Claude Perplexity skill: 10 search-augmented LLM recipes

The prompt

Use the perplexity skill. Call sonar-pro with search_recency_filter="day" and ask for a 7-bullet brief on the topic 'AI agents shipping this week'. Each bullet must end with a [n] marker tied to the citations[] array the API returns. Save the rendered markdown to ~/notes/daily-ai-agents-{today}.md and print the citation list at the bottom. Skip anything older than 24 hours.

What slides.md looks like

$ curl -s https://api.perplexity.ai/chat/completions \
    -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{"model":"sonar-pro",
         "messages":[{"role":"user","content":"7-bullet brief on AI agents shipping this week, each bullet ends [n] tied to citations"}],
         "search_recency_filter":"day"}' | jq '.choices[0].message.content, .citations'
→ choices[0].message.content: 7 bullets, each with a [1]–[7] marker
→ citations: ["https://...", ...]  (one URL per marker)

Pairs with

The prompt

Use the perplexity skill. For each competitor in @competitors.txt, call sonar-pro with search_recency_filter="month" and a structured prompt: 'pricing changes, new features shipped, hiring signals (open roles), funding events.' Return one row per competitor with explicit None markers when a category is empty. Save to docs/competitive-2026-04.md.

What slides.md looks like

import requests, json
for c in open('competitors.txt').read().splitlines():
    r = requests.post('https://api.perplexity.ai/chat/completions',
      headers={'Authorization': f'Bearer {KEY}'},
      json={'model':'sonar-pro',
            'messages':[{'role':'user','content':
              f'For {c}: pricing changes, new features, hiring, funding — last 30d. Mark None when empty.'}],
            'search_recency_filter':'month'}).json()
    print(c, '|', r['choices'][0]['message']['content'])

Pairs with

The prompt

Use the perplexity skill. Call sonar with model='sonar' (not Pro — fundamentals are stable enough), prompt: 'For ticker {SYMBOL}: current price, last-quarter EPS surprise vs consensus, three material headlines from the past 7 days. Cite each fact.' Add search_domain_filter=["reuters.com","bloomberg.com","sec.gov","wsj.com"] so we only quote primary sources.

What slides.md looks like

$ curl -s https://api.perplexity.ai/chat/completions \
    -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
    -d '{"model":"sonar",
         "messages":[{"role":"user","content":
           "For NVDA: current price, last-quarter EPS surprise, 3 material headlines past 7d. Cite each fact."}],
         "search_domain_filter":["reuters.com","bloomberg.com","sec.gov","wsj.com"],
         "search_recency_filter":"week"}'
→ message.content: price + EPS line + 3 headlines, [1]–[5] markers
→ citations: 5 URLs, all from the four allow-listed domains

Pairs with

The prompt

Use the perplexity skill. Given the claim '{CLAIM}', call sonar-reasoning-pro with reasoning_effort='medium' and ask: 'Verify this claim. Return a JSON object with verdict (CONFIRMED, CONTRADICTED, MIXED, UNVERIFIABLE), confidence (0–1), and an evidence array of {source_url, supports_or_contradicts, quote} entries.' Keep at least three sources before returning a non-UNVERIFIABLE verdict.

What slides.md looks like

import requests, json
claim = "OpenAI cut sonar prices by 50% in March 2026."
r = requests.post('https://api.perplexity.ai/chat/completions',
  headers={'Authorization': f'Bearer {KEY}'},
  json={'model':'sonar-reasoning-pro',
        'reasoning_effort':'medium',
        'messages':[{'role':'user','content':
          f'Verify: {claim!r}. Return JSON: verdict, confidence, evidence[].'}],
        'search_recency_filter':'month'}).json()
print(json.loads(r['choices'][0]['message']['content']))

Pairs with

The prompt

Use the perplexity skill. Call sonar-deep-research with the question 'Write a 1500-word report on the state of agentic browsers in 2026. Cover security, market entrants, real adoption, regulatory gestures.' Save the markdown to reports/agentic-browsers-2026.md and append the citations[] array as a bibliography. Verify every [n] marker in the body resolves to a URL.

What slides.md looks like

$ curl -s https://api.perplexity.ai/chat/completions \
    -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
    -d '{"model":"sonar-deep-research",
         "messages":[{"role":"user","content":
           "1500-word report: state of agentic browsers 2026. Security, entrants, adoption, regulation."}]}'
→ choices[0].message.content: 1500w with [1]…[18] markers
→ citations: 18 URLs (book + 6 vendor blogs + 11 news pieces)
# cost: ~$2 per deep-research call (token-heavy + per-search fees)

Pairs with

The prompt

Use the perplexity skill. As I write, when I type a sentence ending with `??`, replace the `??` with the closest source URL the sonar model returns. Use model='sonar' (low latency), search_recency_filter='year', and only inject when confidence is high. Keep the original sentence intact; append the citation in markdown footnote syntax [^n].

