Weak Versus Strong Prompts · ai-skills

Asking Clearly · Beginner to Intermediate

On this page

What makes a prompt weak

Weak prompts omit the artifact, the question, or the shape. The model must guess at least one of them.

Strong prompts name all three plus one bound. Guessing has nowhere left to hide.

Missing piece	Symptom	Fix
Artifact	Answer about the wrong block.	Name the module and version.
Question	A tour instead of an answer.	Name the decision: function, bug class, or grouping.
Shape	Prose that cannot be checked.	Name keys, tables, or line counts.

Explaining an RTL block

Weak form names no block and no question. Strong form names the block, the keys, and the length bound.

Form	Prompt text
Weak	`Explain this RTL.`
Strong	`Explain fifo.sv rev 12 in FUNCTION, DEPTH, SIGNALS keys. Context: 8-deep, active-low reset. Stay under 120 words.`

Reviewing a testbench

Weak form names no file and no bug class. Strong form names the file, the focus, and the output table.

Form	Prompt text
Weak	`Review my testbench.`
Strong	`Review tb_fifo.sv for reset coverage only. List each register missing an async-assert sync-deassert check as DUT \| REGISTER \| MISSING CHECK.`

Summarising a log

Weak form leaves grouping and fields open. Strong form fixes the grouping key and the columns.

Replies below are schematic illustrations only, never output from a specific system. They show the shape a strong prompt requests.

plaintext

schematic illustration only, not a real model reply
weak shape: one paragraph of prose with mixed codes and times
strong shape, three fixed columns:
ERR_TIMEOUT | first 09:14:02 | count 17
ERR_OVERFLOW | first 09:15:44 | count 3
ERR_CRC | first 09:18:10 | count 41
total errors = 17 + 3 + 41 = 61
the strong shape groups by code and carries counts a script can sum

plaintext

assume 1 million input tokens cost 3 units (stated example parameter)
weak prompt uses 50 input tokens, strong prompt uses 200 input tokens
batch of 1000 calls
weak total = 1000 x 50 = 50000 tokens
weak cost = 50000 / 1000000 x 3 = 0.05 x 3 = 0.15 units
strong total = 1000 x 200 = 200000 tokens
strong cost = 200000 / 1000000 x 3 = 0.2 x 3 = 0.60 units
extra cost of the strong form = 0.60 - 0.15 = 0.45 units for the batch

Check yourself A strong prompt still returns prose instead of the requested table. What is the cheapest next fix?

Add one example row in the requested shape. One example fixes key names better than repeating the instruction.

Check yourself In the costed calc, what is the extra batch cost of the strong form over the weak form?

0.45 units for 1000 calls. Strong costs 0.60 units and weak costs 0.15 units under the stated example price.

Related topics

Up nextPrompt Template Builder

Practice this lesson Report an error

What makes a prompt weak

Weak prompts omit the artifact, the question, or the shape. The model must guess at least one of them.

Strong prompts name all three plus one bound. Guessing has nowhere left to hide.

Missing piece	Symptom	Fix
Artifact	Answer about the wrong block.	Name the module and version.
Question	A tour instead of an answer.	Name the decision: function, bug class, or grouping.
Shape	Prose that cannot be checked.	Name keys, tables, or line counts.

Form

Prompt text

Weak

Explain this RTL.

Strong

Explain fifo.sv rev 12 in FUNCTION, DEPTH, SIGNALS keys. Context: 8-deep, active-low reset. Stay under 120 words.

Form

Prompt text

Weak

Review my testbench.

Strong

Review tb_fifo.sv for reset coverage only. List each register missing an async-assert sync-deassert check as DUT | REGISTER | MISSING CHECK.

Summarising a log

Weak form leaves grouping and fields open. Strong form fixes the grouping key and the columns.

Replies below are schematic illustrations only, never output from a specific system. They show the shape a strong prompt requests.

plaintext

schematic illustration only, not a real model reply
weak shape: one paragraph of prose with mixed codes and times
strong shape, three fixed columns:
ERR_TIMEOUT | first 09:14:02 | count 17
ERR_OVERFLOW | first 09:15:44 | count 3
ERR_CRC | first 09:18:10 | count 41
total errors = 17 + 3 + 41 = 61
the strong shape groups by code and carries counts a script can sum

plaintext

assume 1 million input tokens cost 3 units (stated example parameter)
weak prompt uses 50 input tokens, strong prompt uses 200 input tokens
batch of 1000 calls
weak total = 1000 x 50 = 50000 tokens
weak cost = 50000 / 1000000 x 3 = 0.05 x 3 = 0.15 units
strong total = 1000 x 200 = 200000 tokens
strong cost = 200000 / 1000000 x 3 = 0.2 x 3 = 0.60 units
extra cost of the strong form = 0.60 - 0.15 = 0.45 units for the batch

Check yourself A strong prompt still returns prose instead of the requested table. What is the cheapest next fix?

Add one example row in the requested shape. One example fixes key names better than repeating the instruction.

Check yourself In the costed calc, what is the extra batch cost of the strong form over the weak form?

0.45 units for 1000 calls. Strong costs 0.60 units and weak costs 0.15 units under the stated example price.