Software Search

Search Engineering

Build search features that return the right results, fast.

Quick Reference

When to Use This Skill

• Choosing a search engine or evaluating whether PostgreSQL search is sufficient
• Building full-text search, autocomplete, or faceted filtering
• Designing an indexing pipeline from source data to search index
• Tuning relevance scoring, synonyms, or ranking signals
• Implementing search analytics to measure and improve quality
• Debugging search quality issues (missing results, poor ranking, slow queries)

When NOT to Use This Skill

• RAG and retrieval for LLM context augmentation → ai-rag
• Database query optimization (SQL performance) → data-sql-optimization
• Marketing SEO and search visibility → marketing-seo
• Product analytics and event tracking → marketing-product-analytics
• Backend API design and architecture → software-backend

Workflow

1. Confirm the search problem: engine choice, indexing pipeline, relevance, autocomplete, or analytics.
2. Route RAG, database tuning, SEO, or API-architecture questions to the adjacent skill when product search is not the real problem.
3. Choose PostgreSQL, a dedicated search engine, vector search, or hybrid search from the decision tree.
4. Apply the relevant guidance for indexing, ranking, facets, autocomplete, and measurement.
5. Verify current engine capabilities and hosted-service behavior through the navigation references before final recommendations.

ASCII Flow

Search task
  -> Define corpus, query intent, filters, and freshness needs
  -> Choose database search, dedicated engine, vector, or hybrid retrieval
  -> Design indexing, schema, ranking, synonyms, and hydration strategy
  -> Add relevance evals, analytics, and regression checks
  -> Verify engine-specific behavior and limits
  -> Report quality tradeoffs and rollout plan

Decision Tree

Which search engine?
├── Small dataset (<100K docs), PostgreSQL already in stack?
│   └── YES → PostgreSQL full-text search (pg_trgm + tsvector)
│       └── Outgrowing it? (facets, typo tolerance, sub-50ms at scale)
│           └── YES → Move to dedicated search engine (below)
├── Need instant search-as-you-type with zero ops?
│   └── YES → Algolia (managed, fastest DX)
├── Need full control, complex queries, large scale?
│   └── YES → Elasticsearch or OpenSearch
├── Developer-friendly, simpler than Elastic?
│   └── YES → Typesense or Meilisearch
├── Client-side search (static site, small dataset)?
│   └── YES → Lunr.js, Pagefind, or FlexSearch
├── Need semantic/meaning-based search?
│   └── YES → Vector search (pgvector, Pinecone, Qdrant, Weaviate)
└── Need both keyword AND semantic?
    └── YES → Hybrid search (keyword + vector + reciprocal rank fusion)

Engine Capability Matrix

Capability availability shifts release to release (Algolia added native vector fusion; Qdrant and Weaviate added native BM25). Reverify each engine's current docs before finalizing a recommendation — do not rely on this table's exact wording beyond "capability exists in some form."

PostgreSQL Search (Start Here)

For most applications, PostgreSQL is good enough. Evaluate dedicated engines only when you hit real limits.

tsvector/tsquery — full-text search with language-aware stemming, ranking, and phrase matching. Create a tsvector column, build a GIN index, query with tsquery. Supports ts_rank for relevance scoring and ts_headline for result highlighting.

pg_trgm — trigram-based fuzzy matching. Handles typos and partial matches. Create a GIN index with gin_trgm_ops. Use similarity() or word_similarity() for ranking. Combine with tsvector for both exact and fuzzy results.

GIN indexes — generalized inverted indexes that make full-text and trigram queries fast. Essential for any non-trivial search workload in PostgreSQL.

When to outgrow PostgreSQL search:

• You need faceted search with filter counts (aggregation queries are expensive in PG)
• Sub-50ms latency requirements at scale (>1M docs with complex queries)
• Complex relevance tuning with field boosting, custom scoring, decay functions
• Search-as-you-type with typo tolerance and instant feedback
• You need synonyms, stemming, and language analysis beyond what tsvector provides

Search Engine Architecture

Indexing pipeline: Extract data from source (database, CMS, API) → transform into search documents (flatten, denormalize, enrich) → push to search index. Keep the pipeline idempotent — re-running should produce the same index state.