What slides.md looks like

# Streaming hook around your editor (Neovim, Cursor, etc.)
# Trigger on save: scan the buffer for sentences ending '??'
for sent in re.findall(r'[^.]+\?\?', buffer):
    r = requests.post('https://api.perplexity.ai/chat/completions',
      headers={'Authorization': f'Bearer {KEY}'},
      json={'model':'sonar',
            'messages':[{'role':'user','content':
              f'Closest authoritative source URL for this claim: {sent!r}'}]})
    cite = r.json()['citations'][0]
    buffer = buffer.replace(sent, sent.replace('??', f'[^{i}]') + f'\n[^{i}]: {cite}')

Pairs with

The prompt

Use the perplexity skill. For each row in queue.csv, call sonar with: 'Is the following statement factually consistent with current public sources? Return one of {LIKELY_TRUE, LIKELY_FALSE, UNVERIFIABLE} and a confidence 0–1.' Append the verdict and the top citation URL back to the CSV. Rate-limit at 3 req/s (Sonar's free-tier ceiling).

What slides.md looks like

import csv, requests, time
with open('queue.csv') as f, open('queue.out.csv','w') as g:
    rows = list(csv.DictReader(f))
    for r in rows:
        resp = requests.post('https://api.perplexity.ai/chat/completions',
          headers={'Authorization': f'Bearer {KEY}'},
          json={'model':'sonar',
                'messages':[{'role':'user','content':
                  f'Factually consistent? {r["text"]!r} → {{verdict, confidence}}'}]}).json()
        r['verdict'] = resp['choices'][0]['message']['content']
        r['top_cite'] = (resp.get('citations') or [None])[0]
        time.sleep(0.34)  # ~3 req/s
    csv.DictWriter(g, fieldnames=rows[0].keys()).writerows(rows)

Pairs with

The prompt

Use the perplexity skill. Build a `pq` CLI: takes a single question argument, calls sonar-pro with stream=true, prints content tokens to stdout as they arrive, then prints the citations[] array under a 'Sources:' header. Handle Ctrl-C cleanly. Save as ~/bin/pq and chmod +x.

What slides.md looks like

#!/usr/bin/env python3
import sys, json, requests
q = ' '.join(sys.argv[1:])
r = requests.post('https://api.perplexity.ai/chat/completions',
  headers={'Authorization': f'Bearer {KEY}'},
  json={'model':'sonar-pro','stream':True,
        'messages':[{'role':'user','content':q}]}, stream=True)
cites = []
for line in r.iter_lines():
    if not line.startswith(b'data: '): continue
    chunk = json.loads(line[6:])
    sys.stdout.write(chunk['choices'][0]['delta'].get('content',''))
    sys.stdout.flush()
    cites = chunk.get('citations') or cites
print('\n\nSources:'); [print(f'  - {c}') for c in cites]

Pairs with

The prompt

Use the perplexity skill. Call sonar-pro with search_domain_filter=["*.gov","*.edu","who.int"] for medical questions, or ["*.gov","law.cornell.edu","justia.com","sec.gov"] for legal. If the resulting citations[] array has fewer than 3 entries, retry once without the filter and tag the answer LOW_AUTHORITY in the output.

What slides.md looks like

$ curl -s https://api.perplexity.ai/chat/completions \
    -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
    -d '{"model":"sonar-pro",
         "messages":[{"role":"user","content":
           "Current FDA stance on at-home semaglutide compounding. Cite primary sources."}],
         "search_domain_filter":["*.gov","*.edu","who.int"]}'
→ citations: 4 URLs, all on fda.gov / nih.gov / cdc.gov
→ message.content: synthesised answer, [1]–[4] markers tied to citations

Pairs with

The prompt

Use the perplexity skill. Step 1: call sonar with the broad question, return the top 5 themes. Step 2: pick the highest-confidence theme, call sonar-pro with a focused follow-up that quotes the theme verbatim. Step 3: call sonar-reasoning-pro with the focused answer plus 'Identify any factual claims and verify each independently. Flag disagreements.' Save the three transcripts as steps-1.md, 2.md, 3.md.

What slides.md looks like

# Step 1 — broad scan (model: sonar)
themes = ask('sonar', 'Top 5 themes in agentic browser security 2026.')
# Step 2 — focused follow-up on theme[0] (model: sonar-pro)
deep = ask('sonar-pro', f'Focused on theme {themes[0]!r}: who has shipped, what broke?')
# Step 3 — verification pass (model: sonar-reasoning-pro, reasoning_effort=high)
verify = ask('sonar-reasoning-pro', f'Identify and verify each factual claim: {deep!r}',
             reasoning_effort='high')
open('steps-3.md','w').write(verify)

Pairs with