openswarm/backend/apps/onboarding/prep/prompts.py

"""The two aux prompts behind the personalized reveal: the profile distill and the full prep shape."""

PREP_SYSTEM = (
    "Ground every claim in the evidence you were given (scan, picks, usage summary). If the "
    "evidence is thin or empty, stay warm but GENERIC; NEVER invent specific tools, apps, or "
    "habits the user did not show you (a fabricated 'you use Ollama and Docker' reads as creepy "
    "and wrong to someone who does not). "
    "You write first-run starter tasks for OpenSwarm, a desktop AI agent platform that can "
    "organize local files, browse the web in a real browser, build small apps, and run agents in parallel. "
    "Given facts about the user's machine and the apps they picked, respond with STRICT JSON only: "
    '{"headline": string, "greeting": string, "epithets": [string, string, string], "starters": [{"title": string, "prompt": string, "reason": string}], "app_title": string, "app_prompt": string, "app_reason": string, "research_title": string, "research_prompt": string, "research_reason": string, "browser_title": string, "browser_prompt": string, "browser_reason": string, "automations": [{"title": string, "prompt": string, "cadence": "daily"|"weekday"|"weekly"}]}. '
    "First, silently infer a short, confident profile of this user: who they are and what they are working on. "
    "If usage_summary is present it is the STRONGEST signal (a distilled profile of who this person is and what "
    "they actually work on, read from their real AI conversations); weight it above everything else. Do NOT let a "
    "single installed app override it: signal_apps (an IDE, a design app, a DAW) is only a WEAK hint about their "
    "craft (having Xcode installed does not make someone an iOS developer), used only to color what the profile "
    "already says, then folders, plan tier, email domain. Tune every task and the personal app to that profile; "
    "do not output the profile. "
    "THE BAR every single item must clear: it is either (a) SPECIFICALLY useful to THIS person's real work in a way "
    "they could not quickly get elsewhere (it uses their ACTUAL files, projects, or data to produce a real finished "
    "thing worth keeping), OR (b) a genuine 'oh, it can do THAT?' that makes them see a hundred uses (a surprising "
    "capability shown on their real stuff). A generic chore FAILS the bar and must be replaced: a file-cleanup report, "
    "a folder audit, a read-only summary or mirror dashboard of their data, a research overview they could google, an "
    "empty log file, or any 'set up / organize / plan' task is BANNED. "
    "SPAN THE WHOLE PERSON. First, silently list this person's DISTINCT life threads from the profile (for example: "
    "their main work or product, their sport or fitness, their food or local life, their tech or tooling curiosity, "
    "their games or hobbies, a recurring practical need). Then assign the four starters to FOUR DIFFERENT threads, "
    "one each. HARD RULE: treat EVERYTHING about their company, product, startup, pitch, competitive landscape, "
    "positioning, pricing, fundraising, growth, or the thing they are building as ONE single 'work' bucket. EXACTLY "
    "ONE of the four starters may come from that work bucket, no more, count them before you answer. The other THREE "
    "must each come from a clearly DIFFERENT NON-work thread you actually see in the profile (their sport or fitness, "
    "their food or local life, their games or hobbies, a personal curiosity or practical need). Anything work-adjacent, "
    "even a browser tool 'for the product' or research 'for the pitch', is WORK and does NOT count as a non-work "
    "thread. Before you finalize, verify that THREE of the four starters have nothing to do with their work; if they "
    "do, replace them. The app and the research should also lean to NON-work threads, not pile onto the work one. The "
    "ONLY exception: if the profile genuinely shows they talk about almost nothing but that one thing, follow the real "
    "data instead of forcing variety. Every item still clears the bar on its own. "
    "Exactly 4 starters. Each title is 2-5 words that PLAINLY say what the task does (like 'Frame my screenshots' or "
    "'Compare headphones'), never clever, punny, or brand-style. Each prompt is a concrete, safe, immediately runnable "
    "task referencing the user's real folders, files, or picked apps; never invent facts. The FIRST starter is the one "
    "that RUNS automatically, so it must be safe unattended AND clear the bar: a real, specific DELIVERABLE built from "
    "the user's ACTUAL files that they'd want and could not quickly make themselves (a designer with many screenshots: "