Index schema design: Define fields, types, and which fields are searchable vs. filterable vs. stored-only. Denormalize aggressively — search indexes are not relational databases. Include all data needed for display in search results to avoid hydration round-trips.

Analyzers and tokenizers: Control how text is broken into searchable tokens. Standard analyzer handles most Western languages. Configure language-specific analyzers for stemming. Add custom analyzers for domain-specific tokenization (email addresses, part numbers, code identifiers).

Synonyms and stop words: Maintain a synonym list for domain terms (e.g., "laptop" = "notebook"). Remove low-value stop words from indexing but keep them in phrase queries. Synonym expansion happens at index time or query time — query-time is more flexible, index-time is faster.

Index lifecycle: Never mutate a live index schema in production. Use index aliases: build new index → swap alias → delete old index. This gives zero-downtime reindexing. For incremental updates, use upsert operations keyed on document ID.

Relevance Tuning

BM25 scoring — the default ranking algorithm in most search engines. Balances term frequency (how often the term appears in a document) against inverse document frequency (how rare the term is across all documents). Handles document length normalization automatically.

Field boosting — weight fields differently. Title matches are typically 3-5x more important than body matches. Boost exact matches over partial matches. Common hierarchy: title > headings > tags > description > body.

Custom ranking signals — layer business logic onto relevance scores. Common signals: popularity (views, purchases), recency (newer content ranked higher via decay function), editorial boost (curated/featured content), user behavior (personalized ranking from click history).

Query understanding — improve what the user meant, not just what they typed. Spell correction (did-you-mean). Intent detection (navigational vs. informational queries). Query expansion (add related terms). Query relaxation (broaden if too few results).

Relevance tuning loop — ship a baseline, measure with analytics, tune iteratively, repeat. Each iteration should move a measurable metric (zero-result rate, MRR, CTR at position 1) not just "feel better."

Query Classification Without a Trained Classifier

Before reaching for a trained intent classifier or an LLM call, check whether the index itself can classify the query. If documents already carry a category or classification field, a semantic-knowledge-graph (SKG) traversal runs a k-nearest-neighbour search against that field and returns the categories most related to the query — no training set, no model to serve. Grainger et al. describe this as asking the graph to "find the category with the highest relatedness to my starting node," where the starting node is the user's query.

What it buys you: classification at index-lookup latency and index-lookup cost, using the corpus you already have. Because the score is computed per query against whatever terms the query actually contains, added context shifts the classification without any retraining — the book's worked example moves driver from a travel reading to a devops reading once install is added to the query.

A second traversal disambiguates. Traverse query → category → keywords and you get a contextualised related-terms list per sense, which separates polysemous terms ("server" as restaurant staff vs. as a machine) into distinct meanings. Grainger et al. warn against the lazy fallback here: given multiple plausible senses, group results by meaning, pick the most likely one, interleave deliberately, or offer alternative query suggestions — an intentional choice beats lumping the senses together.

Where to apply the classification: as an auto-applied filter, as a relevance boost, as a route to a context-specific ranking algorithm or landing page, or as input to term disambiguation.

Limits and guardrails:

• The graph is statistical, not curated — relationships exist only because terms co-occur in the corpus, so expect noise. Set a minimum-occurrence threshold above 1 to suppress false positives.
• Efficacy depends on how well user queries overlap the indexed content. If most queries are for a vocabulary your corpus barely covers, content-derived classification will misread them; user-signal-derived relationships are the complement for that case.
• Scores are comparative, not calibrated probabilities. Treat a negative or near-zero relatedness score as "this category is not the sense," and prefer the known user context over the top score whenever context is available.

The 2026 alternative: an LLM call classifies query intent with no category field and no corpus overlap requirement, and handles queries whose vocabulary the index has never seen. It costs a model call on the query path. Where the latency budget is tight (autocomplete, high-QPS product search) or per-query cost matters, the index-side traversal is the cheaper leg; where budget permits and query vocabulary is open-ended, an LLM classifier is the more capable one. Measure both against the same judged-query set before choosing.

Vector Search API Pattern

Use this pattern when semantic search is a product feature, not just an LLM context retriever.

