project-plugin/skills/project-refocus/SKILL.md
Refresh the plan to focus on the task at hand. Use when context grew, completed steps muddy it, or you want to clear context and continue in auto mode.
npx skillsauth add laurigates/claude-plugins refocusInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Refresh the plan so it focuses only on the task that remains, then surface it for approval so the user can clear context and continue in auto mode from a clean slate.
| Use this skill when... | Use the alternative when... |
|------------------------|-----------------------------|
| Context has grown and the completed early steps now muddy it | Resuming after a break with no live context to trim → /project:continue |
| You want a tightened, forward-only plan before continuing | Capturing session learnings into rules/recipes → /project:distill |
| You want to clear context and continue in auto mode without losing the thread | Entering an unfamiliar codebase needing orientation → /project:discovery |
| The user says "refresh the plan", "refocus", "let's focus on what's left", "trim the context" | Picking the next blueprint action → /blueprint:execute |
The output is deliberately self-contained: it must survive a context clear. Approving it from the ExitPlanMode dialog is where Claude Code offers "clear context and continue in auto mode."
git branch --show-currentgit status --porcelain=v2 --branchgit log --format='%h %s' --max-count=8Parse $ARGUMENTS as an optional free-text focus directive that biases what
counts as Remaining vs Stale and what the Objective emphasizes (e.g. "focus on
the API layer, ignore the docs work").
Execute this plan-refresh workflow. Reason over the live conversation, not just git — git grounds what landed on disk; the conversation holds the decisions and the remaining intent.
Scan the session so far and sort everything into three buckets:
Ground "Done" against git log / git status from Context so you don't trust a claim that never reached disk.
If a focus directive was given (see Parameters): bias the bucketing toward it. Work the directive names becomes Remaining-weighted; areas it de-emphasizes ("ignore the docs work") become Stale-eligible — moved to Stale unless the conversation shows them as a still-active user boundary or an unfinished dependency of the focused work. A directive cannot promote to Done something git/conversation shows unfinished, nor demote to Stale a boundary the user still holds.
From the Done and Remaining buckets, pull the facts the remaining work depends on — the things that would be lost on a context clear:
.claude/rules/auto-mode.md on conversation-stated boundaries).Write a plan that reads as a standalone brief — assume the reader has none of the current context. Use no "as discussed above" / "the file we edited" references; name everything explicitly.
## Objective
<one sentence: the task at hand, restated plainly>
## Already done (do not redo)
- <completed step> → <commit / file proving it landed>
## Constraints & decisions (carry forward)
- <decision or boundary that the remaining work must honor>
## Remaining steps
1. <imperative step naming the exact file/path and what changes>
2. ...
## Verification
- <how we confirm done: test command, behavior to observe>
Drop the Stale bucket entirely — that pruning is the point.
When a focus directive was given, the Objective leads with the focused task and explicitly scopes out the de-emphasized areas, so the cleared-context reader inherits the narrowing.
Call ExitPlanMode with the plan from Step 3 as its content. This presents the refreshed plan for approval and surfaces the continuation options — including clear context and continue in auto mode. Do not start executing the remaining steps before approval.
If the session is not already in plan mode, present the plan as your response and offer to refocus from there; the self-contained plan text is the deliverable either way.
When the user opts to clear context and continue in auto mode, the plan from Step 3 is all that survives — which is exactly why it was written standalone. After the clear, seed a fresh TodoWrite list from the "Remaining steps" and proceed.
Under auto mode, keep the narrow Bash(<command> *) permissions this skill already declares — they carry over and skip the classifier round-trip (see .claude/rules/auto-mode.md).
Before calling ExitPlanMode, confirm the plan:
| Context | Command |
|---------|---------|
| What landed on disk | git log --format='%h %s' --max-count=8 |
| Uncommitted surface | git status --porcelain=v2 --branch |
| Confirm a "done" claim | git diff --stat against the named files |
| Seed remaining work | Rebuild TodoWrite from the plan's "Remaining steps" |
development
Debug HTTP APIs: trace requests, inspect headers. Use when a request fails: check status first.
documentation
Render architecture diagrams from text sources. Use when documenting system topology.
tools
Inspect JSON payloads and extract nested fields. Use when parsing API responses.
tools
--- name: no-description allowed-tools: Read --- # No Description This skill has no description and must be dropped with a warning.