context-management

Files (1)

• SKILL.md

SKILL.md (first lines)

# Context Management

Protect the working context for the decisions that remain. Do not spend it on
raw material that will not affect the next action.

## Read With A Question

Before opening a large file, log, diff, document, image set, or search result,
state the question it must answer. Then read only the relevant section, fields,
or examples.

- Prefer a targeted search, range, filter, or screenshot page over reading the
  whole source.
- Do not reread material already understood. Keep a short note with the fact,
  evidence location, uncertainty, and next action.
- Read more only when the current evidence cannot support the decision. Do not
  collect background merely because it is available.

## Keep A Compact Working Record

When a task has several phases, preserve only what the next phase needs:

- the goal and constraints;
- decisions made and why;
- direct evidence and its location;
- open questions, risks, and the next action.

Summarise verbose tool output in those terms. Do not paste a long transcript
into later work unless its exact wording is necessary.

## Delegate Only When It Helps

Delegate a large investigation when it has a clear question, can be explored
independently, and the main task can proceed from a concise evidence-backed
summary.

- Give the delegate a bounded scope and the facts it must return.
- Ask for source locations and uncertainty, not a retelling of every file read.
- Inspect important evidence yourself before making a consequential decision.
- Keep small, tightly coupled reads in the main thread. Do not create agents
  merely to avoid reading a few relevant files.

### Write The Handoff

State the intent and the check for done. The delegate chooses the steps.

- State the goal and a done check that can pass or fail. "Make it better" is
  not a check.
- Name the proof it must show: real command output or a before-and-after
  number.
- Say what is already known, such as a symptom, repro, log, or link, and the
  real constraints, such as "repro first" or "zero behavior change".
- Leave out the how and any hand-ordered list of steps. A fixed order drops or
  reorders steps and blocks a better route.

Before an unattended run, also:

- Write done as checks the delegate can run itself.
- Answer in advance the questions it would stop to ask.
- Ask for a decision log to audit later.
- Give it a way out: what to do or report if it cannot finish.

## Load Instructions Progressively

Keep rules used on nearly every invocation in the active skill. Keep variants,
language details, and rare procedures in directly linked references. Load one
reference when the task needs it; do not load a catalog speculatively.

## Reassess Before Context Becomes A Problem

Pause before a broad new phase, a long fan-out, or a large raw payload. Restate
the current decision and choose the smallest next read or delegated task that
can resolve it. If the context is already crowded, make a compact handoff
record before continuing.

## Repeated Checks

When agents judge the same kind of result more than once, script the check
instead. See `scripted-verification`.

## Sources

- [Pstack guard the context window](https://github.com/cursor/plugins/blob/main/pstack/skills/principle-guard-the-context-window/SKILL.md)
- [Pstack poteto-help prompting](https://github.com/cursor/plugins/blob/main/pstack/skills/poteto-help/references/prompting.md)
- [OpenAI skill authoring guidance](https://github.com/openai/skills/blob/main/skills/.system/skill-creator/SKILL.md)

From AnthonyPoschen/agent-skills, MIT.

More from AnthonyPoschen/agent-skills