Lesson 27 of 32

The Unbroken Chain

Prompt Chaining

Instead of asking for the whole job in one prompt, break it into a short sequence - research, outline, draft, polish, translate - where each step's output becomes the next step's input, and you inspect it before passing it on. It works because a small step is something you can actually judge, and a mistake caught in the research never becomes a mistake printed in three languages.

Reach for it whenWhen the thing you want is too big to judge in a single answer - a plan, a script, a leaflet - and especially when a later step will translate, print or publish whatever an earlier step got wrong.

State the whole chain in your first message, then order one link and nothing more - "four steps; reply with step one only and stop". That instruction matters more here than on the other platforms, because the eagerness to be helpful shows up as running ahead and handing you the finished business plan when you asked for a list of five pickles. Run the drafting and polishing links in Canvas, where you can correct the text yourself at the checkpoint rather than describing the correction in prose. And when one link comes out wrong, edit that message and continue from there - the conversation branches at the bad link instead of carrying it forward.

From the book Run the drafting and polishing links in Canvas, where you can correct the text yourself at each checkpoint before handing it on. If one link goes wrong, edit that message and continue from there rather than starting the whole chain again.
Worth knowingTwo things to watch. First, it will often ignore "stop and wait" on the first attempt; if it delivers the whole chain anyway, do not try to use the parts - say "I asked for step one only" and start again, because a draft built on unapproved research is exactly the compounding error the chapter warns about. Second, saved memory can carry assumptions between chains: a decision you made for last month's leaflet may quietly shape this month's. If a link produces something you did not ask for and cannot explain, check what it remembers about you.
  • Say "step one only, then stop" - and repeat it if it runs ahead.
  • Keep drafting and polishing links in Canvas, not in chat.
  • Edit the bad link's message to branch instead of restarting.
  • Never use output from a step that skipped your checkpoint.
  • Save the finished sequence as a Custom GPT with blanks.
The Monster Prompt vs. The Approved Links

Instead of

Write a business plan for our SHG pickle business.

ChatGPT version

We are a self-help group of 12 women near Warangal, planning a small pickle business.

## The chain we will run
1. Shortlist the products.
2. Outline the one-page plan.
3. Write the plan.
4. Translate it into Telugu.

## Step 1, and nothing further
List 5 pickles we could make from produce grown around us. For each one give me:

| Pickle | Rough raw material cost per kg (Rs) | Shelf life without a cold chain | How easy to sell locally (1-5, with one line why) |

Use the cheapest reasonable local sourcing, and say in a line underneath which costs you are least sure about.

**Stop there.** Do not outline anything, do not write a plan, do not mention Telugu. I will approve three of the five and only then order step 2. If you reply with more than this table, you have made the whole chain unusable to me.
Open ChatGPT 810 characters

Then loop it

  1. Approved: mango, gongura and lemon. Using only those three rows of your table, outline the one-page plan - products, costs, who does what, where we sell, first-month target. An outline only: headings and one line under each. Stop again.
  2. Good, but move "where we sell" to the top; nothing else changes. Using that corrected outline, now write the full one-page plan in simple English, and open it in Canvas so I can fix the wording myself before we translate.
  3. The Canvas version is final. Now translate exactly that text into simple Telugu our members can read aloud at the meeting. Keep the headings in the same order, and leave the rupee figures as numerals.

Why it worksEach link has one job and a checkpoint, and every handoff names exactly which output to build on, so mistakes are caught long before the final Telugu version.

The Vague "Now Write It" vs. The Named Baton

Instead of

Write a YouTube video about Gulmarg in winter.

ChatGPT version

I make Hindi-English travel videos for budget travellers. Next up: a 10-minute guide to a winter day trip from Srinagar to Gulmarg.

## The chain
Research, then outline, then script, then title and description. Four links. We are on link 1.

## Link 1 only
Give me 12 facts and tips a viewer would actually want, covering: how to get there and back in a day, rough costs in Rs, what to wear at that altitude in winter, gondola booking, and the mistakes first-timers make.

Format each as one line. Put **[check]** at the end of any line whose fact could have changed this season - fares, timings, road conditions, booking rules - because I will verify those locally before they go anywhere near a script.

Do not write an outline. Do not write a script. Do not suggest titles. Reply with the 12 lines and stop.
Open ChatGPT 810 characters

Then loop it

  1. I have been through the [check] items locally; 4 and 11 were wrong and I have dropped them. Using facts 1, 3, 5, 7 to 10 and 12 only, build a scene-by-scene outline for 10 minutes with timings. Outline alone - stop before the script.
  2. The outline works. Using it, write the full script in casual Hinglish with B-roll notes in brackets, and put it in Canvas - I will tighten the spoken lines myself before anything is recorded.
  3. That script in the Canvas is locked. Based on it, give me 5 title options under 60 characters and a 150-word description with chapter timestamps that match the outline's timings.

Why it worksEvery handoff names its baton - which facts, which outline, which script - and the [check] marks create a checkpoint before anything reaches the camera.

The Unchecked Chain vs. The Baton Summary

Instead of

Make a dengue awareness leaflet in Marathi.

ChatGPT version

NGO volunteer, Pune. We run monsoon awareness sessions in housing colonies and need a one-page leaflet for residents.

## The contract for this chain
| Link | Output | Who approves it |
|---|---|---|
| 1 | Eight key messages | Our local health centre |
| 2 | English leaflet | Me, in Canvas |
| 3 | Review pass | Me |
| 4 | Marathi version | A native-speaking volunteer |

Nothing moves to the next row until the approver in that row has signed off. Treat a missing approval as a reason to stop and ask, not a reason to continue.

## Link 1, and nothing else
Eight things a resident can do this week to stop mosquitoes breeding around their home. Plain words, one line each, and in brackets the authority behind it - municipal, national programme, WHO - so the health centre can scan the list in five minutes.

Prevention of breeding only; the single clinical sentence this leaflet will ever carry is "see a doctor if fever continues".

Then stop. These eight go to the health centre first, because an error here is an error printed on every door in the colony.
Open ChatGPT 1,061 characters

Then loop it

  1. Health centre says tips 1 to 6 are fine and has added "empty flower-pot trays every week". That is row one approved. From those 7, write the leaflet in Canvas: headline, 7 short points, closing line with a space for [helpline number].
  2. Row two approved - I have edited two lines in the Canvas myself. Before the Marathi row, write me a baton summary of this chain: aim, audience, the 7 tips as they now stand, the scope rule, the tone. Short enough to paste at the top of a fresh chat.
  3. (New chat.) [Paste the baton summary, then the Canvas leaflet text.] Render this in simple spoken Marathi for colony residents - same headings, same order, [helpline number] left as it is. Our volunteer approves this row, not you, so flag anything you are unsure of rather than smoothing it over.

Why it worksThe checkpoint lets a professional verify the facts before they flow into every later link, and the baton summary carries the chain cleanly into a fresh chat.

Loop it

Treat every link as its own small loop. Read each output before you pass it on: are the facts right, is anything missing, is the order sensible? Fix it there - "Change point 3, then show me the corrected outline" - and only then hand off with a named baton: "Using the corrected outline above, now…". If a later link goes wrong, return to that link alone instead of restarting the chain. When the conversation grows long, ask for a baton summary and continue in a new chat. Save the finished sequence as a chain recipe.