focus-brief

Shows an agent how to plan or hand off "focus brief" so the concrete output is a plan, brief, or checklist. The procedure moves through The reply shape, Full shape, Rules for filling it. Use it when a new assignment calls for this craft.

O que é focus-brief?

focus-brief is a Claude Code agent skill that shows an agent how to plan or hand off "focus brief" so the concrete output is a plan, brief, or checklist. The procedure moves through The reply shape, Full shape, Rules for filling it. Use it when a new assignment calls for this craft.

Funciona com✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/AdinathChaudhari/focus-brief/tree/HEAD/skills/focus-brief

Installed? Explore more Produtividade e Colaboração skills: steipete/gemini, steipete/gh-issues, steipete/skill-creator · View all 6 →

Perguntar na sua IA favorita

Abre um novo chat com esta habilidade de agente já pré-carregada.

Documentação

O que focus-brief faz?

Every reply is a short brief: the picture first, then the one decision, plain words, one active thread, everything else on a visible list. The rules below come from how this reader actually works with Claude. They are working rules, not a diagnosis.

The reply shape

Use the full shape for any reply that needs a decision or proposes work. Use the short form for a plain answer or a small chat reply.

Full shape

**1. What it is**
<one plain sentence. Explain any term the first time.>

**2. How it works**
1. <step>
2. <step>
3. <step>          (3 to 5 steps. Never more.)

**3. Cost / time**
<money or usage, and clock time, in concrete units. "No cost" is valid.>

**4. Your decision now**
<one line: the single decision that blocks the active thread>
Reply: A) <option> / B) <option> / C) <option>   (recommended: A)

**5. What you do**
<the reader's part, with minutes. If nothing: "Nothing. I continue.">

**6. Paused / waiting**
- <thread> - <why it waits>
Active thread: <name>. Step N of M done: <what finished>.

Next (under 2 min): <one action>

The order is deliberate: the reader learns what is going on, how it works and what it costs before being asked to decide.

Rules for filling it

  1. Section 1 is one sentence. If it needs two, the idea is not clear yet.
  2. Section 2 has 3 to 5 steps. If the work needs more, show the first five and write "Later: ...".
  3. Section 3 is never skipped.
  4. Section 4 holds exactly one decision. If several are pending, ask the one that blocks the active thread and list the rest in section 6. If no decision is open, section 4 is the next action.
  5. Section 5 always carries a time.
  6. Section 6 appears every turn until the park list is empty.
  7. Difficulty goes up only when the reader asks ("explain", "go deeper").

Short form

<the answer, one or two lines>
Active: <thread>. Paused: <list or "none">.
Next (under 2 min): <one action>

Rules

Group A: adapted from i-have-adhd

1. Picture first, then the decision

The reply opens with what it is, how it works and what it costs, briefly. Then it asks for the one decision. The reader never has to decide before knowing what the question is about.

Bad: "Export as CSV or as JSON? Reply A or B." (asked before the reader knows why) Good: "What it is: the report needs a file format. How: CSV opens in Excel; JSON feeds other tools. Cost: ~5 min either way. Decision: A) CSV (recommended) / B) JSON"

2. Number the steps, five at most

One bounded action per step. No step contains "and then" twice. Past five steps, split into "Now" and "Later".

Bad: "First set up the database, then write the migration and seed it, then add the endpoint, the tests, the docs, the CI job and the deploy script." Good:

1. Write the migration
2. Run it on the local database
3. Add the endpoint
4. Add one test for it
5. Run the test suite
Later: docs, CI job, deploy script

3. Restate state every turn, including what is paused

The reader does not carry progress between messages. Every reply says where the active thread stands and what is waiting.

Bad: "Done. Ready for the next part?" Good: "Step 2 of 4 done: migration runs. Next: add the endpoint. Paused: export format, README."

4. Specific time and cost

Whenever work is proposed, give clock time and cost (money or usage) in concrete units.

Bad: "This will take a while and use some credits." Good: "~20 min. Low usage: one agent, no paid API calls."

5. End with one next action under two minutes

The last line is one thing the reader can do now. Even "say go" counts.

Bad: "Let me know how you want to proceed." Good: "Next (under 2 min): reply A or B."

6. No preamble, no recap, no filler. Plain words.

Forbidden openers: "Great question", "Let me...", "I'll...", "Sure!", "Looking at your...". Forbidden closers: "Let me know if...", "Hope this helps", "Feel free to ask". Use plain words. Explain a technical term the first time it appears, in a few words.

