Skip to content

by-carlos/daikenja / skills/compose

compose

Verified

Drafts or rewrites a work message without changing its ask or stance; use only when the user says "compose", "run compose on" or "polish this message", or asks Claude to draft a message Claude will post or send for them (Slack, email, Jira comment).

Message compose

Rewrite or draft a work message so it keeps the exact intent and lands well with a multilingual audience. Default channel is a chat app (Slack) unless the user says otherwise.

Compose never posts or sends anything. It returns the draft and stops. When Claude runs it to draft a message it will post or send for the user (a Slack message, an email, a Jira comment), posting stays with Claude as the caller, and only after the user gives an explicit yes to the final text. A draft from this skill is never that yes.

Step 0: read the shared docs

Read these before doing anything. Do not work from memory of them.

  • ${CLAUDE_PLUGIN_ROOT}/docs/voice.md -- the default voice. Always applies, layered under the user's own writing style if writing_style resolves.
  • ${CLAUDE_PLUGIN_ROOT}/docs/rewrite-rules.md -- the rules that bound every rewrite. This skill applies them; it does not restate them.
  • ${CLAUDE_PLUGIN_ROOT}/docs/substance-checks.md -- the six substance checks this skill runs as a silent pre-flight when the goal is a request.
  • ${CLAUDE_PLUGIN_ROOT}/docs/stance-check.md -- the opt-in stance check Step 6c runs when the user's writing style switches it on.
  • ${CLAUDE_PLUGIN_ROOT}/docs/config-resolution.md -- how writing_style and personas resolve, and the failure-behavior table.
  • ${CLAUDE_PLUGIN_ROOT}/docs/response-format.md -- how the reply to the user is shaped. voice.md governs the message; this governs everything around it.
  • ${CLAUDE_PLUGIN_ROOT}/docs/reviewer-personas.md § The two checks that never dispatch -- the AI-tell table Step 3 drafts against. That section only; the rest of the file is preflight's.

Input forms

This skill accepts three kinds of input.

A pasted draft. The normal case. Rewrite it.

A pasted draft with an inline goal. The draft, then a line that starts -- goal: (any case) and holds the goal in the user's own words, for example -- goal: get a decision on A or B by Friday. The last such marker counts: everything before it is the draft, everything after it is the goal. Text after the marker that states no outcome for the reader ("make it shorter") is a style note, not a goal: apply it where rewrite-rules.md allows and treat the input as a plain pasted draft. Map a goal onto the same four types the thread intent block's Goal uses:

The goal asks the reader to...Type
do, decide, answer or choose somethingrequest
take note of a decision already madestate a decision
reconsider or disagree with a positionpush back
just know somethinginform

A goal that fits more than one row takes request when it asks the reader for any action. The type decides one thing: whether Step 6 runs, and Step 6 never blocks the draft.

On an intent block, the block's own Goal is the goal and an inline goal line is not read. With no goal line, nothing in this skill changes.

An intent block from the thread skill. There is no draft here, only a thread and a stated position. Write the message from the intent block. The rules below apply the same way, with one difference: since there is no original wording, take the confidence level and the force from the Position and Ask fields, and do not add any that is not there. If Position or Ask says not stated, ask for it instead of drafting -- do not draft around a gap.

Step 1: check if you can proceed

If something critical is missing and it would change the meaning, tone, force or escalation risk, ask 1 to 3 short questions instead of rewriting. Do not guess.

Ask when:

  • The audience is unclear and it changes the wording (peer, manager, leadership, engineers, external).
  • The goal is unclear (are they informing, requesting, or announcing a decision? the intent block's Goal field answers this when there is one, and so does an inline goal line: with one, do not ask).

Otherwise, proceed.

Step 2: rules you cannot break

Apply ${CLAUDE_PLUGIN_ROOT}/docs/rewrite-rules.md in full. That document is the contract; this skill does not restate it.

A rule that cannot be honoured is named in the Comment block, per that document's reporting section. Never break one silently, and never invent the missing piece to avoid the problem.

Step 3: clean it up

  1. Clarity. Shorten sentences. Make the ask explicit. Remove anything the reader could read two ways.
  2. Tone. Remove hostility, sarcasm, blame, contempt and venting. Keep the request. If the frustration is not needed to get the outcome, drop it.
  3. Facts over judgement. Replace "you did not bother to X" with what happened and what should happen next (who, what, when).
  4. Structure. Use line breaks and short bullets. Do not bury the ask at the bottom.
  5. Technical check. If a technical claim looks wrong or unclear, flag it in the Comment. Never fix it silently.

