The actor is looking at the thing they want to change, and the way to change it is somewhere else: a form, a dialog, another page. Getting there means translating what they see into someone else’s arrangement of the same information, making the change without its surroundings in view, and coming back to find out whether it landed the way they meant.

Forces

• Context vs. room — changing something where it sits keeps its surroundings visible: the neighbouring values, the layout it has to fit, the comparison that prompted the change. A separate surface has room for labels, help, and grouping, and no way to show any of that context.
• Reading vs. changing — the same surface now does two jobs. An affordance obvious enough to invite editing is noise for whoever came to read; one quiet enough to stay out of the way is missed by whoever came to change something.
• Clean surface vs. reach — revealing controls on approach keeps the reading surface uncluttered and hands the feature to pointer users. Touch has no hover, and keyboard access has to be built deliberately rather than inherited.
• Fluency vs. stray input — when changing is always one gesture away, so is changing by accident.

Solution

Let the change happen on the rendered artefact: the thing that displays a value is the thing that accepts a new one. Where the display cannot host the control, or the change only makes sense as part of a set, route it to a separate surface.

How the affordance appears

Always live. The surface is permanently editable and nothing marks the difference between reading and changing. Cheapest to build and lowest friction to use, with the most exposure to accidental change. Fits surfaces whose main use is writing.

On activation. The display becomes a control when the actor clicks, taps, or focuses it. Keeps the reading surface clean and puts one deliberate act in front of every change.

On approach. The affordance appears on hover or focus. A workable compromise for pointer input that needs an explicit equivalent for touch and keyboard, or it does not exist for those actors at all.

Granularity

What is editable in place — a single value (a cell, an attribute in a detail view), a fragment (a paragraph, a block), or the whole artefact — decides both what a change costs to make and how much of the record a stray gesture can reach. It also decides where feedback fits: with the display doubling as the control, there is no labelled field to hang a validation message under.

When the change belongs elsewhere

• The input needs room the display has not got: help text, a legend, a picker that would cover the thing it is meant to change.
• The fields only make sense together, and applying half of them would leave the record in a state nobody asked for.
• The change applies across a set — “every row where…” — which is a description of a collection, not a thing being looked at. Describing is a symbolic surface’s job (a query, a script, a form over the set); sets and repetition are the canonical limit of direct manipulation.
• The actor needs to review the changes as a group before committing. A review step needs pending state, and editing in place does not produce any.
• The display form cannot host a control: a map pin, a mark on a chart, a compressed row with nowhere to put a date picker.
• The change is destructive or expensive to recover, where action consequences asks for a step rather than one fewer.

Resources & references

• Ventura (2019) Advancing the block interface — the Gutenberg team on the tension between making editing affordances clear and not distracting people who are writing
• Hutchins, Hollan & Norman (1985) Direct manipulation interfaces — names the two arrangements this pattern contrasts: a model world acted on directly versus a conversation with an intermediary about a hidden one; also the limits — sets, repetition, and precision are where description beats manipulation
• Beaudouin-Lafon (2000) Instrumental interaction — makes the in-place/form choice measurable as spatial and temporal offset between control and object: dialogs sit high on both, selection handles low, inspectors in between; the short temporal offset of committing piecemeal is what produces the sense of causality
• Sellen, Kurtenbach & Buxton (1992) The prevention of mode errors through sensory feedback — accidental edits as mode errors; continuous kinesthetic or visual feedback about which mode a surface is in reduces them
• Gaver (1991) Technology affordances — the appearance modes as an affordance taxonomy: perceptible, hidden (always-live with nothing marking editability, on-approach before the hover), and false
• Masson et al. (2024) DirectGPT — editing an LLM’s output in place: participants were 50% faster with 72% shorter prompts than steering the same edits through the conversation
• Editing with AI: How Doctors Refine LLM-Generated Answers to Patient Queries (2025) — doctors refining LLM drafts: instructing changes cost less effort but introduced errors, editing directly was precise but more work

To-do

• The read-heavy case needs somewhere to go: a surface actors rely on rather than write to wants something other than always-live editing. The mode-error literature points at a reintroduced read/edit boundary with continuous feedback about which side the actor is on, rather than restriction or a saved view.
• The inspector/property panel sits between this move and form — it stays open, acts on the current selection, and applies immediately — and currently has no home in the catalogue.