idea-discovery-workflow - SKILL.md Agent Skill

name: idea-discovery-workflow description: Use when running or designing recurring product/open-source idea discovery, daily or scheduled idea mining, thesis-first AI product/OSS bet discovery, 3-bucket final selection across ai_oss, ai_product, and ai_prosumer, Product Shape gates, independent Report Reader checks, candidate-ledger replenish proof, Promotion Gate workflows, multi-role Thesis Scout/Signal Scout/Critic/Competitor/CEO evaluation, evidence-to-decision pipelines, idea backlog memory, idea handoffs, handoff-ready idea dossiers, or automation prompts for finding AI-native products, high-star open-source projects, complete AI product directions, CLI, MCP, Skill, agent workflow, and developer-tool ideas. This skill orchestrates discovery and uses ai-founder-playbook for judgment, pressure testing, competitor analysis, and market reasoning.

Idea Discovery Workflow

Use this skill as the orchestration layer for recurring idea discovery. It turns AI-era theses, product/OSS bet sketches, community signals, product/platform news, competitor gaps, and open-source ecosystem changes into a small number of rigorously reviewed product/OSS bets.

This skill is intentionally separate from ai-founder-playbook:

idea-discovery-workflow owns workflow, roles, evidence memory, run artifacts, report shape, and automation integration.
ai-founder-playbook owns founder/open-source judgment: pressure tests, market scans, competitor reasoning, scoring, and launch advice.

Core Run

For a normal run, load only the references needed for the task:

Read source-policy.md to build the thesis-first discovery plan, evidence sweep, AI relevance gate, and product/OSS promotion gate.
Read workflow.md for the DAG, thesis generation, promotion gates, history relation gate, and iteration gates.
Read role-contracts.md when assigning real or simulated roles.
Read report-format.md before rendering the final report.
Read memory-schema.md when persisting or reading historical signals, ideas, competitors, decisions, or graph edges.
Read handoff-mode.md when the user asks to hand off one or more stored ideas.

Use ai-founder-playbook during Candidate Draft, Critic Review, Competitor Check, and CEO Decision. Do not duplicate its rubrics here.

Built-In Scripts

scripts/idea-scout-kit.mjs [topic...] generates a thesis-first discovery plan, thesis seeds, Product Shape templates, AI relevance and promotion gates, source modules for evidence sweep, candidate-ledger templates, independent reader-review prompts, history-relation tables, and Red Team questions. When topics are provided explicitly, it treats them as thesis constraints and also generates topic-guided evidence queries. It does not fetch the web; use it to plan and normalize searches.
scripts/init-store.mjs creates the local JSONL evidence store under ${IDEA_MINER_HOME:-$HOME/.idea-miner}. It also honors the legacy CODEX_IDEA_DISCOVERY_HOME variable for existing installs.
scripts/idea-handoff.mjs [--session-prompt] <idea name...> resolves one or more stored idea dossiers, copies them into temporary handoff files, and can write prompt files suitable for new Codex sessions. It does not browse the web or create sessions by itself; the skill should call host session tools when available.
scripts/check-run-artifacts.mjs <run_dir> checks that a completed run has a report, source notes, handoff index, independent reader review, candidate-ledger proof when needed, per-idea JSON/Markdown dossiers, Product Shape, and source-backed claims. Run it after writing artifacts whenever shell is available.

Execution Rules

Always use current web/realtime tools for current claims, competitors, adoption, pricing, releases, and source freshness.
For any web/realtime/search step, first try the Grok search MCP (prefer mcp__grok_search.grok_web_search for web search; older runtimes may expose grok_search.grok_ask / mcp__grok_search.grok_ask with search: "web"). If that MCP is unavailable, times out, fails, or cannot cover the needed source class, fall back to Codex's built-in web/search/browser/GitHub tools and record the fallback in source notes or coverage notes. Do not assume native X/Twitter search works through the CLI-backed Grok MCP; use mcp__grok_search.grok_x_search only when it is exposed, otherwise use web search or another available tool and mark X-only coverage as limited when needed.
If a source cannot be accessed, mark it 未覆盖/受限; do not infer content.
Default discovery is thesis-first and imagination-led, not evidence-first or complaint-mining-first. Unless the user asks for a narrow market scan, generate a thesis portfolio and product/OSS bet sketches before collecting evidence.
Treat evidence as a brake and sharpening tool: use current sources to support, challenge, kill, or refine bets after they exist. Do not reward "one complaint -> one small checker" as a final idea.
Use fit gates to exclude unrelated raw opportunities such as physical goods, local services, inventory, hardware manufacturing, or pure operations plays unless they can be reframed as a complete product or high-star OSS opportunity.
Default final selection is bucketed: up to 3 ai_oss, up to 3 ai_product, and up to 3 ai_prosumer. Each final idea should be AI-core or AI-native workflow by default; AI-leveraged ideas need unusually strong product/OSS proof, and non-AI ideas stay out unless the user explicitly widens scope.
GitHub Actions, CI gates, PR comments, templates, hooks, checklists, and thin wrappers are integration surfaces only. They cannot be the body of a final idea unless attached to a broader complete product or high-star OSS project.
Before final selection, compare every candidate with stored ideas and label it as new, update_existing, duplicate_of, revives, merged_from, splits_from, or adjacent_to.
Default to a rigorous replenish workflow. Use real sub-agents or multi-agent tools when they help cover more sources, run independent critique, or refill the candidate pool after vetoes; otherwise simulate roles and state that.
Use expensive debate only for high-disagreement or high-value candidates.
Aim to return up to 9 product/OSS bets that pass the current standard: at most 3 per final bucket. Do not lower the bar to fill any bucket. Existing ideas with only incremental evidence should be reported as backlog updates, not counted as new final ideas. If a bucket has fewer than 3 new or meaningfully changed bets, treat that bucket as underfilled and write candidate-ledger.jsonl rows showing the thesis/source replenish rounds and why coverage was exhausted.
Keep full reports as normal Markdown before any host-specific control block.
Before saving final artifacts, run the Product Shape Gate and independent Report Reader check. A reader who did not participate in discovery must be able to explain each final idea as: product/repo form, target user and task, inputs or permissions, core objects, outputs or state, user actions, first-version boundary, and why it has product/OSS body beyond a prompt, checker, Action, wrapper, dashboard, or platform hook recipe. Save that review as reader-review.md or reader-review.json. Rewrite or reject ideas that remain abstract, jargon-heavy, field-filled, story-theater, or integration-only.
Final reports and reader reviews must be readable Chinese. Necessary proper nouns, product names, repo names, and common technical abbreviations may stay in English, but uncommon English terms or invented object names must get a Chinese parenthetical gloss or an explanatory Chinese sentence the first time they carry the explanation. Do not let a list of English object names do the work that Chinese product prose should do.
For every final idea, persist a handoff-ready dossier. Persist a paused dossier only for a strong idea with one clearly named unresolved issue; do not save vetoed, weak, internal-only, or "could be a small tool" ideas as handoff work. Later handoff requests should read the stored dossier first and should not repeat source discovery unless the user explicitly asks for a current refresh.
After persistence, run the artifact checker when shell is available. If it fails, fix the files and rerun it; do not present a failed artifact check as a normal success.

Automation Prompt Boundary

Automations should stay thin: schedule, destination, language, display rule, final bucket target, and the instruction to run this skill using ai-founder-playbook. The detailed workflow should live in this skill, not inside the automation prompt. In Codex, recurring full discovery runs should be cron workspace jobs so each run starts with a fresh execution context. Use thread heartbeats only for thin reminders or controllers.