Flakback

From natural explanation to visual brief in real time.

One shortcut. Explain it naturally. Get one Flakback.

1. 1

   Capture the intent

   Use one shortcut, then speak, point, circle, and demonstrate in the order that makes sense to you.

2. 2

   Ground every instruction

   Flakback binds each statement to the relevant crop, marker, and moment, preserving what “this” and “there” mean.

3. 3

   Send one complete brief

   Share the evidence, the requested actions, the reasoning and the sequence as one portable Flakback.

It was never about saving tokens.
It's about being understood the first time.

Compressing a screenshot into a card saves roughly 5% of a multi-turn bill. The round trip you never had to take saves about forty times more. Here is the measured version.

• 25.7 → 86.4

  A mark lets a model name a region instead of guessing at it

  In the paper that introduced the technique, overlaying numbered marks on an image took GPT‑4V from 25.7 to 86.4 acc@0.5 on referring-expression comprehension. That is exactly Flakback's method — a numbered pin on the spot you pointed at, at the moment you said it.

  Yang et al., Set-of-Mark Prompting, arXiv:2310.11441. A 177-instance RefCOCOg sample; the marked condition also supplied pre-segmented candidates to choose from.
• −39%

  Explaining across turns costs you a third of the model

  Over 200,000 simulated conversations, 15 models and six generation tasks: the same information delivered turn by turn instead of all at once cost an average 39% of performance. And “when LLMs take a wrong turn in a conversation, they get lost and do not recover.”

  Laban et al., LLMs Get Lost In Multi-Turn Conversation, arXiv:2505.06120.
• 0.325 → 0.150

  The screenshot-and-arrows habit actively backfires

  Draw a pointer on a screenshot, paste it in chat, leave the instruction in the message box. The one study that tested that exact workflow found GPT‑4V did worse than with no pointer at all, and asked for more information in over 78% of cases. The mark and the words have to arrive together, on the same surface.

  VRPTEST, arXiv:2312.04087. GPT‑4V, 2023; not reproduced on current models.
• 2,691 → 361

  And yes — it is also smaller

  A 1920×1080 screenshot costs 2,691 input tokens on Claude Opus 5. A 512×512 Flakback costs 361, already cropped, ordered and labelled. Anthropic prices images by area — ceil(w/28) × ceil(h/28) visual tokens — so every pixel you don't send is a token you don't pay for.

  Computed from each provider's published formula, July 2026. Model-dependent: the same comparison is ~4.6× on standard-resolution Claude, ~1.6× on GPT‑5 and ~1.1× on Gemini 3, which price images closer to flat.

The arithmetic, in full: a Flakback costs about $0.0016 more in input tokens than three sentences of prose. One avoided clarification saves about $0.0690 — roughly 42×. It only has to prevent a clarification 2.4% of the time to pay for itself. Modelled on an 800-token system prompt, 12,000 tokens of context and a 1,500-token answer, at Anthropic list pricing, 2026‑07‑26. Your shape will differ; the direction won't.

One portable Flakback.
Complete working context.

What you showed, drew, said, and did becomes one visual brief. The PNG works anywhere image input is supported. Compatible tools can also use the structured story carried with it — including order, exact language, and spatial references.

• 

  Relevant evidence

  Focused crops retain enough surrounding context to orient the reader without sending an entire screen.

• Grounded instructions

  Matching markers bind every requested action to its target, in language that remains natural and explicit.

• Portable by default

  Generated locally, readable as an ordinary image, and controlled by you from capture through delivery.

Before you buy.