corkboard

Goal: a corkboard, one page where a plan's items are pinned as cards, its numbers are drawn as charts, and each card opens into its own page (popup, then Expand to a full page, or New tab). The plan is shown on the boards its data earns, picked from a catalog of well-known formats in references/boards.md:

• Overview: the whole plan for whoever runs it. Headline numbers, a milestone strip, one tile per group with progress, next date and what's stuck.
• Timeline: a Gantt chart on a true date axis, empty stretches folded into marked breaks.
• Decisions: each open choice with its options, who backs which, and the thread behind it.
• Kanban: columns by state.
• Flow: phases in columns, groups in lanes, arrows for what waits on what.
• Moodboard: everything at once, each card wearing its evidence.

Every board zooms: far out it shows summaries, and a click on one zooms into it.

The code is fixed: assets/ holds the boards, the item page and the charts; scripts/build.py combines them with the plan. Each run writes only data: plan.json and the evidence in media/. Scripts run from the folder holding <plan-slug>/, as <skill-dir>/scripts/<name>.py.

Ground rule: provenance

Every pin traces back to the corpus or to a page the corpus names.

• Pictures show the item's own subject: a photo from the corpus; a screenshot of the item's own page (the venue, the listing, the review, the comment); a Wikimedia Commons photo of the landmark or place the item names, with credit and licence.
• Charts draw numbers the corpus states: prices, votes, counts, payments. A physical arrangement (seating, floor plan, route) appears as a screenshot of the venue's real plan.
• An item without evidence carries no picture, and what the corpus doesn't settle becomes a line in gaps.

The two misses this rule exists for, both from a trip board: generic photos pinned to items they weren't of (a dim sum shot on a banquet card, captioned "not the banquet"), and a "row 3" seating diagram drawn from a guess at the venue's layout. Both looked informative and told the reader nothing true.

build.py refuses a picture with no source, a screenshot or Commons photo with no URL, and a Commons photo with no credit or licence. Whether a picture shows this item is your check: view every image before pinning it.

Paths

• Build: a corpus and no plan folder yet. Run steps 1–6.
• Update: an existing plan folder plus new material. Run Update.

Steps

1. Read the corpus. Every file, every sheet tab, the whole chat export. Take only what bears on the plan; health details, addresses, codes and anything else personal stay out of plan.json.

   Done when: every source has been read end to end and is listed in plan.sources with the date it was read.

2. Write plan.json in <plan-slug>/ under the current working directory, to the contract in references/plan-schema.md. Every decision, booking, task and open question becomes an item. Copy facts as the source writes them (dates, prices, names); quotes are verbatim with who and when. Groups, icons, links between items and onDay mappings are your reading; keep a list of them for the hand-off.

   Mark the dates the plan turns on milestone: true, and every choice still being argued, or settled after argument, as a decision with its options and who backs each. These feed the Overview, Timeline and Decisions boards; an unmarked plan gets none of them.

   Done when: every actionable line in the corpus maps to an item or a gap, and every item field traces to a source line.

3. Collect evidence into media/, one media/media.json entry per file (fields in the schema reference):

   • Photos and videos in the corpus: kind: "photo" / "video".
   • The item's own web page: uv run --with playwright python -I <skill-dir>/scripts/capture.py <url> <plan-slug>/media/<key>.png [--selector CSS]. It prints the entry's source fields. A page behind a login, a script wall or a bot check: screenshot it in a real browser (the Neko browser MCP where available).
   • A named landmark or place: a Commons photo, kind: "commons", with credit, licence and the file page URL.

   Open every image. A cookie wall, an error page, a "no article by this name" page, or a different place is not evidence: recapture or drop it.

   Done when: every item that names a place, venue, listing or post has evidence pinned or a line in the hand-off saying why not, and every image has been viewed.

4. Choose the charts. 3–5 dashboard panels for the whole plan, plus chart blocks on the items whose numbers deserve one. A chart needs two or more comparable numbers or a part of a whole; a lone number is a stat or a tile. Follow the dataviz skill where it is installed.

   Done when: every value in every chart traces to a source line.

5. Build and look.

   uv run --with pillow python -I <skill-dir>/scripts/build.py <plan-slug>
   uv run --with playwright python -I <skill-dir>/scripts/check.py <plan-slug>

   build.py prints every rule break at once; fix them all and rebuild. It then prints the board ranking with each board's reason and builds the top four. Set plan.boards only to overrule the ranking, and say why in the hand-off. check.py shoots every built board, a popup and a full page at desktop (light) and phone (dark) into <plan-slug>/check/.

   Done when: check.py passes and every shot has been viewed with no overlapping cards, clipped labels or empty panels.

6. Hand off. Where an Artifact tool is available, publish artifact.html with any video files as supporting files at media/<file>, then rebuild with --artifact-url <url> and republish so New tab links resolve.

   Done when: the user has the folder path (and link), the boards built and why, the list of additions from step 2, every item left without evidence, and the gaps.

Update

1. Read the new material and the current plan.json.
2. Change plan.json: new items, changed states, dates and prices, gaps the new material settles. Keep existing ids: saved card positions are keyed by them.
3. Run steps 3–6 for the changed items.

Done when: every change in the new material is in plan.json and listed in the hand-off, and the board is rebuilt and checked.

Style

• Accent: one colour per plan, light and dark (plan.accent), taken from the plan's own materials where they have one (a wedding's palette).
• Layout lives in the data: the order of groups, items and dashboard arranges the boards. On Moodboard, readers drag cards to rearrange, and their moves stay in their browser. Boards whose positions encode data (Timeline's dates, Kanban's states) don't drag.
• Visual style is fixed for now: one design language with a per-plan accent. Board format is the choice that matters.