---
name: summarise
description: Use when the user types /summarise, or asks for a summary, recap, or "in short" version of what the agent just said or asked. Restates the agent's own immediately preceding turn under the plainly rules: precise, concise, plain English, every claim already evidenced in that turn. When that turn concluded an investigation, the summary carries exactly two things: the diagnosis and the proposed resolution.
---

# summarise

Invoked by the admin agent directly.

The subject is the agent's own last turn, not the codebase, not the conversation as a whole, and not the user's message. Whatever was just said or asked gets restated so a reader who skipped the turn ends up with the same working understanding, in a fraction of the words.

## What success looks like

The summary is precise, concise, plain English, with every claim already evidenced in the turn being summarised. It also satisfies the shipped `plainly` skill (`skills/plainly/SKILL.md`): spoken register, no acronyms, no formulas, no AI tells, no em-dashes. Read `plainly` before writing; do not summarise from memory of what it says.

Three sentences is the ceiling for an ordinary turn. If the turn ended in a question, the summary states the question and what turns on the answer, and nothing else.

## Investigations

When the preceding turn concluded an investigation, the summary has exactly two parts, in this order:

1. **Diagnosis.** What is actually wrong, and the evidence that establishes it. Name the cause, not the layer where the symptom surfaced.
2. **Proposed resolution.** What would fix that cause, and its scope boundary.

Both parts are mandatory. A diagnosis with no proposed resolution is an unfinished summary; a proposed resolution with no diagnosis is speculation wearing a summary's clothes. Neither part introduces a finding, a fix, or a caveat that was absent from the turn being summarised. If the investigation reached a diagnosis but no resolution, say that in the resolution slot rather than inventing one.

## Constraints

- **No new content.** The summary is a compression, never an extension. Anything not in the turn being summarised does not belong here.
- **Drop the scaffolding, keep the load.** File paths, tool names, and step-by-step narration go unless the reader needs them to act. The outcome stays.
- **No hedging.** The original turn either established a claim or it did not. Do not soften a confirmed finding into "appears to", and do not harden a speculation into a fact.
- **No follow-up offers.** End on the summary.

## When NOT to use

- The user wants the summary of a document, a file, or someone else's text: that is ordinary reading work, and the `plainly` skill fits better.
- The last turn was already one or two plain sentences: there is nothing to compress, so say so and stop.