Request contract:

• query: required non-empty string, with length and character-class limits
• limit: bounded integer, default 10, hard max 50
• offset or cursor: optional pagination, only if the engine supports stable ordering
• filters: allowlisted fields only; never pass arbitrary filter JSON through to the search engine

Response contract:

• stable result ID and display fields
• relevance score or rank, clearly marked as diagnostic when the score is not user-meaningful
• matched source metadata needed for display
• applied query preprocessing version

Operational rules:

• Keep the endpoint stateless and idempotent even if implemented as POST.
• Validate and sanitize input before embedding or query construction.
• Rate-limit by authenticated actor or API key; IP-only limits are weak for logged-in products.
• Add structured errors for invalid input, unavailable embedder, search timeout, and backend failure.
• Log raw query, cleaned query, retrieval mode, top result IDs, latency, and result count.
• Start with exact or small-corpus search to debug embeddings, then move to an index once quality is proven.

Ranking Signal Mix

For product search, semantic similarity is usually one leg of ranking, not the whole ranker. Typical final scoring candidates:

• vector or hybrid relevance
• recency decay
• popularity or engagement
• editorial boost or business rule
• personalization, only when user consent and isolation rules are clear

Do not hand-pick weights from intuition. Calibrate weights against judged queries and analytics slices. If scores come from different systems and cannot be normalized safely, prefer rank-based fusion such as RRF before applying business boosts.

Index-Time vs Query-Time Signals Boosting

Once popularity or engagement signals are part of the ranker, there is a second decision: where the boost is applied. Grainger et al. frame it as scale versus flexibility.

Query-time boosting keeps signals in a separate sidecar collection. Each incoming query first looks up its boosts there, then the boosts are injected into the main query. Because the collections stay separate, signals for one query can be updated by touching one document, boosting can be switched off by simply skipping the lookup, and a different boosting algorithm can be swapped in at any time. That flexibility — and the ease of incorporating real-time signals and running ranking experiments — is the reason it is the more common implementation.

Its costs are structural, not incidental:

• Every search becomes two searches back-to-back; the main query waits on the lookup.
• Only a top-N slice of boosted documents can be injected before query cost becomes unreasonable, so relevance is traded against scalability. A query with hundreds of documents carrying signals will boost only the handful that fit.
• Paging degrades. Covering page 2 means loading more boosts than page 1 did, page 10 more still, so deep paging gets progressively slower and can time out. Worse, boost is only one scoring factor: as the boost set grows between pages, documents can jump onto a page the user already passed or reappear on a later one, producing skipped and duplicated results.

Index-time boosting inverts the problem — instead of boosting popular documents for a query at query time, it writes the popular queries and their boost values into a field on each document at indexing time, and the query simply searches that field. The same signals aggregation feeds both; only the final application step differs. This removes the second query, keeps query cost flat as the number of boosted documents grows, and fixes paging outright, because every matching document carries its boost rather than just the top-N that fit in a query string.

Its costs land on the indexing side:

• Adding or removing a keyword from the model requires reindexing every document associated with that keyword. Incremental per-keyword updates can therefore mean continuous reindexing; batch regeneration can mean reindexing the whole corpus.
• Changing the boosting function needs a migration, not an edit. Reweighting click versus purchase signals means writing a second boost field, reindexing into it, then cutting the query over — otherwise scores fluctuate while the corpus is half-updated.
• Under sustained indexing pressure, separate the servers that index from the servers that serve queries, or indexing CPU and memory will degrade query latency. Several engines expose a mechanism for this (replica types, follower indexes); confirm the specific mechanism and its current behaviour in your engine's own docs.

Choosing: take query-time boosting when the ranking function is still moving — active experimentation, real-time signals, boosts that need to be toggled per request. Take index-time boosting when the model has stabilised and scale is the constraint: deep paging matters, per-query boost sets are large, or query latency is the budget under pressure. The book's own summary of the tradeoff is that query-time is more flexible while index-time is more scalable and gives more consistent relevance ranking.

For the full learning-to-rank pipeline that sits above these signal decisions — feature logging, judgment lists, model training, and reranking — see ai-rag, whose references/learning-to-rank-pipeline.md covers it end to end.

Vector Memory Sizing (Worked Example)

Estimate before choosing a vector index type — memory, not disk, is usually the binding constraint for in-memory ANN indexes (HNSW).

