To turn tasks into AI prompts in PAL, capture each request as a Task, select a related set from Inbox, and Brew it through a Blend with a {{tasks}} slot. PAL places the eligible Tasks into a numbered block in board order, combines that block with the Blend's instructions, and saves the assembled prompt to Project History. You copy the text into the LLM you use. The model's answers come back to you, so the Tasks board remains the place where you check the work.
Capture enough detail for the next pass
In the authenticated app, press T to open quick-add from any page. The shortcut works when you are outside an input, textarea, select, or other editable field, which keeps it from firing while you type. You can also open quick-add from the navigation.
One save creates one Task and closes the modal. Press T again, or click the quick-add control again, when you have another Task to capture.
The Task has one body field. Give your future self enough to judge the answer. "Rewrite client email" leaves too much to reconstruct. A more useful Task says what is wrong and what should change:
Client email sounds defensive. Keep the deadline and the two factual corrections, but make the reply direct and calm.
That body can pass through {{tasks}} without a second round of explanation. It also gives you a concrete check when the answer returns.
Select a related batch in Inbox
Each Project has its own Tasks board with Inbox, In review, and Done lanes. When you are ready to work, turn on Select in Inbox and check the Tasks that belong in the same prompt. You can choose a few Tasks or use Select all for the visible Inbox result.
The numbered block follows board position, not click order. Drag Tasks into the sequence you want, or use Move up and Move down, before you Brew. Put an item first because it should be answered first, not because you happened to select it first.
When PAL takes the generation snapshot, it keeps only active Tasks still in this Project's Inbox. A Task moved to Done or archived earlier is absent from the prompt, the History link, and the transition to In review.
Before the transition, PAL reads the status of each included Task again. If a Task left Inbox after the snapshot, its text and History link remain because it was part of the assembled prompt. Its current board state wins, so PAL does not move it to In review.
Let the Blend carry the repeated instructions
From the board, Brew selected opens a picker with every active Project Blend and Global Blend available in the current Project. The Tasks badge marks only a direct {{tasks}} slot in that Blend's own text. A parent Blend can inherit the slot from a referenced Blend without showing the badge.
After you choose a Blend, the Generation modal checks its full reference tree. It opens the Tasks picker without a missing-slot warning when it finds {{tasks}} either in the selected Blend or in a referenced Blend. If the token is absent throughout the tree, the modal shows the warning and offers the Insert {{tasks}} slot action.
The slot supplies the numbered work list. The rest of the Blend tells the model how to handle it. For example, a shared editorial-rules Blend can appear before {{tasks}}, while a short output contract after it can require one numbered answer for each numbered Task and a clear note when information is missing. PAL resolves referenced Blends into the final text and then fills the variable slots.
Quick Brew supports the same reserved slot when you need a temporary composition. Add {{tasks}}, combine it with existing Blend references or one-off text, and Brew without first saving a new Blend. Save as Blend remains a separate action if the wrapper proves useful enough to keep.
For the general variable rules behind the reserved slot, read Prompt Variables. For the structure of reusable and referenced Blends, see What Is a Reusable Prompt System.
Brew in PAL, run the prompt elsewhere
Brew assembles the text and writes that text to Project History. The saved record contains the rendered prompt, including the numbered Task block, and links the included Tasks in the same order. The Copy control puts the prompt on your clipboard.
PAL does not send the prompt to a model or receive the model's answer. Paste the copied text into the external LLM you already use. History records what PAL assembled, not the answer that came back from that provider.
PAL also does not read provider token counts or billing data. Its local token estimate uses an approximate character-count formula. Treat it as a rough size signal rather than a provider usage figure.
Keep each batch tied to shared context
Related Tasks can share instructions that would otherwise be repeated. Several content revisions for the same voice guide may fit one prompt because the rules and output contract apply to all of them. A contract review and a code-debugging request probably need different context and are easier to check separately.
Research on batch prompting supports this narrower reason for batching. Cheng, Kasai, and Yu found lower token and time costs in their evaluated settings, while also reporting that batch size and task complexity affected performance in Batch Prompting: Efficient Inference with Large Language Model APIs. Lin and colleagues reported fewer input tokens and model calls in their benchmark experiments, but found that naive or larger batches could lose quality and that item position and order affected answers in BatchPrompt: Accomplish More with Less.
Those papers are evidence about the prompting methods and models they tested. PAL has not reproduced their benchmarks. Putting related Tasks under shared instructions can avoid repeating the same input context, but actual savings and answer quality depend on the model, provider, Task mix, batch size, output length, and how much review or retry the work needs.
Start with a batch whose Tasks genuinely share context. Split it when the instructions become full of exceptions or the answers become hard to match back to individual Tasks.
Review each Task before you close it
After a successful Brew, PAL moves each included Task to In review only if it is still in Inbox when that transition runs. Open the external model's response and check each numbered answer against its Task body.
If the result is correct, mark that Task Done. If it needs another pass, edit the body to record what failed and what the next answer should change, set the Task's Column back to Inbox, and save it. A body edit resets the age shown on the card. Moving a Task, changing its status, archiving it, Brewing it, or returning it to Inbox does not reset that content age.
The Brew panel also offers Return to Inbox for Tasks it just moved, provided they are still in In review. That action changes the board lane but keeps the generation and its Task links in History.
Capture stays quick because the Blend holds repeated instructions and History keeps the assembled prompt. The board still requires an explicit check before each Task reaches Done.