| name | coding-wrapped |
| description | Creates or refreshes a private local Coding Wrapped website from Claude Code and Codex session history. Use for Coding Wrapped, coding-agent fun facts, AI-coding behavior insights, a personal localhost coding dashboard, new insight cards, or an export of an existing Wrapped. Scans locally, excludes raw transcripts from generated state, and produces one language with aggregate metrics, a Coding Overview, and four-at-a-time illustrated insights. |
Coding Wrapped
Turn local coding-agent history into a private localhost site. Make the default
flow short: infer the language, scan standard local sources, generate the site,
open it, and explain the controls.
Product contract
- Local first — Read local session files and write state only to the user's
machine. Bind the site to
127.0.0.1.
- Facts before stories — Derive every claim from aggregate metrics. Never
invent a number because it makes a better card.
- Fun before judgment — Lead with recognizable habits and memorable
moments. Give light suggestions, never scores, rankings, or diagnoses.
- One language — Infer Chinese or English from the invoking prompt and
generate the whole experience in that language.
- Four distinct scenes — Generate insights four at a time. Their pixel-art
illustrations must use different compositions, not four variations of a
person, a hub, and connecting lines.
This is not a raw transcript viewer, employee-monitoring tool, productivity
score, or cloud analytics service.
Runtime
Require Python 3.9+ and a local web browser. The bundled scripts make no
outbound network requests and need no separate API key; copy and illustration
generation use the active agent when available. Keep the workflow
platform-neutral. Do not require a Codex-only or Claude-only tool.
First-run decisions
- Infer
zh or en from the user's current prompt. Do not ask for a language
choice unless the prompt is genuinely ambiguous.
- Reuse the saved display name when
config.json exists. Otherwise infer a
suitable short name; ask only if no reasonable name is available.
- Treat an explicit request to scan or generate Coding Wrapped as authorization
to read the standard local Claude Code and Codex session directories. If the
user only asks what the Skill does, ask before scanning.
- Do not ask about themes, metrics, layout, or data sources during normal
onboarding. Use the defaults and report which sources were found.
- Use one interface language. Do not expose a language toggle. A later request
to change language is a regeneration request.
- Do not present an implementation plan during the normal first run. Build the
site, show it, then let the user react to something concrete.
Run the workflow
Set SKILL_DIR to this Skill folder. Use a custom CODING_WRAPPED_HOME only
when the user requests one; otherwise state lives in ~/.coding-wrapped/.
1. Scan local history
Run:
python3 "$SKILL_DIR/scripts/build_metrics.py" \
--display-name "<name>" \
--locale "<zh-or-en>"
This creates 7d, 30d, and all aggregate snapshots. It may read:
~/.claude/projects/**/*.jsonl
~/.codex/sessions/**/*.jsonl
Never copy raw transcript text into the generated site. Read
references/privacy-policy.md before changing the
scanner or sending any scan-derived payload to a model.
2. Generate the Coding Overview
Run the brief command:
python3 "$SKILL_DIR/scripts/generate_overview.py" brief \
--output /tmp/coding-wrapped-overview-brief.json
Read the brief. Write one overview in its requested locale with no score or
ranking and up to three sourced recommendations. Keep the summary to two
sentences: select two or three representative facts instead of replaying every
metric, then state the behavior they reveal. Aim for roughly two desktop lines
(45–75 Chinese characters or 18–34 English words). Save the result as:
Read references/coding-best-practices.md
when writing recommendations. Choose only from the brief's
tip_context.candidate_practices. Match the practice to an observed behavior
signal and use its localized action as the factual basis. Do not invent a
generic recommendation when no candidate fits.
{
"copy": {
"<locale>": {
"eyebrow": "...",
"title": "...",
"summary": "...",
"recommendations": [
{
"id": "...",
"practice_id": "an-allow-listed-practice-id",
"title": "...",
"body": "..."
}
]
}
}
}
Persist it:
python3 "$SKILL_DIR/scripts/generate_overview.py" persist \
--input /tmp/coding-wrapped-overview.json
3. Generate exactly four insights
Run:
python3 "$SKILL_DIR/scripts/generate_insights.py" brief \
--output /tmp/coding-wrapped-insight-brief.json
Read references/insight-writing.md and
references/visual-system.md. Read
references/coding-best-practices.md when
writing each Light Tip. Choose the four most personal and visually expressive
findings, not merely the four largest values. Use exactly one allow-listed
tip_practice_id per insight; persistence derives its source links from the
catalog. A tip must match an observed aggregate behavior signal.
When an image-generation tool is available:
- Use the
cw-pixel-diorama-v1 contract and required prompt prefix from
references/visual-system.md.
- Resolve the matching theme reference under
$SKILL_DIR/assets/frontend-template/assets/ and attach it as an actual
image reference. If the tool accepts multiple references, also attach
agent-orchestra-warm.png to anchor the shared rendering language.
- Create exactly four 1536 × 1024 PNG illustrations with four different
compositions.
- Run the reference comparison and six-point visual QA gate before
persistence. Regenerate a drifting image; do not accept it merely because
it is pixel art.
- Pass each approved absolute PNG path as
image_source.
When no image-generation tool is available, omit image_source; the
persistence script uses the four approved fallback illustrations. Do not block
the site on image generation.
Write a batch matching the requested locale and persist it:
python3 "$SKILL_DIR/scripts/generate_insights.py" persist \
--input /tmp/coding-wrapped-insights.json
Every batch must contain exactly four insights. The first real batch replaces
the demo seed; later batches append and survive page refreshes.
4. Serve and open the site
Run the server as a persistent process:
python3 "$SKILL_DIR/scripts/serve_site.py" --open
Default URL: http://127.0.0.1:4173/.
Use the local browser to confirm:
- the configured name and language appear;
- the Overview loads;
- four generated insights load;
- factual metric cards load;
- generated images resolve;
- the Insight deck changes cards with mouse drag and touch swipe while vertical
page scrolling still works;
- the Dashboard scrolls vertically at
1280 × 720, 1440 × 900, and
1024 × 768, without an unexpected horizontal scrollbar;
- export-card routes remain fixed-size and do not shift;
- no raw prompt, code, project name, or local path is visible.
5. Hand off a tiny manual
After generation, give the user the URL and explain only these controls:
- Scan data refreshes factual aggregates and does not regenerate insights.
- Generate insights uses model allowance, usually takes 1–3 minutes, and
always adds four.
- Insight deck supports click, arrow keys, mouse drag, and horizontal swipe.
- Customize chooses which factual metric blocks are shown.
- Export creates a portable local archive:
python3 "$SKILL_DIR/scripts/export_wrapped.py"
End with: raw conversations stay out of the website; only aggregates, generated
copy, approved source links, and generated illustrations are saved.
Refresh rules
- Refresh facts whenever requested; this is deterministic and model-free.
- Do not regenerate the Coding Overview more than weekly unless the user asks.
- Auto-refresh the Overview only after it is at least seven days old and there
are at least three new sessions or twenty new messages.
- Generate new insights only after explicit user action. Warn only that it uses
model allowance and usually takes 1–3 minutes.
- Generate four, never one.
Recovery
- If one source directory is missing, continue with the other and report the
reduced coverage.
- If both are missing, explain the expected default paths and stop before
generating claims.
- If image generation fails, use the fallback illustrations.
- If port
4173 is busy, choose another local port and report the exact URL.
- Preserve existing state on every retry; writes are atomic.