AskNow — Turn every Q&A with Claude into a knowledge asset

English · 简体中文

Just chat with Claude. It quietly turns every conversation into searchable, linked, source-traceable Markdown notes. Three months in, you'll have your own second brain — and Claude will actually read it the next time you ask.

---

Why this exists

Everyone has lived this:

• You asked a great technical question last week. This week you can't recall the details.
• The same question gets asked across three different chats — each time from scratch.
• You're researching a topic / reading a book / tracking a domain, and your notes scatter across Notion, Apple Notes, browser bookmarks. Nothing connects.
• AI gives a great answer, but the answer only exists inside that one window. Close the tab and it's gone.

AskNow solves two things:

1. Capture — turn AI output from a one-shot consumable into a compounding asset.
2. Reuse (the bigger one) — make the AI actually read your past research before it answers the next question, so it gives you advice that fits your project, not generic boilerplate.

It is not a new note-taking app and not a chat skin. It's a set of working rules injected into Claude Code: Claude follows them to capture after every answer, and to consult your notes as long-term context on every future turn. Your past benchmarks, the ideas you killed, the calls you made — all retrievable. The knowledge base lives on your disk, opens in any editor, and can sync to Lark/Feishu (Wiki or Docs) for teams.

---

One-line summary

AskNow = Claude Code + 1 orchestrator skill + 3 executor skills (optionally + a CLAUDE.md snippet for always-on mode).

cd into the project, run claude, chat normally. Claude will:

1. Search — scan your existing notes before answering; pull in matching entries.
2. Answer — quality first, with cited sources where it matters.
3. Capture — decide what's worth keeping, then write it as structured Markdown under the right category, maintaining indexes, tags, and a relationship graph.
4. Tail step — auto-sync to Feishu (default on) — if this turn produced new or updated notes, incrementally push them to your chosen Feishu target (Wiki space or Docs folder) and append the links to the final reply.

The flow doesn't interrupt the conversation. The base grows on its own — and the Feishu mirror grows with it.

---

What it looks like

The base lives at the project root, organized by the nature of the knowledge, not by date:

./
├── INDEX.md              # Global index (categories + tags)
├── GRAPH.md              # Mermaid relationship graph
├── TAGS.md               # Controlled vocabulary
├── STATUS.md             # Cross-session counters & queues
├── FEISHU_MAP.md         # Feishu sync map (only when feishu-sync is enabled)
├── concepts/             # WHAT — concepts, definitions, models
├── howto/                # HOW — flows, steps, guides
├── insights/             # WHY — principles, takes, trends
├── references/           # LOOKUP — params, checklists, tables
├── decisions/            # WHICH — comparisons, choices, calls
├── cases/                # CASES — postmortems, debugging, experiments
└── journal/              # Daily Q&A log
    └── archive/          # Quarterly archive

Every note is a self-contained Markdown file with YAML frontmatter:

---
title: Harvey Deep Dive — How a Vertical Legal AI Builds a Moat
category: cases
tags: [ai-startup, vertical, moat, legal-tech]
related:
  - path: concepts/ai-application-moat-defensibility.md
    type: applies
  - path: insights/ai-application-startup-landscape-2026.md
    type: relates-to
created: 2026-05-08
updated: 2026-05-08
importance: high
---

# Harvey Deep Dive

## TL;DR
3–5 sentence summary ...

## Body
Background → product evolution → customers & pricing → moat analysis → takeaways ...

Notes link via [[wiki-link]]s, which auto-feed GRAPH.md to form a graph that's uniquely yours.

---

How it differs from the alternatives

Dimension	Notion / Obsidian	Plain ChatGPT	AskNow
Content production	You type everything	AI writes well, nothing persists	AI captures while answering
Structure	You design folders	None	Forced taxonomy + templates
Linking	Manual	None	AI maintains bidirectional links
Capture depth	Depends on your mood	0	The knowledge-write skill enforces 4-dimension thinking (≥ 3 of 4) + word-count floors
Data ownership	Vendor	OpenAI / Anthropic	Local Markdown — yours
Team sharing	One vault each	One chat each	Auto-sync to Lark/Feishu Wiki or Docs

It is not an Obsidian replacement. If you already use Obsidian, point it at this directory — Obsidian's graph, backlinks, and search will all work. AskNow just keeps writing new notes for you.

