Define Publication

SkillAI & models

Interview the user to define or revise the standing premise of one writing medium, such as a product blog, newsletter, or brand site, then preserve it in docs/publications/<slug>.md. Always use for a new publication or any change to its readers, promised change, voice, coverage, medium conventions, or how a finished piece is evidenced, even when the file is not named. Start from a rough medium direction. Do not use to brief or draft one piece, and do not use for the app's product meaning, which belongs to define-product.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Define Publication skill

What this skill tells your AI

The instructions your AI receives, as published by toy-crane/claude-code-playbook-template in .agents/skills/define-publication/SKILL.md and read by ahel’s review.

Draw out what one publication is for and how it works, then preserve it as the single current premise later pieces read. A rough direction starts the interview; it does not supply the missing meaning.

Require a direction seed

Start with an existing publication file or at least one rough medium the user wants to write for: "a product blog", "a newsletter for customers", "copy for the brand site". When none exists, state the missing input and stop without inventing readers or creating a file.

Treat a medium name as enough to begin the interview, not as evidence for an unstated reader, situation, promise, or voice.

Read the current context

Read docs/publications/ when it exists. If a file already covers the medium the user names, that file is the current meaning until the user confirms a change; a new medium gets a new file. Read root PRODUCT.md when it exists so product facts can be referenced, and GLOSSARY.md plus relevant subjects from docs/decisions/README.md when present, so the premise uses the project's terms and reuses settled style decisions instead of restating them.

Keep these sources distinct:

  • The user's explicit statements provide intended meaning.
  • The current publication file provides the existing meaning until the user confirms a change.
  • Existing pieces show what has been written, not what was intended.
  • Repository evidence and authoritative external sources provide facts, not editorial choices.

Check only facts that can change the next question or the final premise. When the existing meaning, a new request, and existing pieces disagree, show the difference plainly and ask which meaning should govern. Preserve the current file until that choice is clear.

Draw out the premise

Use the following areas as a check on the final premise, not as a questionnaire the user must fill out:

  • primary readers and the situations in which they read;
  • the change the publication promises its readers;
  • voice and tone;
  • what the publication covers and what it leaves out;
  • medium conventions: the form and typical length of a piece, where finished pieces live in the repository, and how the rendered result is previewed;
  • how a finished piece is evidenced: at minimum, that a reader with no other context can answer the piece's reader questions from the piece alone, and, for pieces carrying code, that every embedded code block and command runs as written;
  • observable success signals;
  • material assumptions and unknowns.

When the reader, situation, or promise is unclear, begin with one concrete reading scene. Ask who meets the writing, in what situation, and what they read instead today. Ask one or two closely related questions in a round; when one answer determines the next question, ask only that one and wait.

Ask open questions when drawing out meaning the user may already have in mind, except for the sample-based voice recommendation below. Do not recommend an answer or supply leading choices in those questions. If an answer stays broad, ask for a concrete reader or an observable change rather than translating words such as "developers", "friendly", or "useful" into settled meaning.

After each answer, reflect only what became newly clear and move to the next meaningful gap. Keep the full list of missing areas internal until the final draft, and do not ask again for meaning the user already supplied.

Recommend voice through prose

Once the readers and purpose are known, propose one short illustrative paragraph and a concise reason it suits them. The user need not choose tone adjectives or presets first. A personal record may serve recollection without promising to change its readers. Keep illustrative details distinguishable from known facts.

Invite a reaction to the actual sentences. When the user corrects the feeling, revise the prose; when comparison would help, add an alternative with the same content and claims. Describe the difference briefly. The sample remains a proposal until accepted or settled within explicit delegation.

Retain the accepted sample and the characteristics the user confirmed, including its intended scope and the choice that confirmed it. Update or remove that example when later feedback replaces or rejects it. Preserve an already settled voice, including concrete instructions in a complete explicit write request; do not invent an accepted example or reopen it just to obtain one.

Handle other choices the user has not made

When the user says a choice has not been made, ask whether to leave it unknown for now or decide it together. Offer options and a recommendation only after the user asks for help making that choice.

When the user delegates a choice, make its scope clear: name the exact choice, the criteria that matter, and any limits. Then disclose the selected answer and the reason. A broad "you decide" does not grant authority over unrelated meaning.

Keep these states separate while working:

  • meaning the user directly confirmed;
  • a choice made within the user's explicit delegation;
  • an unsupported belief the user accepted as an assumption;
  • an important point the user chose to leave unknown;
  • an AI proposal that is not settled yet.

Silence does not settle, delegate, assume, or defer a choice. Record an assumption or unknown only after the user agrees to leave it in that state.

Decide when the premise is ready

Continue the interview while an unanswered point could materially change the readers, the promised change, the voice, the coverage, or how a piece is evidenced, and write the file only once the central reader, situation, promise, and voice are confirmed rather than guessed; until then, say what is still missing and keep going. Ask about other areas only when their answer would change later pieces; do not manufacture detail merely to fill a section. A less central gap may remain only when the user accepts it as an assumption or intentionally leaves it unknown.

Show the full premise once before writing. Clearly separate what the user confirmed, what was decided under delegation, and what remains an accepted assumption or intentional unknown. Ask the user to confirm or correct that draft.

A detailed initial request may count as that confirmation only when it explicitly asks for the file, every material statement can be written without guessing, and no existing meaning conflicts with it. In that case, write without repeating questions or asking for another approval.

Write the current premise

Create or update docs/publications/<slug>.md, one file per medium with a kebab-case slug, using the publication format. Rewrite the current premise in place; do not append chronology, create dated versions, or preserve obsolete meaning beside the current one.

Keep the file about this medium:

  • Reference PRODUCT.md for product facts instead of copying them.
  • Put the medium's accepted sample and confirmed voice in its Voice section. Voice explicitly shared across the author's publications belongs in the existing style subject under docs/decisions/; link to it instead of copying the rule or sample. Use project-knowledge when available, or update that subject directly with the rule, reason, scope, representative example, and confirmation evidence. When no subject exists, create one and link it once from docs/decisions/README.md. A pointer to an unwritten record is not preservation. Durable writes stay within these premise and style owners.
  • Leave individual pieces, their theses, outlines, and briefs to the per-piece workflow.

Do not brief or draft a piece as part of this skill.

Finish

Finish when the user has confirmed the complete premise and a later agent can read the file and brief a piece for this publication without mistaking a guess for intent. Report the file written, the material meaning established or changed, choices made under delegation, and accepted assumptions or intentional unknowns.

Signals

GitHub stars
60
Forks
8
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
define-publication
Source
github.com/toy-crane/claude-code-playbook-template