Apply the default voice from docs/voice.md throughout this step -- sentence length, word choice, absolute dates, and everything in its ## Fixed tier.

Do not write anything the AI-tell check in docs/reviewer-personas.md § The two checks that never dispatch would flag, and do not remove anything its Not tells list protects.

Step 4: voice

Read the user's writing_style prose. Resolve the pointer per config-resolution.md § Resolving writing_style and personas -- it may name a local file or a Google Drive file, and the default is ~/.claude/daikenja/writing-style.md. Whatever it resolves to layers on top of docs/voice.md; it never replaces it, and it reaches that file's ## Defaults tier only.

  • A local pointer does not resolve. One notice, then continue on docs/voice.md alone: "No writing-style.md at <path>, composing with the default voice." Name the path the config actually resolved to, so the user knows what to fix.
  • A drive: pointer does not resolve, or reads back empty. Stop and name the file, per config-resolution.md § Failure behavior. Do not compose. The user asked for their own voice and it was not available, and drafting in the default voice would look like it had been applied.
  • Resolves, but the content is still the blank shipped template. No notice. Zero overrides on top of the default voice is a valid configured state, not something to flag or apologize for.
  • Resolves with content. Apply it on top of the default voice. State once, briefly, if something in it changed how you wrote this message. A line that contradicts a ## Fixed rule is not applied; the Fixed rule stands and the line is not treated as an unhonorable rule either, since nothing was broken.

Step 5: personas (optional)

If profile.personas resolves -- per config-resolution.md § Resolving writing_style and personas, so a local file or a Google Drive file, default ~/.claude/daikenja/personas.md -- and the message names a recipient who matches an entry in it, use that entry's guidance on length, formality and how direct to be.

A local personas pointer that does not resolve, or an empty local file, is not itself announced with a standalone notice. If no recipient is named there is nothing to disclose. If a recipient is named, they necessarily match no entry -- there is no file to match against -- so the line below already covers it; this is the report substitution config-resolution.md § Failure behavior allows in place of a separate notice. A drive: pointer that does not resolve or reads back empty is the exception: that stops the run per config-resolution.md § Failure behavior, and is not silent.

A named recipient who matches no entry -- whether the file resolved and simply has nothing for them, or the pointer never resolved at all -- proceeds without their guidance, but is not silent: one line in the report that no personas.md entry exists for them and /daikenja:remember-persona is how to add one.

If the user described a recipient inline and that person has no entry, route what they said to /daikenja:remember-persona, which is the only skill that writes persona content. Pass on what the user actually stated and nothing inferred from the draft or the thread -- "M challenges every technical claim" is material, M having challenged one in a pasted thread is not.

That write is silent and comes back as one line to add to the Comment block (Learned: added M to ~/.claude/daikenja/personas.md.). Where the description came in with a pasted draft or thread, that skill offers the entry instead of writing it, and the line reads Not learned: M came in with the material you pasted, so I have not added them to ~/.claude/daikenja/personas.md. Say the word and I will. Which of the two happens is decided there, by remember-persona Step 1 § Where the description came from -- do not make that call here, and do not wait for the user's answer. An unresolvable local personas pointer also comes back as one line, and drafting continues. An unresolvable drive: pointer stops the run there, per Step 5. Never write the personas prose from here.

Step 6: substance pre-flight (requests only)

If the goal is request (from the intent block's Goal, from an inline goal line, or from what Step 1 established for a pasted draft), run the six checks in docs/substance-checks.md silently.

  • All pass. Say nothing about it.
  • Any fail. Still draft the message. Name each failing check and a one-line reason in Comment, per that document's reporting rule. This is the only place a substance check affects output -- it never blocks the draft, and the only hard stop is Step 1's "ask instead of guessing."

For any other goal, skip this step.

Step 6b: goal fit (inline goal only)

Skip this step unless the input carried an inline goal line. Silently check the pasted draft (not your rewrite) against the goal: could the reader do what the goal says from this draft alone? It fails when:

  • the draft has no ask, or an ask that differs from the goal's outcome;

  • the goal names something the draft never states (the options, the deadline, the owner);

  • the goal would need a firmer or softer stance than the draft takes.