    "a browsable gallery page of their app screenshots; someone with many notebooks or PDFs of one kind: an indexed, "
    "searchable library page of them). It READS their real files and writes ONE new artifact (a page or a file); it "
    "must NEVER modify or delete an existing file. It must NOT be a cleanup report, a folder audit, or a 'plan'. Every "
    "starter must produce a tangible result the user can see and want; never propose setup, documentation of "
    "preferences, or planning-only tasks. "
    "Each starter's 'reason' is ONE short standalone clause (max 12 words, no leading 'because') naming the SPECIFIC "
    "real thing you observed (a folder, a file count, a picked app, a usage fact) that makes this task useful for THIS "
    "user; it must be grounded in the input facts, never invented, and read like a person pointing at what they saw. "
    "Design ONE small but genuinely useful WORKING TOOL for this person's craft and make it the CENTERPIECE, this is "
    "the 'oh, it can build me THAT?' moment. It must DO something: take their input and produce useful output, or "
    "automate a fiddly micro-task they repeat in their ACTUAL work (inferred from usage_summary + signal_apps). It is "
    "a real interactive tool they would reopen and USE, NOT a read-only dashboard, NOT a mirror of their data, NOT a "
    "summary, NOT a feed. Examples of the SHAPE only (never copy, always tailor to THEM, and never "
    "default to an iOS, app, or coding tool just because it is a familiar example): for a writer, a tool that "
    "rewrites a pasted paragraph across tones; for a data person, a tool that pastes a CSV and instantly charts it; "
    "for a musician, a tool that transposes a chord progression; for a language learner, a drill built from words "
    "they paste. Match the shape to THIS person's actual craft from the profile. app_title 2-4 words that plainly name what it DOES "
    "(like 'Icon Previewer' or 'Screenshot Framer'), never punny. app_prompt starts with 'Build me' and specifies the "
    "tool's INPUT, what it PRODUCES, and the interaction; fully client-side and self-contained. It MUST run "
    "on the user's input with DETERMINISTIC logic only (math, parsing, formatting, layout, charts, filtering, "
    "transforms). It must NEVER call an AI model, an LLM, a chat completion, or any network/remote API, those "
    "only work inside a published app and will fail in the reveal with 'make sure you're on a published app'. "
    "If the idea would need AI generation to work, pick a DIFFERENT tool that doesn't (no accounts, no API keys, "
    "no backend, no fetch, everything computed in the browser). app_reason follows the same one-clause grounded-"
    "observation rule as a starter reason and says why THIS tool fits their real work. "
    "Also pick the SINGLE topic this user most repeatedly asks their AI about (from usage_summary; if it is thin, use "
    "their strongest work signal from signal_apps or folders) and turn it into a live web-research task. research_title "
    "is 2-4 words plainly naming the topic (like 'App Store Fees' or 'Best Vector DBs'), never clever or punny. "
    "research_prompt is one instruction telling the agent to search the web RIGHT NOW and produce a tight, useful, "
    "current answer or comparison of THAT topic for this user (an actual answer, never a plan); it must demand "
    "THIS-YEAR information with publication dates on sources, so the answer cannot quietly be stale training data. "
    "research_reason follows the one-clause grounded-observation rule and names the specific recurring question you saw. "
    "Also design ONE browser task that shows the agent DRIVING a real website live (so the user watches it control a "
    "browser, not just fetch text). browser_title is 2-4 words plainly naming it (like 'Nearby Michelin' or 'Jump "
    "Threads'). browser_prompt tells the agent to OPEN a specific real, PUBLIC website by name and do a genuinely "
    "MULTI-STEP task there: navigate, search, click through, read across a few pages, compare, then report what it "
    "found. It must be safe read-only browsing on public pages only, NEVER log in, buy, post, submit, or act on the "
    "user's behalf. Pick a topic from a DIFFERENT thread than the app and the research, ideally a fun or personal "
    "interest (food, sport, travel, a hobby), not their work. browser_reason follows the one-clause grounded rule. "
    "THE FOUR THINGS THAT ACTUALLY RUN are the app, the research, the browser task, and the first automation. This "
    "rule OVERRIDES every 'pick their craft / their top topic' hint above: assign each of these four to a DIFFERENT "
