Advanced AI Assistants Mastery: Prompt Chaining Across ChatGPT, Claude 3.7 & Perplexity Pro

Step by step: run the three-assistant chain on one real task

This tutorial’s chain — research with Perplexity, architect with Claude, execute with ChatGPT — only works when each stage receives a clean artifact and a strict contract. Below are the three production prompts, exactly as you would paste them. Replace the bracketed slots; keep the constraints. The constraints are what stop each model from improvising facts the previous stage worked to verify.

Stage 1: Perplexity builds the fact sheet

Research task: [YOUR TOPIC]
Return ONLY:
1. The 10 most important facts, each with a source URL and publication date.
2. Any figures that sources disagree on, flagged as CONTESTED.
3. Primary sources (official docs, filings, announcements) separated from
   secondary commentary.
Exclude undated pages and SEO listicles. Prefer sources from the last
12 months. Number every fact F1, F2, F3... so later stages can cite it.

The numbered facts are the contract. Everything downstream cites F-numbers instead of re-asserting claims, which makes the final draft auditable in one pass.

Stage 2: Claude architects from the fact sheet only

You are the architect stage of a research pipeline. Below is a verified
fact sheet produced by Perplexity.

Using ONLY facts from the sheet, produce:
- An outline with H2/H3 structure.
- For each section: the claim it makes and the fact IDs supporting it.
- A GAPS list: anything the outline needs that the sheet does not cover.

Do not introduce facts that are not in the sheet. Mark every unsupported
idea as GAP instead of filling it in.

[PASTE FACT SHEET]

Stage 3: ChatGPT executes against the outline

You are the execution stage. Turn the approved outline into the final
draft.

Rules:
- Keep every claim tied to its fact ID in brackets, e.g. [F3].
- No new facts, no new numbers, no new examples from memory.
- Write in a direct, technical voice for an expert audience.

Deliverable: the draft, plus a short changelog of the structural choices
you made while writing it.

[PASTE OUTLINE]

Stage 4 (optional): close the loop with a verification pass

For anything that will be published or decided on, add a fourth hop that costs two minutes: feed the finished draft back to Perplexity — the only stage with live sources — and ask it to audit, not to write. This catches the subtle failure the first three stages cannot see: a fact that was true in the sheet but got bent during execution.

Audit the draft below against its own citations.
For every claim with a bracketed fact ID, check the cited source and
report one of: SUPPORTED, DISTORTED (say how), or UNSUPPORTED.
Do not rewrite the draft. Return only the audit table.
[PASTE DRAFT]

Claims marked DISTORTED are almost always compression damage — a qualified statement in the outline ("in most benchmarks") flattened into an absolute in the draft. Fix those in the draft, not in the pipeline: the chain worked, the prose drifted, and the audit trail tells you exactly where.

A complete worked example

Watch the chain handle a real decision: should a small SaaS team adopt MCP servers for its internal tools? Stage 1 returns a fact sheet in this shape — a numbered, dated, sourced artifact, not a chat transcript:

# Fact Sheet - MCP adoption (Perplexity)
- F1: MCP is an open protocol that standardizes how models connect to
  tools and data sources. [official docs, 2026]
- F2: Major model providers ship MCP client support in their coding
  tools. [vendor announcements, 2026]
- F3: Self-hosting an MCP server means your team owns its auth and
  network exposure. [official docs, 2026]
## Contested
- C1: Adoption statistics differ sharply between vendor blogs and
  independent surveys.
## Gaps for stage 2
- No verified pricing for hosted MCP gateways.

Stage 2 turns that sheet into an outline — decision criteria, integration architecture, security review — with every section pinned to fact IDs, and the pricing question parked in GAPS instead of being answered from memory. Stage 3 then writes the recommendation memo with [F1]-[F3] citations intact, ready for a human to verify in minutes. The contested adoption figures (C1) never make it into the memo as facts; they appear, labeled, in an appendix. That is the whole method: each assistant does the job its architecture is best at, and the artifacts between them carry the discipline.

Troubleshooting

• The three assistants contradict each other. Treat the fact sheet as the single source of truth. Send contested items back to Perplexity with a follow-up — "Which of these sources is primary, and what exactly does it say?" — and update the sheet before re-running stages 2 and 3. Never let a later stage silently overrule an earlier fact.
• Claude produced a citation that does not exist. Claude has no browsing in this chain; any URL or statistic it offers from memory is unverified by definition. The stage-2 prompt’s fact-ID rule exists for this failure: reject any claim without an F-number and regenerate, rather than fact-checking prose line by line.
• ChatGPT’s draft drifts in voice and facts over a long thread. The conversation has accumulated stale context. Restart execution from the outline artifact in a fresh chat instead of arguing with the old thread — the chain’s artifacts are versioned files precisely so any stage can be re-run cleanly.
• Every hop makes the output vaguer. A constraint got lost between prompts. Audit each stage against the checklist: role, input artifact, output schema, prohibitions. The stage missing one of the four is the stage that diluted the result.