feat: AI job filter — score jobs 0-10 with Claude Haiku before applying

- lib/filter.mjs: batch scoring engine (10 jobs/call, Claude Haiku) - job_filter.mjs: standalone CLI with --dry-run and --stats flags - Threshold configurable globally + per-search in search_config.json (filter_min_score, default 5) - Job profiles (gtm/ae) passed as context via settings.filter.job_profiles - Filtered jobs get status='filtered' with filter_score + filter_reason - Filter errors pass jobs through (never block applications) - status.mjs: added 'AI filtered' line to report
2026-03-06 10:01:15 +00:00
parent ff59316abc
commit 9bf904dada
4 changed files with 303 additions and 0 deletions
--- a/config/search_config.example.json
+++ b/config/search_config.example.json
@@ -1,10 +1,12 @@
 {
  "_note": "Configure your job searches here. Each search runs on both listed platforms.",
  "first_run_days": 90,
+  "filter_min_score": 5,
  "searches": [
    {
      "name": "Founding GTM",
      "track": "gtm",
+      "filter_min_score": 5,
      "keywords": [
        "founding account executive",
        "first sales hire",
@@ -23,6 +25,7 @@
    {
      "name": "Enterprise AE",
      "track": "ae",
+      "filter_min_score": 5,
      "keywords": [
        "enterprise account executive SaaS remote",
        "senior account executive technical SaaS remote",
--- a/job_filter.mjs
+++ b/job_filter.mjs
@@ -0,0 +1,127 @@
+#!/usr/bin/env node
+import { loadEnv } from './lib/env.mjs';
+loadEnv();
+
+/**
+ * job_filter.mjs — claw-apply AI Job Filter
+ * Scores all queued 'new' jobs 0-10 against candidate profile using Claude Haiku
+ * Jobs below filter_min_score (default 5, configurable per-search in search_config.json)
+ * are marked 'filtered' and skipped by the applier
+ *
+ * Usage:
+ *   node job_filter.mjs           — filter all new jobs
+ *   node job_filter.mjs --dry-run — score without writing status changes
+ *   node job_filter.mjs --stats   — show filter stats only (no re-filter)
+ */
+
+import { dirname, resolve } from 'path';
+import { fileURLToPath } from 'url';
+
+const __dir = dirname(fileURLToPath(import.meta.url));
+
+import { getJobsByStatus, updateJobStatus, loadConfig } from './lib/queue.mjs';
+import { acquireLock } from './lib/lock.mjs';
+import { runFilter } from './lib/filter.mjs';
+
+const isDryRun = process.argv.includes('--dry-run');
+const isStats = process.argv.includes('--stats');
+
+async function showStats() {
+  const all = getJobsByStatus(['new', 'filtered']);
+  const filtered = all.filter(j => j.status === 'filtered');
+  const scored = all.filter(j => j.filter_score != null);
+
+  console.log(`📊 Filter Stats\n`);
+  console.log(`  Filtered (blocked): ${filtered.length}`);
+  console.log(`  New (passed/unscored): ${all.length - filtered.length}`);
+  console.log(`  Total scored: ${scored.length}\n`);
+
+  if (filtered.length > 0) {
+    console.log(`Sample filtered jobs:`);
+    filtered.slice(0, 10).forEach(j => {
+      console.log(`  [${j.filter_score}/10] ${j.title} @ ${j.company} — ${j.filter_reason}`);
+    });
+  }
+}
+
+async function main() {
+  if (isStats) {
+    await showStats();
+    return;
+  }
+
+  const apiKey = process.env.ANTHROPIC_API_KEY;
+  if (!apiKey) {
+    console.error('❌ ANTHROPIC_API_KEY not set — filter requires Anthropic API');
+    process.exit(1);
+  }
+
+  const lock = acquireLock('filter', resolve(__dir, 'data'));
+
+  const settings = loadConfig(resolve(__dir, 'config/settings.json'));
+  const searchConfig = loadConfig(resolve(__dir, 'config/search_config.json'));
+  const candidateProfile = loadConfig(resolve(__dir, 'config/profile.json'));
+
+  const jobs = getJobsByStatus('new');
+  const globalMin = searchConfig.filter_min_score ?? 5;
+
+  console.log(`🔍 claw-apply: AI Job Filter${isDryRun ? ' (DRY RUN)' : ''}\n`);
+  console.log(`  Jobs to score: ${jobs.length}`);
+  console.log(`  Default threshold: ${globalMin}/10\n`);
+
+  if (jobs.length === 0) {
+    console.log('Nothing to filter.');
+    return;
+  }
+
+  let passed = 0, filtered = 0, errors = 0;
+  const filterLog = [];
+
+  const results = await runFilter(jobs, searchConfig, settings, candidateProfile, apiKey, {
+    onProgress: (done, total, track) => {
+      process.stdout.write(`\r  [${track}] ${done}/${total} scored...`);
+    }
+  });
+
+  console.log('\n');
