Myfocus

MyFocus Skill

Surface the single highest-priority next task across recent coding-agent sessions — the one thing that deserves the user's next 30 minutes.

Usage

/myfocus                                       # last 5h, all agents, verbose beginner-friendly output
/myfocus --window 24h                          # widen window
/myfocus --cwd $(pwd)                          # only sessions in current directory tree
/myfocus --agents claude_code,codex            # restrict to subset
/myfocus --brief                               # one-screen terse output (legacy minimal layout)
/myfocus --sections top_priority,first_step             # pick exactly which sections to render
/myfocus --list-agents                         # show what scanners are registered
/myfocus --save                                # snapshot to ~/.claude/myfocus/YYYY-MM-DD-HHMM.md
/myfocus --json                                # raw scanner JSON, skip LLM synthesis

What this skill does

When invoked, perform these steps:

1. Run the scanner

Run from the repo root — walk up from this file to the directory holding scripts/collect.py, so the scripts package resolves:

python3 -m scripts.collect --window <window> [--cwd <cwd>] [--agents <list>] [--max-files-per-agent N] [--max-chars-per-session N]

The scanner takes those flags plus --list-agents; forward only those. --brief, --sections, --save, and --json are rendering flags you apply yourself after the scan — the scanner exits 2 on an unrecognized flag.

If --json was passed, print the script's stdout verbatim and stop.

2. Parse and cluster

Parse the scanner's JSON. Group sessions[] by cwd (treat null cwd as its own bucket). Within each cwd, infer the dominant topic from the most recent user_excerpts and assistant_excerpts. Done when every entry in sessions[] sits in exactly one topic.

3. Apply the priority heuristic

For each topic, compute a score:

Score every topic; the highest score becomes top_priority. Tie-break: most recent mtime, then the agent the skill is running inside.

4. Render — composable sections

The output is a sequence of named sections. Each section is independent: render any subset, in any order.

Section catalog (built-in)

Add a new section by adding a row here, defining its renderer below, and accepting its name under --sections. Contributors should send PRs that update only this catalog + a single Markdown template; no Python code change is required to add a section.

Modes

• Default (no flag): emits top_priority, description, why_now, first_step, references, deferred, context — the beginner-friendly verbose layout.
• --brief: emits top_priority, why_now, first_step, deferred, context — terse single-screen output for power users.
• --sections a,b,c: emits exactly those, in that order. Overrides --brief and the default.

Verbose default — beginner-friendly template

Render exactly this layout — plain Markdown, section markers the only emoji:

🎯 TOP PRIORITY: <one-sentence task title>

📝 What this task is
   <2-4 sentences in plain language. Name the code/file/concept; say what
   problem it addresses. Avoid acronyms unless you immediately expand them.
   Write for someone seeing the codebase for the first time.>

🔍 Why this is the top priority right now
   1. <signal name>: <one sentence — what we observed and why it counted>
   2. <signal name>: <one sentence>
   3. <signal name>: <one sentence>
   (Include only the signals that actually fired. Skip those that didn't.)

🚶 First step you can take in the next 30 seconds
   1. <concrete action: cd path, command, or specific decision>
   2. <next action>
   3. <next action>
   (Use real shell commands the user can paste. Show file paths with line
   numbers when relevant.)

🔗 References
   · Session: <agent> / <session_id>
   · Repo: <cwd>
   · Ticket/issue: <if mentioned in transcripts>
   · IDs / artifacts: <any concrete identifiers from transcripts>

📦 Deferred candidates (do these later)
   · <topic title> — <one-liner: why it can wait>
   · <topic title> — <one-liner>
   · <topic title> — <one-liner>

📊 Context
   Window: last <window>
   Sessions scanned: <agent>=N, <agent>=M, ...
   Cross-cutting theme: <only if there is one — otherwise omit this line>

Brief template (--brief)

TOP PRIORITY
  <one-sentence task title>
  Why  <one sentence: which signals made this win>
  First step  <one concrete action — file path, command, or first sub-decision>

DEFERRED
  · <one-liner>
  · <one-liner>

CONTEXT
  Window: last <window>  ·  Sessions: <agent>=N, <agent>=M, ...

Edge cases

• If fewer than 2 distinct deferrable topics exist, omit the deferred section.
• If top_priority was forced (only one viable topic), still include why_now with a single line saying "Only one in-flight topic in this window — no comparison needed."
• If the scanner returned zero sessions, output:

No activity in last <window>. Try /myfocus --window 24h to widen the search.

5. Save snapshot (if --save)

Create ~/.claude/myfocus/ if missing. Write the rendered markdown to ~/.claude/myfocus/YYYY-MM-DD-HHMM.md. Print the absolute path on a final line.

Boundaries

• Never execute the task itself. The skill only points; the user (or the next skill invocation) acts.
• Never write to source code as part of /myfocus. The only write target is the optional snapshot file.
• Single laptop only. No remote fetching, no cross-machine aggregation.
• No external uploads. Do not send session contents to another service. Use only the bounded scanner output needed for synthesis and redact secrets/personal data from excerpts and errors; do not promise the active model itself is local.

Failure modes