---

Quick start

Prerequisites

• Claude Code installed and signed in
• Node.js ≥ 18 (only if you use the one-line install or Feishu sync)
• macOS / Linux / Windows — any
• Optional: lark-cli — only if you want Feishu sync; the first-time wizard installs it for you

One-line install (recommended — works for 54 agent harnesses)

This project is a Claude Code skill pack (1 orchestrator + 3 executors), distributed via the vercel-labs/skills installer. That CLI supports 54 agent harnesses (Claude Code / OpenClaw / Hermes-Agent / Cursor / Codex / OpenCode / Windsurf …). Any directory with .claude/skills/*/SKILL.md works — this repo qualifies natively.

Prerequisite: the repo is already pushed to public GitHub (Carlos-Zen/asknow). npx skills pulls from there via the <owner>/<repo> shorthand. If you fork and want to install from your fork, just swap Carlos-Zen for your GitHub username.

# Claude Code, project-scope (writes to ./.claude/skills/)
npx skills add Carlos-Zen/asknow -p

# Claude Code, global (writes to ~/.claude/skills/)
npx skills add Carlos-Zen/asknow -g

# Multiple agent harnesses at once (auto-detect also works)
npx skills add Carlos-Zen/asknow -a claude-code -a openclaw -a hermes-agent

# Pick specific skills
npx skills add Carlos-Zen/asknow --skill asknow --skill knowledge-write

# Inspect without installing
npx skills add Carlos-Zen/asknow --list

What npx skills does: pulls .claude/skills/*/SKILL.md from the GitHub repo and drops them under the target agent's skill directory. -p → ./<agent>/skills/, -g → ~/<agent>/skills/, -a <agent> declares the target agent.

What npx skills does NOT do: run scripts/bootstrap.sh, bind lark-cli, or edit CLAUDE.md. Bootstrap happens via either of:

1. Automatic (zero-friction): Just open Claude Code in the target project and ask anything → the asknow skill activates → it checks the knowledge-base skeleton under ./ and offers to create it inline (5 meta files + 7 category dirs). No command needed.
2. Manual: Run the bootstrap script once:

   curl -fsSL https://raw.githubusercontent.com/Carlos-Zen/asknow/main/scripts/bootstrap.sh | bash

Feishu sync init does not require an installer — it always kicks off the first time you say "sync to Feishu", via the feishu-sync skill's interactive wizard (pick wiki vs docs → paste a Feishu URL → permission check → write FEISHU_MAP.md). Full script: .claude/skills/feishu-sync/SKILL.md.

Always-on mode (optional): want Claude to auto-run the 4-step flow on every conversation (including auto-sync)? Paste the claude-md-snippet field from skills.json (~5 lines of Markdown) into your target project's CLAUDE.md. Full skill-pack manifest with harness adapters: skills.json.

Manual three-step setup (without an installer)

1. Clone

git clone https://github.com/Carlos-Zen/asknow.git ~/think/asknow
cd ~/think/asknow
bash scripts/bootstrap.sh           # creates the wiki skeleton, idempotent

Or copy CLAUDE.md and .claude/ into any directory you want as a knowledge base.

2. Launch Claude Code

cd ~/think/asknow
claude

Claude Code auto-loads the root CLAUDE.md — from this moment it's in Knowledge Base Builder mode.

3. Just talk to it

Example — researching the AI startup landscape:

"Walk me through the 2026 AI application-layer landscape. Which verticals already produced winners like Harvey or Cursor? What do their moats look like?"

Claude answers (citing reports and public data), then quietly:

1. Decides this is worth capturing → yes
2. Calls the knowledge-write skill
3. Creates insights/ai-application-startup-landscape-2026.md (the landscape)
4. Creates cases/harvey-deep-dive.md, cases/cursor-deep-dive.md (benchmark teardowns)
5. Creates concepts/ai-application-moat-defensibility.md (the concept)
6. Wires applies / relates / contrasts edges between them
7. Updates INDEX.md, STATUS.md, journal/2026-05-09.md
8. (If auto-sync is on) Pushes the new files to your Feishu target and appends the resulting links to its reply

The point comes the next day. You ask: "Based on Harvey and Cursor that I already studied, plus my read on the application layer, design me a cold-start plan for an agentic-commerce product." — Claude pulls those notes as context. The plan extends your prior research instead of starting from zero. That's the whole point.

---

In-conversation commands (all natural language)

Just say it. No slash prefix needed. Skills match these via their description keywords.