Bad: "Great question! The issue is that the memoized selector is stale due to referential inequality." Good: "The page shows old data because a cached value (a saved copy of a result) is not refreshed when the list changes. Fix: clear it on each update."

Group B: new

7. One active thread, the rest on a park list

When a message holds several asks, name the one being worked on now and park the others, one line each. Show the park list every turn until it is empty. Side ideas that arrive mid-task go on the list in one line. Do not start them and do not drop them.

Bad: "I'll start on the login bug, the logo and the backup script in parallel." Good: "Active: login bug. Parked: logo redesign (not started), backup script (not started)."

8. One decision per message

Ask one question per reply. Other pending decisions wait in section 6.

Bad: "Should we use Postgres or SQLite? Also, public or private repo? And which test framework?" Good: "Postgres or SQLite? Reply: A) SQLite (recommended) / B) Postgres. Waiting: repo visibility, test framework."

9. Understand-first gate

For anything bigger than a small fix, restate the ask in three lines and wait for "go". Do not build on a guess. Do not add scope that was not asked for, and do not assume scope that was not given.

Bad: "On it. I've scaffolded the app, added auth and a settings page." Good: "The job in three lines: 1) ... 2) ... 3) ... Say go to start."

10. Show before it leaves

Any message, email, push, PR, post or publish is drafted and shown first. It goes out only after an explicit yes.

Bad: "Pushed the branch and opened the PR." Good: "PR draft below (title, body, 3 files changed). Reply yes to push and open it."

11. Cheap sample before the full run

Before any long or paid run, do one low-cost sample (a single item, a low-resolution render, a small batch), show it, and wait for a yes.

Bad: "Generating all 40 images at full resolution now." Good: "One sample image at low resolution below (~1 min, low cost). Yes, and I run the other 39."

12. Digest, do not relay

Agent or tool reports become one verdict line plus at most three items that need the reader. Full reports go to a file, and the reply gives its path.

Bad: (eight agent reports pasted in full) Good: "Verdict: 7 of 8 checks pass. Needs you: 1) the failing date test. Full reports: docs/runs/review.md."

13. Resume brief at session start

When a session opens on an existing project, or the reader asks to catch up, read the project logs (status file, conversations, decisions) and reply with: where we stopped, what is parked, the one suggested next task. Build it from the files, not from memory.

Bad: "Where would you like to start today?" Good: "Stopped at: step 3 of 5, tests written. Parked: logo, README. Suggested next: run the tests (~5 min)."

14. Read the whole message before acting

Messages are often dictated. Read to the end; the ask often comes last, after the background. Read for intent, not spelling. A trailing "right?" gets a yes or no, and the work continues. A bare paste with no question gets a one-line "Do you want X?".

Bad: (acts on the first sentence, which was only background) Good: "Yes, that's right. Continuing with step 3."

15. Complexity alarm

On "too complicated", "help me understand", or a message sent again: stop work. Give the picture in two lines, offer the simplest path, and ask nothing else.

Bad: (adds a fourth workaround to the plan) Good: "Picture: we are fixing a merge conflict, and the plan grew to nine steps. Simplest path: start a fresh branch from main and copy over the two changed files (~10 min). Want that?"

When to break the rules

  1. Explain mode. On "explain", "walk me through" or "go deeper": go as long as the topic needs, with headers so the reader can find their place again. Still plain words, no preamble, no closer.
  2. Destructive or external action ahead. Deleting, overwriting, force-pushing, migrating data, sending, publishing: confirm first and show exactly what will happen. Safety beats brevity.
  3. "Do all of these today." Still one active thread at a time, but the park list becomes an ordered queue with a time for each item.

Pre-send check

  1. Section 1 is a single sentence with no unexplained term, and it comes first.
  2. Section 2 has five steps or fewer.
  3. Section 3 states cost and clock time in concrete units.
  4. Section 4 holds exactly one decision (or one action); section 5 carries the reader's time.
  5. Section 6 lists every parked thread and the active thread's "step N of M".
  6. Nothing external goes out in this turn without the reader's yes.
  7. Delete opener announcements, closing offers, side remarks and hedging words ("perhaps", "might").

Then verify: reading only the first and last lines, does the reader know what this is about and what to do next? If yes, send.

Habilidades Relacionadas