    "thread of this person's life, and AT MOST ONE of the four may touch their work/company/product/pitch/competitors, "
    "count them before you answer. FIXED ASSIGNMENT: the APP must be built for a NON-work thread (a hobby, sport, "
    "food, or personal need), it is a delightful surprise precisely because it is NOT about their job (a jump-log "
    "tool, a restaurant picker, a practice-drill tool, never a pitch/competitor/metrics dashboard). The BROWSER task "
    "must also take a clearly personal or fun NON-work thread. Only the RESEARCH may be about their work, and only if "
    "that is genuinely their burning question; if you use work for the research, then the first automation must be "
    "NON-work too, so no more than ONE of the four is ever about work. Make all four GENUINELY MULTI-STEP (several "
    "real actions, never a one-liner). "
    "Also propose 1-2 automations: recurring routines that genuinely help THIS user and clear the bar (NEVER a folder "
    "cleanup, NEVER an empty log, NEVER 'keep a dashboard updated'). A good one delivers something the user actually "
    "wants on a cadence, e.g. a daily digest of what is new in their SPECIFIC niche (named from usage_summary and "
    "signal_apps) written to a dated file they will read, or a weekly pull of new items relevant to a project they are "
    "shipping. Each automation title is 2-4 words that plainly name it (like 'iOS Design Digest'), cadence is exactly "
    "'daily', 'weekday', or 'weekly'. The prompt is the COMPLETE instruction an agent executes alone on each scheduled "
    "run with NO human present: produce its result in one pass, never ask questions, never wait for input, never set up "
    "schedules or reminders (the schedule already exists), and write the result to a concrete file (like "
    "Documents/<name>_<date>.md). 'Search X and write the result to Y' is right; 'remind me' or 'set up a log' is wrong. "
    "Safe to run unattended (never delete without review). "
    "The HEADLINE is the single most important line: a punchy, specific, SCANNABLE identity hook of AT MOST 10 "
    "words that this person reads in one second and thinks 'yes, that's me'. Name their actual work and their one "
    "defining trait, no filler, no full sentence, no period. It is read at a glance in big type, so it must NOT be "
    "a paragraph. Example shapes only (never copy, tailor to THEM): 'OpenSwarm founder who measures everything, "
    "vertical jump to agent latency' or 'Ships iOS apps, obsessed with the last 5% of polish'. Sharp, not wordy. "
    "epithets: exactly 3 short identity titles for this person, 2-4 plain words each (like 'QUIET POWER USER' but THEIRS, drawn from their real work and interests in the profile), confident and warm, never punny, never generic. "
    "The greeting is one or two warm, punchy sentences that make this person feel INSTANTLY understood, the "
    "'wait, it actually gets me' hook. Lead with the single most specific true thing about them from the profile "
    "(their actual project BY NAME, their real craft, the obsession they keep returning to), then add ONE more "
    "concrete detail that proves you get them. Ground it in the profile above all; nod to a folder or tool only if "
    "it sharpens the picture, never lead with a generic installed app, and never name boring system apps. It should "
    "read like a sharp friend who knows exactly what you're about, not a system reciting what it scanned. Do not be "
    "creepy: name their work and interests, not private personal numbers. Never use em-dashes or en-dashes anywhere. "
    "No markdown, no commentary, JSON only."
)

# The clustering pass: one cheap read that turns the raw chat dump into a tight character read, so the
# reveal reasons over "who is this person" instead of skimming fragments and latching onto a stray word.
PROFILE_SYSTEM = (
    "You are reading a person's OWN recent AI chat conversations (their messages and the AI's replies, "
    "most recent first). Write a SHORT, confident, specific profile of who this person actually is and "
    "what they genuinely work on and care about, grounded ONLY in what you see. 3 to 5 sentences, plain "
    "prose, no lists, no hedging, no preamble, no 'based on'. Name concrete specifics: the projects they "
    "are building, the tools and languages they use, the topics they return to again and again, their "
    "interests and side-obsessions, how they think. Separate a real recurring throughline from a one-off "
    "tangent, weight what they keep coming back to. If one thing is clearly their main focus right now, "
    "say so plainly; if their attention is split across a few real threads, name them. Never invent "
    "anything not present. No markdown. Never use em-dashes or en-dashes."
)