Formula: raw_bytes = num_vectors × dims × bytes_per_value. Add HNSW graph overhead on top (graph edges + metadata); treat 20-50% of raw size as a starting planning range and verify the actual multiplier against the specific engine's current documentation before sizing hardware.

Worked derivation — 1,000,000 documents, 768-dimension embeddings (a common mid-size embedding model output), three storage precisions:

Adding a 30% HNSW graph overhead to the float32 case: 2.86 GiB × 1.3 ≈ 3.72 GiB of working memory for one million 768-dim vectors — before the rest of the document payload (text, metadata) is counted.

How to use this: re-run the same formula with your own num_vectors and dims — never scale a neighboring number instead of recomputing from your corpus size and embedding dimension. Binary and scalar quantization trade recall for memory; validate the recall drop against your judged-query set before committing to a lower precision in production. Confirm current quantization support (halfvec, binary, product quantization) in the specific engine's docs — pgvector, Elasticsearch, OpenSearch, and Qdrant each expose different quantization options and defaults that change across releases.

Common Misdiagnoses

Symptoms that get the wrong fix more often than the right one:

• "Search is slow" → jumping straight to a dedicated engine. Check for a missing GIN index, an N+1 hydration query per result, or an unbounded LIKE '%term%' scan first. Many "we need Elasticsearch" tickets are fixed by an index that was never created.
• "Relevance is bad" → adding vector search. Verify analyzers, stemming, and field boosting are configured correctly before assuming lexical search is semantically incapable. A missing stemmer or unboosted title field often looks identical to "BM25 can't understand meaning."
• "Zero-result spike" → assumed content gap. Check first whether a recent synonym, analyzer, or tokenizer change caused a regression. Content gaps and indexing regressions produce the same symptom but need opposite fixes.
• "Hybrid search will fix our recall" → skipping the judged-query set. Hybrid retrieval reduces but does not eliminate poor recall if the underlying embedding model was never validated against the domain's vocabulary.
• "Facets are slow" → blaming the engine instead of cardinality. Faceting on a free-text or unbounded-cardinality field is usually the actual cause, not an engine limitation.
• "Autocomplete is laggy" → tuning the main index. Autocomplete usually needs its own latency budget and often its own lightweight index or cache; it should not share load or latency budget with full search.

Faceted Search and Filtering

Aggregation queries — compute filter counts alongside search results. Show users how many results match each filter value before they click. This is where PostgreSQL struggles and dedicated engines shine.

Hierarchical facets — nested categories (e.g., Electronics > Phones > Smartphones). Implement with path-based tokens or nested aggregations. Allow drill-down and drill-up navigation.

Range facets — numeric or date ranges (price $0-50, $50-100; last 24 hours, last week). Pre-define meaningful ranges or use dynamic bucketing.

Multi-select vs. single-select — multi-select filters use OR within a facet and AND across facets. Single-select uses exclusive selection. Multi-select requires disjunctive faceting (count all values, not just those matching current filter).

Performance — apply filters before scoring when possible (filter context vs. query context in Elasticsearch). Cache frequently used filter combinations. Pre-compute facet counts for high-traffic pages.

Autocomplete and Search-as-You-Type

Prefix matching — match documents where a field starts with the typed characters. Fast but limited to prefix positions.

Edge n-gram indexing — at index time, generate token prefixes ("search" → "s", "se", "sea", "sear", "searc", "search"). Converts prefix queries into exact match lookups, which are faster.

Completion suggesters — dedicated data structures optimized for prefix completion. Elasticsearch has a built-in completion suggester. Algolia and Typesense handle this natively.

Client-side debouncing — wait 150-300ms after the user stops typing before sending the query. Reduces server load and prevents UI flicker. 200ms is a good default.

Highlight matching terms — show users why a result matched by bolding the matching portion. Most search engines provide highlighting out of the box.

Zero-state and popular suggestions — before the user types, show trending queries, recent searches, or popular categories. Pre-compute these from search analytics data.

Search Analytics

What to track: every query (with timestamp, user ID, session), every click (which result, position clicked), conversions (did the user complete their goal after clicking), zero-result queries, query refinements (user searched again after seeing results).

