syft-storyline-backfill

name

syft-storyline-backfill

description

Use when the user wants to extend an existing storyline trunk or branch with Syft search results across a time window in Codex or Claude Code without local Python scripts. Also use when the user asks to backfill a branch timeline, repair a thin storyline, derive keyword strategies, merge known events with search evidence on one timeline, or extend hobby-interest logic lines such as IP merch cadence or onsen destination updates.

Syft Storyline Backfill

Use this skill after a storyline tree already exists and the user wants one branch, one trunk, or the full tree extended with older or missing events.

Inputs

Prefer to have:

• an existing storyline tree markdown file
• optional html or notes version of the same tree
• profiles/following_topics.md
• profiles/profile_summary.md
• optional guidance rules

What this skill produces

Create:

• storylines/storyline_backfill_<date>.md
• optionally an updated storyline tree markdown file
• optionally an updated storyline tree html file
• optionally an updated storyline tree json file
• optionally storylines/backfill_keyword_strategy_<date>.md

Important:

• storyline_backfill_<date>.md is a working-facing explanation artifact
• the updated storyline tree files are user-facing final artifacts

Do not confuse the two. Working-note language is acceptable in the backfill memo. It is not acceptable in the final tree outputs.

Workflow

1. Choose the target. The user may specify:

   • one branch
   • one trunk
   • all branches

2. Restate the target meaning in one sentence before searching. Do not search from labels alone. If the branch is a hobby or travel line, restate the branch logic precisely.

   Examples:

   • not just Fate

   • but Fate official merch / collab / event cadence

   • not just Japanese onsen

   • but high-quality scenic ryokan / seasonal opening / local cuisine line

3. Use the retrieval ladder. Before broad search, check whether the branch maps to a followed topic id. Default order:

   1. existing known events
   2. topic top via syft top --topic-id <TOPIC_ID> --days <N> --limit 20 --rich
   3. targeted syft search

   Topic top should be the normal second step because it often returns higher-quality branch-local material than free-text search. Start with limit 20. In most branches, the first 20 topic-local items are enough to show whether the branch can be filled cleanly. Only widen the limit if the branch remains obviously thin after triage.

   Use search when:

   • the branch spans multiple topics
   • the topic top is still thin
   • the branch logic is narrower than the topic title

4. Generate 3 to 6 query angles only when search is still needed. Use:

   • main actors
   • causal mechanism
   • market or policy consequence
   • geographic theater
   • domain jargon
   • fandom or collector cadence terms where relevant

5. Run syft search "<query>" --days <N> --limit <M> --rich. Use multiple small queries rather than one overloaded query.

6. Triage results into:

   • primary evidence
   • secondary evidence
   • discard

7. Merge accepted topic-top and search results into the same branch timeline as known events. Do not create a separate extension pane.

8. Sort the merged timeline chronologically.

9. Preserve the existing tree identity unless the user explicitly asks for a reframe. Backfill should usually enrich the existing trunks and branches, not invent a new ontology.

Primary vs secondary evidence

Primary evidence:

• directly advances the branch storyline
• adds a missing step in the time chain
• sharpens causality or escalation
• for hobby branches, may represent a locally important official release, collab, event, or merch step inside that IP or collector world

Secondary evidence:

• does not advance the core chain directly
• but helps explain spillover, validation, or outer context

Discard:

• same domain but wrong storyline
• generic commentary without event value
• noisy duplicates

Query design rules

• For macro branches, include actors plus mechanism.
• For semiconductor branches, include supply, policy, and market phrases.
• For ACG branches, include IP, event, creator, merch, collab, cafe, exhibition, campaign, and official store language.
• For photography branches, include product line, rumored body or lens names, and release language.
• For travel or onsen branches, include place name, ryokan / hotel / onsen terms, seasonality, scenery, food, promotion, reopening, and local experience words.

Importance rule

Do not filter only by mainstream importance.

• For macro branches, ask whether the result advances the structural chain.
• For hobby branches, ask whether the result would be seen as a meaningful update inside that interest world.

Examples:

• an EVA official collab drop can be primary evidence on an EVA merch branch
• a Fate exhibition or campaign launch can be primary evidence on a Fate release branch
• a new onsen plan or scenic ryokan promotion can be primary evidence on a destination-experience branch if that is the branch logic

Voice rule

When you summarize the extended branch, match the lane:

• macro: analytical and causal
• ACG: scene-aware and cadence-aware
• photography: enthusiast and collector aware
• travel: experiential and local-texture aware

Do not summarize a fandom branch in dry institutional language if that would sound alien to the people who actually care about it.

Backfill boundary rule

This skill extends an existing tree. It should not casually rewrite the editorial taxonomy.

Default behavior:

• keep existing trunk ids
• keep existing branch ids
• keep existing trunk and branch names unless the original names are clearly bad
• enrich timelines
• upgrade thin orphan highlights only when the new evidence is truly strong enough

Do not turn a backfill run into a brand-new classification exercise unless the user explicitly asks for that.

Failure modes to avoid

• adding many search hits that do not deepen the branch
• letting secondary evidence drown out the main chain
• leaving known and extended events in separate panes
• forgetting chronological sorting
• assuming hobby branches need war-or-macro style causality before they can be extended
• skipping topic top even when a clean followed topic id exists
• widening topic top too early instead of probing with the first 20 items
• using the backfill memo tone inside the final HTML or markdown tree
• rebuilding the whole tree into a noisier structure during full-tree backfill
• exposing route labels like topic top 路 or targeted search 路 in user-facing output

Reference files

Read:

• references/keyword-strategy.md
• references/merge-template.md
• ../syft-storyline-tree/references/render-template.md