+
+  for (const { job, score, reason, pass, minScore } of results) {
+    if (score === null) {
+      errors++;
+      continue;
+    }
+
+    filterLog.push({ id: job.id, title: job.title, company: job.company, score, reason, pass, minScore });
+
+    if (pass) {
+      passed++;
+      if (!isDryRun) {
+        updateJobStatus(job.id, 'new', { filter_score: score, filter_reason: reason });
+      }
+    } else {
+      filtered++;
+      if (!isDryRun) {
+        updateJobStatus(job.id, 'filtered', { filter_score: score, filter_reason: reason });
+      }
+    }
+  }
+
+  console.log(`✅ Filter complete${isDryRun ? ' (no changes written)' : ''}`);
+  console.log(`  ✅ Passed:   ${passed}`);
+  console.log(`  🚫 Filtered: ${filtered}`);
+  console.log(`  ⚠️  Errors:   ${errors} (passed through)`);
+  console.log(`  📊 Pass rate: ${jobs.length > 0 ? Math.round((passed / jobs.length) * 100) : 0}%\n`);
+
+  if (isDryRun && filterLog.length > 0) {
+    console.log(`Sample scores:`);
+    filterLog.slice(0, 20).forEach(j => {
+      const icon = j.pass ? '✅' : '🚫';
+      console.log(`  ${icon} [${j.score}/10] ${j.title} @ ${j.company} — ${j.reason}`);
+    });
+  }
+}
+
+main().catch(err => {
+  console.error('Fatal:', err.message);
+  process.exit(1);
+});
--- a/lib/filter.mjs
+++ b/lib/filter.mjs
@@ -0,0 +1,171 @@
+/**
+ * filter.mjs — AI job relevance filter
+ * Scores queued jobs 0-10 against candidate profile + job profiles using Claude Haiku
+ * Jobs below filter_min_score are marked 'filtered' and skipped by the applier
+ */
+
+import { readFileSync, existsSync } from 'fs';
+
+const BATCH_SIZE = 10;
+const DESC_MAX_CHARS = 800;
+
+function loadProfile(profilePath) {
+  if (!profilePath || !existsSync(profilePath)) return null;
+  try { return JSON.parse(readFileSync(profilePath, 'utf8')); } catch { return null; }
+}
+
+function buildSystemPrompt(jobProfile, candidateProfile) {
+  const tr = jobProfile.target_role;
+  const exp = jobProfile.experience || {};
+  const highlights = (exp.highlights || []).map(h => `- ${h}`).join('\n');
+
+  return `You are a job relevance scorer. Score each job 0-10 based on how well it matches this candidate.
+
+## Candidate
+- Name: ${candidateProfile.name?.first} ${candidateProfile.name?.last}
+- Location: ${candidateProfile.location?.city}, ${candidateProfile.location?.state} (remote only, will not relocate)
+- Years in sales: ${candidateProfile.years_experience}
+- Desired salary: $${(candidateProfile.desired_salary || 0).toLocaleString()}
+- Background: ${(candidateProfile.cover_letter || '').substring(0, 300)}
+
+## Target Role Criteria
+- Titles: ${(tr.titles || []).join(', ')}
+- Industries: ${(exp.industries || []).join(', ')}
+- Company stage: ${(tr.company_stage || []).join(', ') || 'any'}
+- Company size: ${tr.company_size || 'any'}
+- Salary minimum: $${(tr.salary_min || 0).toLocaleString()}
+- Remote only: ${tr.remote ? 'Yes' : 'No'}
+- Excluded keywords: ${(tr.exclude_keywords || []).join(', ')}
+
+## Experience Highlights
+${highlights}
+
+## Scoring Guide
+10 = Perfect match (exact title, right company stage, right industry, right salary range)
+7-9 = Strong match (right role type, maybe slightly off industry or stage)
+5-6 = Borderline (relevant but some mismatches — wrong industry, wrong seniority, or vague posting)
+3-4 = Weak match (mostly off target but some overlap)
+0-2 = Not relevant (wrong role type, wrong industry, recruiter spam, part-time, etc.)
+
+Penalize heavily for:
+- Part-time roles
+- Wrong industry (insurance, healthcare PR, construction, retail, K-12 education, utilities)
+- Wrong role type (SDR/BDR, customer success, partnerships, marketing, coordinator)
+- Junior or entry-level
+- Staffing agency spam where no real company is named
+- Salary clearly below minimum`;
+}
+
+function buildUserPrompt(jobs) {
+  const jobList = jobs.map((j, i) => {
+    const desc = (j.description || '').substring(0, DESC_MAX_CHARS).replace(/\s+/g, ' ').trim();
+    return `JOB ${i + 1}
+Title: ${j.title}
+Company: ${j.company || 'Unknown'}
+Location: ${j.location || 'Unknown'}
+Description: ${desc}`;
+  }).join('\n\n---\n\n');
+
+  return `Score each of the following ${jobs.length} jobs. Return ONLY a JSON array with one object per job in order:
+[{"score": 7, "reason": "one line explaining score"}, ...]
