs, compile results into an organised report, and return direct links to the original source listings — functionally equivalent to a person manually visiting multiple job boards and noting relevant roles in a spreadsheet. #### Data Sources & How They Are Accessed All listings surfaced are drawn from publicly accessible, search-engine-indexed web pages, published by employers and agencies with the express intent of broad public distribution. No authenticated portals, private databases, or ATS systems are accessed. Where a URL returns an access error (403), the tool flags it with ⚠️ and does not attempt to bypass any access control. Content is retrieved via standard HTTP requests, identical to a web browser visiting a page, and is not stored, cached, or redistributed. #### Intellectual Property & Fair Use Job listings remain the intellectual property of the employers, agencies, and platforms that created them. This tool does not reproduce full job descriptions verbatim. Summarised details (title, rate, tech stack) constitute fair dealing for research and private study under the Copyright, Designs and Patents Act 1988, s.29. Every result includes a direct hyperlink to the original source, driving attribution and traffic back to the originating platform. This tool does not divert or disintermediate traffic — it directs users to original listings more efficiently: #### Always Use the Original Source The original job board listing is always the authoritative source. This report is a reference index only.

• Always click through to verify the full job description before applying
• Confirm the role is still active — filled positions are not always removed promptly (roles >14 days old are flagged 🕐)
• Apply only through official channels on the source platform
• Treat salary and rate figures as indicative — actual compensation may differ

#### MIT Limitation of Liability Results are based on publicly available data at the time of search execution. Accuracy, completeness, and link availability cannot be guaranteed. Users rely on this output at their own discretion.

:100: <i>Thanks!</i>

#### Now, don't be an stranger. Let's stay in touch!

<a href="https://github.com/leolanese" target="_blank" rel="noopener noreferrer"> <img src="https://scastiel.dev/api/image/leolanese?dark&removeLink" alt="leolanese’s GitHub image" width="600" height="314" /> </a>

##### :radio_button: Portfolio: <a href="https://www.leolanese.com" target="_blank">www.leolanese.com</a> ##### :radio_button: Linkedin: <a href="https://www.linkedin.com/in/leolanese/" target="_blank">LeoLanese</a> ##### :radio_button: Twitter: <a href="https://twitter.com/LeoLanese" target="_blank">@LeoLanese</a> ##### :radio_button: DEV.to: <a href="https://www.dev.to/leolanese" target="_blank">Tech Blog</a> ##### :radio_button: Questions / Suggestion / Recommendation: [email protected]