🪞 Teaching Gemini to Write Sentences, Not Word Salad
🎯 The Problem
📅 Every day, a reflection note captures the day’s readings, blog posts, videos, and thoughts.
🏷️ Each reflection deserves a creative, emoji-enriched title that distills the day’s themes at a glance.
✍️ Previously, these titles were crafted manually — time-consuming and easy to forget.
🚫 Untitled reflections shouldn’t be posted to social media.
🎮 The Title Game
📊 We analyzed over 20 existing reflection titles and reverse-engineered the creative game behind them:
| 📅 Date | 🔗 Content Titles | 🏷️ Result |
|---|---|---|
| 📆 2026-03-23 | 📚 A Gentle Afternoon…, 🤖 The Crucible of Constraint…, 🏛️ The Forgotten Commons… | 🕊️ Gentle 🚪 Constraint 🏛️ Commons 📚🐔🤖🏛️📺 |
| 📆 2026-03-10 | 📚 Hold Me Tight…, 🏗️ Functional Refactoring…, 🤖 The Agentic Playbook…, 🔊 Teaching the Robot to Breathe… | 🫂 Tight 🏗️ Functional 🤖 Agentic 🗣️ Breathe 📚 |
🔑 The key insight: each title contributes exactly one word, and those words should form a coherent phrase or sentence — not a keyword dump.
🧠 The Word Salad Problem
🤖 Our first prompt simply said “pick one interesting word from each title and arrange them into a phrase.”
📋 Small models like gemini-3.1-flash-lite-preview produced output like:
⏳Time 🧠Memory ⭕Stewardship 🌅Horizon 🕸️Web 🤖Vault 🖼️Generation 🏛️Democracy 📉Fallout
🫠 That’s a grocery list, not a sentence.
🔤 Even the spacing was inconsistent — emojis were glued to words without spaces.
🏗️ Iteration 1: Emphasize Coherence
📝 We rewrote the prompt to insist on coherent sentences, allow reordering, and permit small filler words.
📐 We added a normalizeEmojiSpacing post-processor to ensure “🕊️ Gentle” not “🕊️Gentle”.
🤏 The spacing fix was simple — a regex that inserts a space between emoji and word characters.
📊 Result: spacing fixed, but still word salad from the lite model.
🏗️ Iteration 2: Structured Sentence Building
💡 The breakthrough idea: separate grammar from word choice.
🎯 If we commit to words too early, we paint ourselves into a corner — there’s no grammatical structure to hang them on.
🧩 The Five-Step Process
- 📖 Full Word Inventory — label ALL words in ALL titles with parts of speech (nouns, verbs, adjectives, etc.) — no choices made yet
- 📐 Sentence Templates — draft 2–3 grammatical structures using only POS labels — focus purely on grammar
- 🧩 Fill Templates — go back to the full inventory and try multiple word combinations per template, one word per title
- ⚖️ Compare & Iterate — if nothing reads naturally, swap words, try new templates, keep iterating
- 🎨 Emoji Enrichment — prefix each chosen word with a relevant emoji; filler words stay bare
🔓 By deferring word selection until after the grammatical skeleton exists, the model has room to explore.
🎯 Any word from a title is fair game — even “of”, “the”, or “in” — because filler words are often the grammatical glue.
🚨 As a last resort, the model may skip a title or double up, but only when coherence truly demands it.
🤖 Model Matters
🧪 We tested the same prompt with two models side by side:
| 🤖 Model | 📅 2026-03-21 (9 titles) | 🎯 Verdict |
|---|---|---|
| 🏃 gemini-3.1-flash-lite-preview | ⏳ Time 🧠 Memory ⭕ Circle 🪞 Mirror 🕸️ Weave 🤖 AI 🎨 Image 🏛️ Democracy 📉 Fallout | 🫠 Keyword list |
| 🧠 gemini-2.5-flash | ✨ Smarter ⏳ Time and 🧠 Memory 🕸️ weave a 🤼 struggling, 💥 dismantling ❤️ heart, ⚙️ driven by 💭 thoughts | ✅ Real sentence! |
🏆 gemini-2.5-flash is now the default — its thinking capability handles the multi-step reasoning the prompt requires.
