gw-media-content

Propose up to 10 varied, unpublished posts about a project (app, updates, development, AI-workflow data) for its own social channels. Use when the user asks what to post, for post ideas, a content pla

SKILL.md

GitHub

gw-media-content

If .claude/gw/gw-media-content.md exists in the project, read it first: it holds what is specific to the project.

You propose; the user picks; gw-media-content-now produces each picked piece. Nothing is ever published: the user posts and records where. Every number in a proposal comes from cifras.py or from a cited file and line or commit, never from memory.

Shared material in .claude/skills/gw-media-content/: catalogo.md (post types), reglas.json (per-network limits), plantilla/ (the piece template), nota.md (the project note), scripts/ (cifras.py, comprobar.py, render.mjs).

Project note

.claude/gw/gw-media-content.md (shape in nota.md) sets <dir>, the folder below, and the active channels and languages.

Steps

  1. Read the ground. Done when you can list what is already published, already a piece or still proposed: every <dir>/propuestas/*.md (its Base: commit, the topic and status of each proposal; the newest file gives the Base: for step 2), and the title, ## Canales and ## Publicado of every ficha.md in <dir>/*/*/.

  2. Gather what happened. From the repo root, run python3 .claude/skills/gw-media-content/scripts/cifras.py with --desde <Base> set to the newest Base: (no proposals file yet: no --desde).

    • Without --desde it starts at the last version tag; when that tag points at HEAD it uses the previous one and says so in desde and in the note «points at HEAD» of ultima_etiqueta.
    • A Base that git no longer knows (rebased, garbage-collected) makes it exit 2: run it again without --desde and tell the user the range changed.
    • A shallow clone exits 2 as well: run git fetch --unshallow --tags, then again.
    • A note «found, but its version is not readable» on version_proyecto: pass --version-file <path> or leave the version out of the proposals. Then read what the figures point at: the CHANGELOG section, the log file, the commits, ADRs, plan or roadmap, test and CI results, /opt-tokens and /retro reports, new screenshots. Done when each candidate has a fact you can cite. What is presented as done is merged work; the roadmap and behind-the-scenes types may use open plans, labelled as plans and never as shipped.
  3. Choose. Match the facts to rows of catalogo.md. Drop what step 1 found published, in a piece or proposed. Done when the set has at most 10 proposals, no more than 2 of one topic, and a topic mix (a set that is all releases has missed the catalog); fewer than 10 when the facts run out.

  4. Write <dir>/propuestas/AAAA-MM-DD.md (today; folder created if missing). A new file starts with # Proposals AAAA-MM-DD and Base: <HEAD short sha>. If today's file exists, append a section that opens with its own Base: line and continue the numbering after its last P<n>; the file is never overwritten and the newest Base: counts. One block per proposal:

    ## P1: <short title>
    - **Topic:** <catalog row>
    - **Hook:** <the angle in one sentence>
    - **Format:** <template source, screenshot or text only>
    - **Networks:** <channels, from the active ones>
    - **Facts:** <each fact> (<cifras.py name, file:line or commit>)
    - **Value now:** <why this week and not another>
    - **Status:** proposed
    

    A Reddit proposal also names the subreddit and says whether the note holds its dated rules; without them the proposal says the user must paste them before the piece.

  5. Present and wait. Show the list as a table (id, topic, hook, networks) and ask which to produce. Each picked one goes to gw-media-content-now with its id (AAAA-MM-DD/P3). Commit the proposals file with the project's commit format and a log row.

Choosing well

  • One audience per post. The project's users and the readers of its AI-workflow data are different audiences.
  • Value first. A proposal gives readers something to use, check or answer; announcement-only posts come last.
  • Honest numbers. A figure that does not flatter stays if it is true; cifras.py is the single source.
  • Channel fit. catalogo.md lists the best networks per topic; reglas.json marks which limits are verified (P, S) and which are not (NV, C). Say so when a proposal leans on an NV limit.