pattern playground

Standing instructions

Added

… the actor works with an agent or assistant across many sessions. Its default behaviour is close but not right, and the same correction comes back each time: the spelling, the length, the tool it should reach for, the thing it must never do. Saying it in the prompt fixes one run; saying it as an instructed revision fixes one output. By the third time the actor wants to say it once and have it hold.

Problem

The system does most things well enough, and a few things wrong in the same way every time. Each time, the actor corrects it. The correction works, and the next session starts from the same defaults, so the actor makes it again. Nobody notices the cost of one correction; the cost is that it never stops.

The defaults cannot be right for everyone, and the preferences that differ are mostly not the kind a settings screen anticipates. An actor who wants British spelling, no headings, one tool rather than another, or a house rule about how tests are run has no axis to set; they have a sentence. Writing that sentence once is cheap. What is not cheap is knowing it needs writing: people often cannot name a preference until they have corrected the same output three times, and a correction made in the flow of work does not announce itself as a candidate for keeping. Where the system writes its own account of the actor, the actor’s handle on it is partial and the words are not theirs; where the actor writes, the words are theirs and the system’s reading of them is what varies.

Solution

Give the actor a place to write, in plain language, what should hold across sessions, and read it at the start of every run. Keep it as the text they wrote, where they can find and change it, rather than as an inference about them. Scope it by where it lives, so a rule for one project or one kind of file does not reach the rest. Make it easy to put a correction there from the place the correction was made.

The instruction is read by a system that interprets, so it holds in the way a note to a colleague holds, not the way a setting holds. Say what it is for as well as what to do, keep each entry short, and do not let it stand in for a rule that has to be exact.

Continuation

What should hold is written down in the actor's words and read at the start of every run, so a correction made once stops recurring, and every piece of generated content from then on carries its mark. Rule composition is the other form a standing correction takes: a condition built from parts the system offers, exact enough to run on its own, where an instruction is prose the system interprets. Selective memory is the neighbouring channel, written by the system rather than the actor; the two need separate controls, or a fact the actor erased survives as behaviour they once asked for. Settings hold the preferences that fit an enumerated axis, and an instruction holds what no axis anticipated. A workflow or a skill is a procedure the actor invokes; an instruction is read whether or not anything invokes it. The instruction accumulates and goes quiet: behaviour set months ago reads later as how the system is, and a file that grew by one line per correction starts to contradict itself with nothing to say which line wins. Reading it does not say what it will do, which is the gap a dry run fills where the effect can be rehearsed. How far the system may act on its own is not an instruction but a grant, and recurring renewals of the grant become standing bounds the way recurring corrections become standing instructions; that is bounded autonomy's side of the same escalation…

All connections
Complements
Selective memory — Selective memory is the neighbouring channel, written by the system rather than the actor; the two need separate controls, or a fact the actor erased survives as behaviour they once asked for.
Settings — Settings hold the preferences that fit an enumerated axis, and an instruction holds what no axis anticipated.
Workflow — A workflow or a skill is a procedure the actor invokes; an instruction is read whether or not anything invokes it.
Alternatives
Rule composition — Rule composition is the other form a standing correction takes: a condition built from parts the system offers, exact enough to run on its own, where an instruction is prose the system interprets.
Related to
Generated content — What should hold is written down in the actor's words and read at the start of every run, so a correction made once stops recurring, and every piece of generated content from then on carries its mark.
Prose
Dry run — Reading it does not say what it will do, which is the gap a dry run fills where the effect can be rehearsed.
Bounded autonomy — How far the system may act on its own is not an instruction but a grant, and recurring renewals of the grant become standing bounds the way recurring corrections become standing instructions; that is bounded autonomy's side of the same escalation
Preceded by
Prompt — Saying it in the prompt fixes one run; saying it as an instructed revision fixes one output.
Instructed revision — the correction that keeps recurring is the one worth making standing
Hosted by
Workspace — “Put the defaults a workspace hands down where the actor can find and change them: its configuration in settings, the shapes new work starts from as templates, and, where a model works inside the workspace, the standing instructions and reference material that every session in it starts with.”
Appears in
Delegating work: Correcting the result, step 2

In practice

Promoting correction to an instruction

  • The actor promotes it by command from the place they made it: a slash command, a phrase (“add this to the instructions”), or a tool that appends what they said to the file.
  • The system saves on its own and shows a receipt (“Saved 2 memories”), which is memory’s territory unless the actor can read and edit what was saved as text.
  • The system drafts an entry and the actor approves, edits or dismisses it before it takes effect

Research on this pattern

Vaithilingam et al. call the artefact an intent specification (“cursorrules, CLAUDE.md files, user directives, AI memory of user intent”), a human-authored document that grounds the system’s decisions, and show that as it grows new entries conflict with old ones and need something like impact analysis before they are committed. Neumann, Pi and Singh find that the actor’s instructions sit at the bottom of a hierarchy they cannot inspect, and that people “consistently desire greater transparency and control” over it. Li et al.’s walkthrough of six assistants finds the same verb (delete, forget) producing local or global effects depending on which layer it reaches, and model customisation records complicating control further. Wang et al. compare labelling examples, writing rules and writing prompts as ways to teach a personal classifier: prompting performs best at first, all three stall at refinement, and people blend them, which is the evidence that instruction, composed rule and demonstration are three moves rather than one. Gao et al. show the system inferring a preference description from repeated edits and showing it for correction, the system-written counterpart of the actor’s text.

Resources & references