+
+${jobList}`;
+}
+
+async function filterBatch(jobs, jobProfile, candidateProfile, apiKey) {
+  const res = await fetch('https://api.anthropic.com/v1/messages', {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'x-api-key': apiKey,
+      'anthropic-version': '2023-06-01'
+    },
+    body: JSON.stringify({
+      model: 'claude-3-haiku-20240307',
+      max_tokens: 1024,
+      system: buildSystemPrompt(jobProfile, candidateProfile),
+      messages: [{ role: 'user', content: buildUserPrompt(jobs) }]
+    })
+  });
+
+  if (!res.ok) throw new Error(`Anthropic API error: ${res.status} ${res.statusText}`);
+
+  const data = await res.json();
+  if (data.error) throw new Error(data.error.message);
+
+  const text = data.content[0].text.trim();
+  const clean = text.replace(/```json\n?|\n?```/g, '').trim();
+  return JSON.parse(clean);
+}
+
+/**
+ * runFilter — score all new jobs and return results
+ * @param {Array} jobs - jobs with status 'new'
+ * @param {Object} searchConfig - search_config.json
+ * @param {Object} settings - settings.json (needs settings.filter.job_profiles)
+ * @param {Object} candidateProfile - profile.json
+ * @param {string} apiKey - Anthropic API key
+ * @param {Object} opts - { onProgress }
+ * @returns {Array} [{ job, score, reason, pass, minScore }]
+ */
+export async function runFilter(jobs, searchConfig, settings, candidateProfile, apiKey, { onProgress } = {}) {
+  const globalMin = searchConfig.filter_min_score ?? 5;
+
+  // Group jobs by track
+  const byTrack = {};
+  for (const job of jobs) {
+    const track = job.track || 'ae';
+    if (!byTrack[track]) byTrack[track] = [];
+    byTrack[track].push(job);
+  }
+
+  const results = [];
+
+  for (const [track, trackJobs] of Object.entries(byTrack)) {
+    const searchEntry = (searchConfig.searches || []).find(s => s.track === track);
+    const minScore = searchEntry?.filter_min_score ?? globalMin;
+
+    const profilePath = settings.filter?.job_profiles?.[track];
+    const jobProfile = loadProfile(profilePath);
+
+    if (!jobProfile) {
+      console.warn(`⚠️  No job profile configured for track "${track}" — passing ${trackJobs.length} jobs through unfiltered`);
+      for (const job of trackJobs) {
+        results.push({ job, score: null, reason: 'no_profile', pass: true, minScore });
+      }
+      continue;
+    }
+
+    let done = 0;
+    for (let i = 0; i < trackJobs.length; i += BATCH_SIZE) {
+      const batch = trackJobs.slice(i, i + BATCH_SIZE);
+
+      try {
+        const scores = await filterBatch(batch, jobProfile, candidateProfile, apiKey);
+
+        for (let j = 0; j < batch.length; j++) {
+          const job = batch[j];
+          const result = scores[j] || { score: 5, reason: 'parse_error' };
+          results.push({
+            job,
+            score: result.score,
+            reason: result.reason,
+            pass: result.score >= minScore,
+            minScore
+          });
+        }
+      } catch (err) {
+        console.error(`\n  Filter batch error (track: ${track}, batch ${i}–${i + batch.length}): ${err.message}`);
+        // On error, pass jobs through — don't block applications
+        for (const job of batch) {
+          results.push({ job, score: null, reason: 'filter_error', pass: true, minScore });
+        }
+      }
+
+      done += batch.length;
+      if (onProgress) onProgress(done, trackJobs.length, track);
+    }
+  }
+
+  return results;
+}
--- a/status.mjs
+++ b/status.mjs
@@ -111,6 +111,7 @@ function buildStatus() {
    queue: {
      total: queue.length,
      new: byStatus['new'] || 0,
+      filtered: byStatus['filtered'] || 0,
      applied: byStatus['applied'] || 0,
      failed: byStatus['failed'] || 0,
      needs_answer: byStatus['needs_answer'] || 0,
@@ -209,6 +210,7 @@ function formatReport(s) {
  }

  lines.push(
+    `  🚫 AI filtered:         ${q.filtered || 0}`,
    `  ✅ Applied:             ${q.applied}`,
    `  🔁 Already applied:     ${q.already_applied || 0}`,
    `  💬 Needs your answer:   ${q.needs_answer}`,