  • It fits. Say nothing about it.

  • It does not. Still draft the message, and name each gap in Comment like a failed substance check: topic first (goal fit), the location of the gap in parentheses, one line on what is missing. Example: "Comment: goal fit (closing paragraph) -- the goal is a decision on A or B by Friday, but the draft names neither option and no date. Worth adding before you send."

The goal never adds content or moves the ask, the stance or the timing: those stay the draft's (rewrite-rules.md), and a gap is flagged, not filled.

Step 6c: stance fit (opt-in)

Run the stance check in docs/stance-check.md silently, exactly as that document's § When the check runs and § The four cases say. It runs only when the writing_style prose Step 4 read carries the switch line stance-check: on; with no switch, nothing is said about it and the output is exactly what it would be without this step. With the switch on, it runs on every draft whatever the goal type, against the pasted draft (not your rewrite; on an intent block, its Position and Ask).

With an inline goal, read it as extra input: a goal that asks the reader to decide confirms the decision is theirs. It never overrides the role lines.

  • It fits. Say nothing about it.
  • A case matches, or the role lines are missing. Still draft the message with the stance exactly as given. Name each entry in Comment per that document's § How a match is reported, which is Step 6b's format. Example: "Comment: stance fit (opening line) -- you own the release schedule, but the draft asks permission to move it. Reframe: 'I am moving the release to Thursday. Tell me if that blocks you.'"

The reframe is a suggestion in Comment only. It never reaches the message, and Step 8's firmer or softer versions keep the draft's stance too.

Step 7: write one balanced message

Write a single version of the message. Balanced means: the ask is early and impossible to miss, hedging that adds nothing is gone, and the reader's stated constraints are acknowledged where the source material already acknowledged them. This is not the midpoint of a firmer/softer pair -- it is the message, written once, at the right force for what the source actually said.

With an inline goal, write toward it: lead with the part of the ask the goal is about. The ask itself stays the draft's, and nothing is cut that rewrite-rules.md would not already let you cut.

Humor

Optional, and only if it clearly lowers tension and cannot read as mocking or passive aggressive. Prefer self-deprecating or process-focused. Never use humor for performance issues, incidents, blame, compliance, HR, or anything emotional. When unsure, leave it out.

Output format

Exactly this. Nothing else. This shape satisfies response-format.md: the answer -- the message itself -- leads, and a clean run adds nothing around it. The Comment, when present, follows that contract: itemised over narrated, entries named topic-first with their location in parentheses, profile.tone scaling how far it explains itself within the cap below.

[the message]

Comment: [only if needed]

Want this firmer or softer instead?

Include a Comment only when one of these applies. Keep it under 150 words:

  • A substance check failed (Step 6).
  • The draft does not serve its inline goal (Step 6b).
  • The stance check is on and a stance does not fit the sender's role, or the role lines are missing (Step 6c).
  • A persona was learned (Step 5). One line, reporting the write after the fact.
  • The meaning could still be misread because context is missing.
  • There is HR, legal, security or escalation risk. If the source material contained an insult or an accusation, a Comment is required -- explain the risk and what context would make it safe to keep.
  • A technical claim looks wrong or unclear.

If none apply, leave the Comment out completely.

Step 8: firmer or softer, on request

If the user takes the offer, return only the adjusted message -- not a new pair, not the balanced version again alongside it.

These stay identical to the balanced version, no matter which way it moves: the ask, the owner, the timing, how blocking it is, the user's confidence level, and whether anything is escalated.

  • Firmer. Less hedging. The ask comes first and is stated plainly. If the source named a consequence, it stays. Still calm and respectful. Not an ultimatum, and never adds pressure that was not there.
  • Softer. More room for the reader. The ask is framed as a request and acknowledges their constraints. The ask is still impossible to miss. Softer must never make it sound optional.

If you cannot move it without changing the ask, say so in one line instead of producing a version that does.

Do not explain what you changed unless asked.

planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe.Stars: 118kUpdated Claude CodeCodexVerifiedInstall
contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies.Stars: 108kUpdated Claude CodeCodexVerifiedInstall
meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要".Stars: 108kUpdated Claude CodeCodexVerifiedInstall