Path map
Detailed content
🏛️ Four design philosophies
Each provider has a signature: the way of instructing that recurs in all its prompts. Anthropic, OpenAI, Google/Microsoft, and xAI side by side.
The set of style decisions a vendor repeats: rule size, how capabilities are declared, block delimiters, and tone calibration.
Anyone who knows only one philosophy applies it everywhere and pays the wrong cost. Comparing them gives you a range of options to choose from based on the task.
Signature · trade-off · choose based on the task · blind excerpt test (identify the provider from 3 style cues).
The justified principle: the rule explains the intent so the model can generalize; consolidation, security by default, and semantic XML tags.
The "since…" clause is the signature — it generalizes to new situations where a bare "never do X" would fail.
Rule with rationale · consolidation · safety default · semantic XML. Evidence: claude-fable-5.md L52.
Short, imperative rules, prioritized by explicit priority; behavior assembled modularly with channel-specific variants (instant/thinking/voice).
A short rule is easy to reuse across products, but it generalizes poorly to new cases and depends on emphasis (CAPS) and hierarchy to break ties.
Short rule · modular persona · channel-specific variant · explicit priority. Evidence: gpt-5.5-instant.md "Priority order".
Procedural signature: capabilities become named schemas and toolkits; behavior becomes step-by-step with sections and headings, with rules illustrated by strong examples.
Shows exactly how to format, when to use each tool, and the word limit — deterministic and auditable, but rigid outside the expected path.
Sectioned (##/###) · schema/toolkit · step-by-step · rule by example. Evidence: gemini-3.1-pro.md "Formatting Toolkit".
The denser, more opinionated signature: short bullets, each with a philosophical stance; persona-driven behavior in a separate file; XML as scaffolding.
Each bullet states “who Grok is,” not just “what to do.” Lightweight and direct, it delegates to the model — but with little explicit coverage and more variance.
Dense concision · opinionated · personas (separate module) · structural XML. Evidence: grok-4.3-beta.md + grok-personas.md.
The four philosophies contrasted along the axes that matter — rule form, capability declaration, structure, tone — each row with real evidence from the collection.
The chart doesn't say "which is better" — it shows what each optimizes for. A coding agent prioritizes structure + why; a companion prioritizes persona + modularity.
Four axes · evidence for each line (actual path + line) · tool for choosing · deliverable from the 4 subscriptions.
🧩 Product Specialization
A shared core is rewritten when it becomes a chatbot, agent, search engine, Office, voice, or CLI. See what each product adds and removes, using real prompts.
Take the core (identity, safety, format, tools) and apply two deltas: +context (channel rules) and −noise (rules the medium makes irrelevant).
You stop “starting from scratch” and start porting: what’s the channel? what does it require? what does it prohibit? — and the prompt almost writes itself.
Inherited core · delta + (channel rules) · delta − (noise) · deliverable changes by product.
Rendered text conversation, with no external environment. The difference is almost entirely in tone and persona; the core comes through more "purely" here.
The chatbot is the baseline: when reading about another product, ask "what does this do differently from a chatbot?" — the answer is exactly the product's delta.
Persona · principled refusal · markdown OK (rendered output) · comparison benchmark.
Rich environment (files, shell, repo). Dominant delta: autonomy (“keep going until resolved”) and strict tool contracts (order, exclusivity).
Autonomy without a stopping criterion leads to an infinite loop; tools without an order produce conflicting calls. The environment is rich and dangerous.
Autonomy · tool contract · stopping criterion · parallelism. Evidence: comet-browser-assistant.md.
Browses the web and returns facts with provenance: each piece of information has an id {type}:{index} and the response cites it inline, next to each item.
Removes "ask for clarification": in a search engine, ambiguity leads to more searching, not a question to the user. Source Preservation taken to the extreme.
source id · inline citation · −clarify (search, not ask) · authority by cross-reference.
The two format extremes. Office: the deliverable is the spreadsheet/doc, the chat is a "cover note." Voice: removes all markdown, cuts length, speaks quickly.
The Concision Budget becomes "be brief because the work is in the document" (office) or "because nobody listens to a long paragraph" (voice). Same principle, channel-specific rationale.
Artifact (doc/spreadsheet) · cover note · −markdown (voice) · talk quickly. Evidence: claude-for-excel.md, voice-assistant.md.
CLI runs in a terminal: plaintext output, space is precious. Delta: output budget (≤100 words) + the reminder that everything will be printed on a command line.
The 6×3 table (6 products × 3 patterns: Concision, Tool Contract, Source Preservation) is the module deliverable — each cell with line-level evidence.
≤100 words · plaintext · 6×3 table · 1 piece of evidence per cell. Evidence: copilot-cli.md.
📊 Comparison matrix
Three universal problems × four providers. The matrix doesn’t say who’s right: it lays out the trade-offs and becomes a reference tool that wraps up the track.
Rows = universal problems (refuse, cite sources, act without asking); columns = the four philosophies of 3.1. Each cell is a real, citable choice.
Reading prompts in isolation teaches patterns; reading them in a matrix teaches you to decide. The value lies in the contrast the grid reveals, not in any single cell.
Rows = problems · columns = provider · cell = citable choice · value = contrast (reveals trade-offs).
How each philosophy states the behavioral limit: justified principle (Anthropic), brief refusal + tone (OpenAI), procedural override (Google), or blunt refusal (xAI).
Anthropic and Google converge on “don’t narrate the detection mechanics”—different approaches (philosophy vs. procedure), same robustness effect.
Principle · refusal tone · override (check first) · short and direct. Evidence: claude-fable-5.md L43–62, gemini-3.1-pro.md L133.
Conversation resource (Anthropic), source of truth with a dedicated tool (OpenAI), contract with a blind agent — text only (Google), or source render (xAI).
Each philosophy protects something: truth integrity, the downstream channel, visual richness, or conversational transparency.
Resource · source of truth · blind agent (text only) · render. Evidence: gpt-5.5-instant.md L53, gemini-3.1-pro.md L230.
When to proceed without permission and when to stop and confirm. Respects the ending (Anthropic), uses context if it helps (OpenAI), scope owner (Google), confirms irreversible actions (xAI).
It’s the problem that most distinguishes chatbots from agents. Good calibration: acts on reversible actions, confirms irreversible ones, and checks its own work.
Respects the ending · uses context · owns the scope · confirms costly actions. Evidence: vscode-copilot-agent.md, grok-4.3-beta.md.
The hidden price of each rule: a principle generalizes but costs tokens and requires a strong model; a procedure is deterministic but rigid with new cases.
It’s the reading that turns the description matrix into a decision-making tool. The mistake: copying the “most prestigious” cell out of context.
Generalizes vs. token · auditable vs. rigid · model-dependent · contextual cell (product > prestige).
Step by step: choose problems (rows), relevant providers (columns), fill each cell with the evidence (file · line), and note the trade-off before deciding.
Without a source, it's not a matrix — it's an opinion. Citing the actual path is what makes it auditable and reusable; the decision (with rationale) goes to Track 4.
Problems → rows · providers → columns · cell with source · decide by cost. Activity: fill in the 5th column (3×5 matrix).