Prompt After Prompt https://promptafterprompt.com/ The practical craft of using AI tools — writing prompts that hold up, building repeatable workflows, and recognising the output you should throw away. en An instruction that only works once was underspecified, not unluckyhttps://promptafterprompt.com/craft/an-instruction-that-only-works-once-was-underspecified-not-unlucky/https://promptafterprompt.com/craft/an-instruction-that-only-works-once-was-underspecified-not-unlucky/Wed, 26 Aug 2026 09:00:00 GMTPrompt CraftAdrian NovakThe gap between a prompt that produced something good and a prompt you can rely on is a list of decisions you left open without noticing. Two worked examples teach a format that a page of rules will nothttps://promptafterprompt.com/craft/two-worked-examples-teach-a-format-that-a-page-of-rules-will-not/https://promptafterprompt.com/craft/two-worked-examples-teach-a-format-that-a-page-of-rules-will-not/Sun, 23 Aug 2026 09:00:00 GMTPrompt CraftDevika MenonRules describe the edges of what you will accept, examples describe the middle of it, and for anything with a shape the middle is what you actually needed to communicate. Saying what you do not want is the half of the instruction people leave outhttps://promptafterprompt.com/craft/saying-what-you-do-not-want-is-the-half-of-the-instruction-people-leave-out/https://promptafterprompt.com/craft/saying-what-you-do-not-want-is-the-half-of-the-instruction-people-leave-out/Fri, 21 Aug 2026 09:00:00 GMTPrompt CraftBhavna DeshpandeDescribing the output you want narrows the field far less than it feels like it does, and the difference is made up by defaults you never asked for. Describe the shape of the answer before you describe the taskhttps://promptafterprompt.com/craft/describe-the-shape-of-the-answer-before-you-describe-the-task/https://promptafterprompt.com/craft/describe-the-shape-of-the-answer-before-you-describe-the-task/Wed, 19 Aug 2026 09:00:00 GMTPrompt CraftNaina SethiOutput structure is the single constraint that decides the most about what comes back, and it is usually the one left to whatever default arrives. Asking the tool to play a role moves the vocabulary more than the judgementhttps://promptafterprompt.com/craft/asking-the-tool-to-play-a-role-moves-the-vocabulary-more-than-the-judgement/https://promptafterprompt.com/craft/asking-the-tool-to-play-a-role-moves-the-vocabulary-more-than-the-judgement/Sat, 15 Aug 2026 09:00:00 GMTPrompt CraftAdrian NovakPersona framing is real and it does something, but what it mostly does is select a register, and mistaking that for added competence is how unverified work acquires the sound of authority. Restarting a draft usually beats the fourth round of correctionshttps://promptafterprompt.com/writing/restarting-a-draft-usually-beats-the-fourth-round-of-corrections/https://promptafterprompt.com/writing/restarting-a-draft-usually-beats-the-fourth-round-of-corrections/Fri, 14 Aug 2026 09:00:00 GMTWriting With AINaina SethiCorrections accumulate as patches on a structure that was wrong to begin with, and there is a point past which every further instruction makes the piece worse in a new way. Handing over the editing is a different bargain from handing over the writinghttps://promptafterprompt.com/writing/handing-over-the-editing-is-a-different-bargain-from-handing-over-the-writing/https://promptafterprompt.com/writing/handing-over-the-editing-is-a-different-bargain-from-handing-over-the-writing/Thu, 13 Aug 2026 09:00:00 GMTWriting With AINaina SethiUsed as a first-draft machine these tools cost you the thinking; used as an editor they cost you almost nothing, and the second arrangement is the one that survives contact with real work. Anything only you know is the part the draft will be missinghttps://promptafterprompt.com/writing/anything-only-you-know-is-the-part-the-draft-will-be-missing/https://promptafterprompt.com/writing/anything-only-you-know-is-the-part-the-draft-will-be-missing/Sun, 09 Aug 2026 09:00:00 GMTWriting With AIMaya ChandrasekarGenerated prose is assembled from what is widely available, so the material that makes a piece worth reading is exactly the material it cannot supply. Generated prose drifts towards the middle of everything it has readhttps://promptafterprompt.com/writing/generated-prose-drifts-towards-the-middle-of-everything-it-has-read/https://promptafterprompt.com/writing/generated-prose-drifts-towards-the-middle-of-everything-it-has-read/Wed, 05 Aug 2026 09:00:00 GMTWriting With AIAdrian NovakThe pull towards the average is not a stylistic flaw to be corrected with better adjectives; it is what the process does, and working with it requires structural choices rather than word choices. The blank page and the last paragraph are not equally good places to askhttps://promptafterprompt.com/writing/the-blank-page-and-the-last-paragraph-are-not-equally-good-places-to-ask/https://promptafterprompt.com/writing/the-blank-page-and-the-last-paragraph-are-not-equally-good-places-to-ask/Tue, 04 Aug 2026 09:00:00 GMTWriting With AIBhavna DeshpandeThese tools