🔄 The fallback chain is gemini-2.5-flash → gemini-2.5-flash-lite → gemini-3.1-flash-lite-preview.
🧹 Post-Processing Hardening
🤖 Models sometimes return unexpected formatting that needs cleanup:
| 🐛 Issue | 🔧 Fix |
|---|---|
🔤 Missing spaces: ⏳Time | 📐 normalizeEmojiSpacing regex inserts space between emoji and word |
📎 Backtick-quoted fillers: `of` | 🧹 Strip all backticks from parsed title |
| 📦 Code fences around output | ✂️ Existing fence-stripping logic |
🏗️ Architecture
🧱 The solution maximizes deterministic computation and minimizes what Gemini must do:
| 📦 Component | 📂 Path | 🎯 Purpose |
|---|---|---|
| 📚 Library | scripts/lib/reflection-title.ts | 🔧 Deterministic extraction + structured AI prompt |
| 🧪 Tests | scripts/lib/reflection-title.test.ts | ✅ 57 tests across 9 suites |
| 🧪 Manual test | scripts/test-reflection-titles.ts | 🔬 Live A/B comparison against existing titles (CI-safe skip) |
| ⏰ Scheduler | scripts/lib/scheduler.ts | 📅 At-or-after scheduling at 10 PM Pacific |
| 🎛️ Orchestrator | scripts/run-scheduled.ts | 🔄 Vault sync, idempotency, yesterday catchup |
| 📋 Spec | specs/reflection-title.md | 📐 Product and engineering specification |
🛡️ Safety Gate
🚫 Untitled reflections are blocked from social media posting:
- 📋
isUntitledReflection()checks if a reflection’s title is just the bare date - ⛔
isPostableContent()returns false for untitled reflections - 🔒
getPriorDayReflectionIfNeeded()skips untitled reflections
🧪 Testing Strategy
✅ 57 tests organized across 9 suites covering all pure functions.
🔬 A separate manual test script reads real reflections, strips their titles, regenerates them, and displays original vs generated side-by-side.
🏗️ The manual test gracefully skips when GEMINI_API_KEY is not set — safe for CI, useful for local iteration.
💡 Lessons Learned
- 🧠 Defer decisions — committing to words before having a grammatical structure produces word salad
- 📐 Grammar first, content second — POS templates create the skeleton; words fill in the bones
- 🤖 Model capability matters — a structured multi-step prompt needs a model that can think through steps
- 🧹 Always post-process — models are creative with formatting in ways you don’t expect
- 🔬 Test with real data — a manual comparison script was invaluable for rapid iteration
📚 Book Recommendations
📖 Similar
- 🧠 Atomic Habits by James Clear
- 📓 The Artist’s Way by Julia Cameron
- 🔤 🦢 The Elements of Style by William Strunk Jr. and E.B. White
🔄 Contrasting
- 🤖 Gödel, Escher, Bach by Douglas Hofstadter
- 📊 🤔🐇🐢 Thinking, Fast and Slow by Daniel Kahneman
🎨 Creatively Related
- 🪞 The Midnight Library by Matt Haig
- ✍️ Bird by Bird by Anne Lamott
- 🔧 💺🚪💡🤔 The Design of Everyday Things by Don Norman
🦋 Bluesky
🪞 Teaching Gemini to Write Sentences, Not Word Salad
AI Q: 🤖 How do you force AI to write like a human instead of a keyword list?
🤖 Large Language Models | ✍️ Creative Writing | 🧠 Prompt Engineering | 🧱 System Design
— Bryan Grounds (@bagrounds.bsky.social) 2026-03-26T07:29:49.194Z
https://bagrounds.org/ai-blog/2026-03-25-reflection-title-generation