Zero-result queries — the most actionable metric. These reveal content gaps (you don't have what users want) or search quality issues (you have it but search can't find it). Review weekly and take action: add content, add synonyms, or fix indexing.

Click position — which position users click in search results. If users consistently click result #4 instead of #1, your relevance ranking is wrong. Use mean reciprocal rank (MRR) as a quality metric.

Build the feedback loop: search query → user clicks result → click signals feed back into relevance tuning (boost documents that get clicked, demote documents that get skipped). This is the core mechanism for search quality improvement over time.

Search Quality Evaluation

Analytics are not enough on their own. Keep a judged-query set for the product's most important search intents and re-run it whenever you change ranking, analyzers, synonyms, or business boosts.

Minimum loop:

• define representative queries across navigational, informational, autocomplete, and zero-result-recovery paths
• rate a small set of expected results for each query
• run the engine's native rank-evaluation API or an equivalent offline harness
• compare score deltas before and after relevance changes
• review failures manually before shipping boosts or synonym changes

Use click data to find candidates for the judged set, but do not let click-through alone define quality. Position bias, sparse traffic, and merchandising effects can hide bad ranking decisions.

Common Anti-Patterns

• LIKE '%query%' at scale — full table scan, no index usage, gets slower linearly with data growth. Use proper full-text search instead.
• No analyzers configured — raw text matching misses stemming, case folding, and accent normalization. Users searching "running" won't find "run."
• Ignoring zero-result queries — the cheapest way to improve search quality, and most teams never look at them.
• Re-indexing full dataset on every change — use incremental updates (upsert by document ID) for individual changes and reserve full reindex for schema changes.
• Coupling search schema to database schema — search documents should be denormalized and optimized for query patterns, not mirror your relational schema.
• Not measuring search quality — if you don't track click-through rate, zero-result rate, and query refinement rate, you cannot improve search. Instrument from day one.
• Over-engineering early — starting with Elasticsearch when PostgreSQL tsvector would handle the workload for the next two years.

Known Traps

• Semantic or vector retrieval added as a replacement for lexical recall instead of a complement to it.
• Faceting on fields with unbounded cardinality, leading to expensive aggregations and unusable filter UX.
• Autocomplete implemented against the main search index with no latency budget, flooding the cluster on every keystroke.
• Synonym expansion shipped without governance, creating silent relevance regressions and impossible-to-debug ranking changes.
• Index freshness assumed to be real-time when ingestion pipelines are actually batched or eventually consistent.
• Relevance tuning done from intuition alone instead of using click data, zero-result analysis, and query reformulation behavior.

Verification Gate

Before calling a search design or implementation ready:

• Engine matches current corpus size, query shape, and latency budget
• Latency budgets defined separately for full search (≤200ms p95) and autocomplete (≤100ms p95)
• Index freshness mechanism documented and measured (real-time, near-real-time, or scheduled batch)
• Zero-result rate instrumented and baseline established (target below 5% for mature catalog)
• Click position and refinement rate tracked from day one
• Judged-query evaluation covers navigational, informational, autocomplete, and zero-result-recovery paths
• Representative queries tested — not only happy-path keyword matches
• Synonym and analyzer changes logged with rollback plan

Scenarios

Recipes keyed to common search implementation moments. Each lists the shortest path using patterns above.

S1 — Hybrid BM25 + vector with RRF for a product catalog

1. Index each product document in both a keyword (BM25) index and a vector index (pgvector, Qdrant, or Weaviate).
2. At query time, run the keyword search and the vector search in parallel; collect both ranked result lists.
3. Apply Reciprocal Rank Fusion: score each document as Σ 1/(k + rank) across both lists, where k=60 is a safe default.
4. Re-rank the merged list by RRF score; apply any business boosts (popularity, recency) on top.
5. Tune the relative weight of keyword vs. vector by evaluating against a judged-query set, not by intuition.
6. Monitor zero-result rate; hybrid rarely returns zero, but confirm semantic recall improves tail queries.

