Cumbersome Annotations lets you highlight exact words in any completed user or AI message, attach an optional note, and carry both into your next request. The original message stays intact. You can point the model at the passage you mean without copying it into another prompt.
I built this for the common case where an AI response is close, but one sentence needs a specific kind of pressure. Maybe the tone is too warm. Maybe a claim needs evidence. Maybe one paragraph misses an important constraint. A vague follow-up like "fix the second part" makes the model guess. An annotation removes that guess.
A one-word revision request
The screenshots use a deliberately small example. The AI writes a three-sentence thank-you note, including the line, "I truly appreciate the time and effort you put in." I select the overgenerous part and add this note:
They barely helped. So be a touch sarcastic.
The next message says only, "Revise." Cumbersome includes the selected passage and annotation as additional context, so the AI knows exactly what to change. The new version ends with, "Every little bit counts." Much better.
This gets more useful as conversations grow. A note stays attached to its source instead of becoming another disconnected instruction ten messages later.
The context stays visible
When you save an annotation, Cumbersome highlights the passage in International Orange and shows the note directly beneath the original message. A compact chip above the composer tells you which new or edited annotations will go with your next message.
After you send that message, the chip disappears and the user turn gets a small backlink. Tap it to jump straight to the original passage and its current annotation. You can edit or delete the note at the source. Editing it queues the new version for the next message. Deleting it removes the annotation and its timeline links.
Annotations work on completed prose in both user and AI messages. The first version excludes code, tables, formulas, media, attachments, tool output, and messages that are still streaming. A selection also can't cross messages or overlap another annotation.
Why annotations belong in an AI conversation
I already treat Cumbersome conversations as working documents, not immutable chat logs. Sometimes editing the original text is the right move. Other times I want to preserve what the model wrote and attach direction to one precise passage. Annotations give me that second option without duplicating text or muddying the transcript.
The feature uses the same conversation-history path as ordinary chat, Ghost/Write, conversation compaction, exports, imports, and duplicated conversations. Cumbersome doesn't send annotation text, selected passages, message identities, or annotation IDs to analytics.
Annotations are also built into Endemic, our on-device AI app. Cumbersome sends the extra context to whichever API model you picked. Endemic sends it to the local model running on your device. The interaction is the same on iPhone, iPad, and Mac.
If you use Cumbersome, highlight a sentence in any finished message and choose Add Annotation. Then tell the AI what that sentence needs.
Bless up! 🙏✨