are most valuable at the two ends of a writing task and least valuable in the middle, and knowing which stage you are in prevents most of the wasted effort. An image prompt describes a picture, not the scene it depictshttps://promptafterprompt.com/media/an-image-prompt-describes-a-picture-not-the-scene-it-depicts/https://promptafterprompt.com/media/an-image-prompt-describes-a-picture-not-the-scene-it-depicts/Sun, 26 Jul 2026 09:00:00 GMTImages & AudioNaina SethiMost disappointing results come from writing a story and expecting a photograph, when what the tool needs is the visual facts a photographer would have decided on. Changing one thing at a time is the only way to learn what an image prompt doeshttps://promptafterprompt.com/media/changing-one-thing-at-a-time-is-the-only-way-to-learn-what-an-image-prompt-does/https://promptafterprompt.com/media/changing-one-thing-at-a-time-is-the-only-way-to-learn-what-an-image-prompt-does/Thu, 23 Jul 2026 09:00:00 GMTImages & AudioAdrian NovakImage prompts accumulate untested words because everything is edited at once, and a habit that costs a few extra generations turns guesswork into something you can actually build on. Anything that has to be counted should not be asked of an image generatorhttps://promptafterprompt.com/media/anything-that-has-to-be-counted-should-not-be-asked-of-an-image-generator/https://promptafterprompt.com/media/anything-that-has-to-be-counted-should-not-be-asked-of-an-image-generator/Sun, 19 Jul 2026 09:00:00 GMTImages & AudioNaina SethiA handful of failures recur across every tool in this class, they share a cause, and the practical response is to design the task so those failures cannot occur rather than to prompt harder. Describing the light and the lens gets further than naming a stylehttps://promptafterprompt.com/media/describing-the-light-and-the-lens-gets-further-than-naming-a-style/https://promptafterprompt.com/media/describing-the-light-and-the-lens-gets-further-than-naming-a-style/Sun, 12 Jul 2026 09:00:00 GMTImages & AudioMaya ChandrasekarStyle words are the least portable and least predictable part of an image prompt, and the physical properties they are shorthand for can be requested directly. Synthetic narration is punctuated speech and the punctuation is the controlhttps://promptafterprompt.com/media/synthetic-narration-is-punctuated-speech-and-the-punctuation-is-the-control/https://promptafterprompt.com/media/synthetic-narration-is-punctuated-speech-and-the-punctuation-is-the-control/Fri, 10 Jul 2026 09:00:00 GMTImages & AudioNaina SethiGenerated voice follows the structure of the text far more closely than it follows any instruction about tone, which makes rewriting the script the main lever available. One long instruction hides which step failedhttps://promptafterprompt.com/workflows/one-long-instruction-hides-which-step-failed/https://promptafterprompt.com/workflows/one-long-instruction-hides-which-step-failed/Tue, 07 Jul 2026 09:00:00 GMTWorkflowsBhavna DeshpandeA single prompt that extracts, judges and writes gives you no way to tell which of those three went wrong, and the fix is to make each one produce something you can look at. A workflow without a checking step ships its worst outputhttps://promptafterprompt.com/workflows/a-workflow-without-a-checking-step-ships-its-worst-output/https://promptafterprompt.com/workflows/a-workflow-without-a-checking-step-ships-its-worst-output/Fri, 03 Jul 2026 09:00:00 GMTWorkflowsAarav SinhaAny process that produces plausible results at volume will eventually produce a plausible wrong one, and whether that reaches a reader is decided entirely by whether a check exists. Long inputs sag in the middle and the fix is structuralhttps://promptafterprompt.com/workflows/long-inputs-sag-in-the-middle-and-the-fix-is-structural/https://promptafterprompt.com/workflows/long-inputs-sag-in-the-middle-and-the-fix-is-structural/Tue, 30 Jun 2026 09:00:00 GMTWorkflowsBhavna DeshpandeSupplying more material does not straightforwardly mean more of it gets used, and building around that produces better results than trusting a large capacity. A prompt you cannot find again is not part of a workflowhttps://promptafterprompt.com/workflows/a-prompt-you-cannot-find-again-is-not-part-of-a-workflow/https://promptafterprompt.com/workflows/a-prompt-you-cannot-find-again-is-not-part-of-a-workflow/Tue, 23 Jun 2026 09:00:00 GMTWorkflowsBhavna DeshpandeInstructions that live in chat history are irrecoverable in practice, and treating them as artefacts with a location and a version is the cheapest reliability improvement available. The step worth automating is the one whose failure you would noticehttps://promptafterprompt.com/workflows/the-step-worth-automating-is-the-one-whose-failure-you-would-notice/https://promptafterprompt.com/workflows/the-step-worth-automating-is-the-one-whose-failure-you-would-notice/Sun, 21 Jun 2026 09:00:00 GMTWorkflowsAarav SinhaAutomation decisions here turn less on how well the tool performs a task than on whether a wrong result would be visible before it mattered.