S2 — Zero-downtime reindex via index alias swap

1. Create a new index with the updated schema (e.g., products_v2); leave the live alias pointing to products_v1.
2. Run the full reindex pipeline against products_v2; writes to products_v1 continue serving production traffic.
3. Verify document count, spot-check relevance on representative queries against products_v2 before swap.
4. Atomically update the alias: remove products_v1, add products_v2 in a single alias-update call.
5. Confirm production traffic is now routing to products_v2; monitor error rate and latency for 10 minutes.
6. Delete products_v1 only after the monitoring window is clean; keep it for one more deploy cycle if in doubt.

S3 — Autocomplete debounce + per-user prefix prefetch

1. Add 200ms client-side debounce before firing autocomplete queries; cancel in-flight requests on each keystroke.
2. On focus of the search input, prefetch popular completions for the empty-string state from a cached endpoint.
3. After the user's first two characters, shift to a live prefix query against a completion suggester or edge-ngram index.
4. Return a maximum of 5–8 suggestions per query; more options slow perceived response and increase cognitive load.
5. Log every prefix query with the user session; feed click data back into the popular-completion cache weekly.
6. Set a strict 100ms server-side latency budget for autocomplete; use a dedicated index or cache layer to meet it.

S4 — Faceted drill-down with cardinality limits

1. Define facet fields explicitly in the index schema; do not facet on free-text or high-cardinality fields.
2. Cap returned facet values per field (e.g., top 20 by count); expose a "show more" call for long-tail values.
3. Use filter context (not query context) for active facet filters so they do not affect relevance scores.
4. Implement disjunctive faceting for multi-select: recompute counts for each facet excluding its own active filter.
5. Test aggregation query latency at expected data scale; move expensive facets to a pre-computed cache if needed.
6. Monitor facet click patterns in search analytics; remove facets that are never interacted with.

S5 — Zero-result rate dashboard + query rewrite trigger

1. Log every search query with its result count; compute the zero-result rate as a daily metric.
2. Build a dashboard showing the top 50 zero-result queries ordered by frequency; review weekly.
3. For each zero-result query, classify: missing content, synonym gap, tokenization mismatch, or indexing bug.
4. Add synonyms or query expansion rules for synonym-gap cases; update the index for indexing bugs.
5. Add a query-rewrite rule (fuzzy match, spell correction, or query relaxation) for tokenization mismatches.
6. Re-evaluate the top-50 list after each change; target zero-result rate below 5% for a mature catalog.

Navigation

References

• Skill Sources: curated primary sources for search engineering guidance.
• Hybrid Search and Reranking: production pipeline — BM25 + dense vector, RRF fusion, cross-encoder reranking, pgvector in-Postgres option, engine capability classes.

Related Skills

Gate before invoking any foundation below: Each foundation has a When to Apply / When to Skip section. If your task matches a skip-condition, route to the foundation it names instead — don't pull in primitives the task doesn't need.

• ai-rag — Retrieval-augmented generation and vector retrieval for LLM context
• data-sql-optimization — Database query performance and indexing
• software-backend — Backend API design and service architecture
• software-frontend — Frontend implementation including search UI components
• software-architecture-design — System design and component boundaries
• foundations-information-theory — Entropy, mutual information, and KL divergence are the math behind BM25, TF-IDF, and the relevance-tuning section

Freshness Protocol

Search engines, managed services, and client libraries evolve frequently. Verify current information before recommending specific versions or providers.

Trigger Conditions

• "Which search engine should I use?"
• "Is Algolia/Typesense/Meilisearch still the best option for...?"
• "What's new in Elasticsearch/OpenSearch?"
• "Should I use PostgreSQL search or a dedicated engine?"
• "How does pgvector compare to Pinecone/Weaviate?"

How to Freshness-Check

1. Start from data/sources.json for official documentation links.
2. Run a targeted web search for the specific engine or library.
3. Prefer official docs and release notes over blog posts for version and feature claims.
4. Prefer official engine docs for rank-evaluation APIs, relevance-debugging tools, and analyzer behavior.

What to Report

• Current landscape: what is stable and widely used now
• Emerging trends: what is gaining traction (and why)
• Deprecated/declining: what is falling out of favor (and why)
• Recommendation: default choice + 1-2 alternatives, with trade-offs

Fact-Checking

• Known bugs, regressions, framework/compiler/runtime footguns, and version-specific crash or workaround guidance must be verified against current primary web sources before being treated as current fact.
• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.

Learnings Loop

Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.