Capture control

You say	Effect
"don't capture this" / "skip the capture" / "no need to record"	Don't write this turn
"file under " / "put it in decisions"	Force category
"tag it as X" / "use X as the tag"	Manual tag
"update " / "append to the existing Y file"	Append instead of new file
"mark important" / "high importance"	Mark high importance (frontmatter)

Sync control

You say	Effect
"skip the sync this time" / "don't sync now"	Skip auto-sync this turn
"turn off auto-sync" / "pause auto-sync"	Persist disable
"turn on auto-sync" / "resume auto-sync"	Persist enable
"switch the sync target" / "use another space" / "sync to "	Re-initialize Feishu target (paste URL to auto-parse; say "list" to enumerate spaces)
"sync now" / "push to Feishu"	Trigger full sync immediately

Maintenance

You say	Effect
"maintain the knowledge base" / "tidy up the wiki" / "clean duplicates"	Invoke knowledge-maintenance for full-library scan & repair

---

The four skills (1 orchestrator + 3 executors)

Skills load on demand, not on every turn — keeps token cost low.

0. asknow — orchestrator

Triggers: the user starts a Q&A in a project that already has the knowledge-base structure; or explicitly says "capture", "tidy up", "knowledge base"; or the target project's CLAUDE.md carries the always-on snippet (every turn activates it). Does: drives the 4-step flow each turn (search → answer → capture → tail-sync), decides when to delegate to the three executor skills, and parses user-control phrases in natural language ("skip the capture", "skip the sync", "switch the sync target", etc. — see "In-conversation commands").

Why an orchestrator skill exists: previously all the orchestration logic lived in the root CLAUDE.md. Lifting it into the asknow skill means the whole project can be distributed as a skill pack to OpenClaw / Hermes / any other Claude Code project, without forcing the target to copy CLAUDE.md. The source repo keeps its CLAUDE.md as the always-on activation config.

1. knowledge-write — capture

Triggers: every turn where content is judged worth keeping (delegated by asknow). Does:

• Classify into one of six folders
• Check for duplicates (avoid sprawl)
• Apply a template, write or update
• Maintain bidirectional related: links
• Append to today's journal, bump STATUS.md

Hard constraint: The knowledge-write skill enforces a depth & length floor — concepts notes 800–2000 words, decisions must include a trade-off table, howto must explain why per step, cases must hit the five-section pattern (symptom / investigation / root cause / fix / takeaway). When Claude can't reach the depth, it appends to an existing file rather than creating a thin new one. That's how the base avoids becoming a stream-of-consciousness graveyard.

2. knowledge-maintenance — upkeep

Triggers:

• You ask ("clean up the knowledge base")
• STATUS.md shows ≥ 20 files and last full maintenance > 30 days (auto-detected by asknow)

Does:

• Full-base scan of all .md files
• Detect: dead links, duplicate content, mis-tagged notes, orphan files, naming drift
• Auto-fix where safe / suggest where not
• Rebuild INDEX.md and GRAPH.md
• Emit a maintenance report

3. feishu-sync — push to Lark/Feishu (optional, dual-target)

Triggers:

• You say "sync to Feishu" / "push to wiki" / "push to Feishu docs" / "sync now" → full sync
• You say "switch sync target" / "use another space" / "another folder" → re-run init
• Automatic: any turn that produced .md changes and FEISHU_MAP.md has auto-sync-on-end: true (default) → incremental push (delegated by asknow)

Two sync targets (chosen at first init):

Type	Destination	Best for
wiki	Feishu Wiki — nested nodes per category	Team sharing, deep taxonomy, hierarchy
docs	Feishu Drive folder — docx (rich-text) or .md files	Personal jottings, flat single-doc structure

First-time init (one-shot):

1. Asks whether to use wiki or docs
2. Asks you to paste a Feishu link (/wiki/space/<id>, /wiki/<node>, or /drive/folder/<token> all parse) or reply list to get the available options
3. Validates permissions, writes the target into FEISHU_MAP.md's top-frontmatter
4. Every subsequent sync defaults here; say "switch the sync target" any time to change

Per-turn auto-sync (no manual trigger):

• Reads FEISHU_MAP.md, gathers the .md files touched this turn
• Strips frontmatter, rewrites [[wiki-link]]s as Feishu URLs
• Creates the file if missing (wiki node or docx/.md), overwrites if local is newer
• Appends the resulting Feishu links to the final reply

Good fit for team collaboration / personal backup: everyone keeps a full local base, the Feishu side stays as the shared mirror; auto-sync means you never forget to push, and never have to push manually.

---

Workflow — one full conversation

You:    "Agentic commerce won't take off short term.
         Should we ship a Shopify plug-in / proxy / price-comparison bot
         to bank traffic and data first?"
  │
  ├── ① Claude scans ./decisions/, ./insights/, ./idea.md
  │     → hits idea.md (your "traffic / data / token" framing) → reads it
  │
  ├── ② Claude answers (your prior take + current first-party data)
  │
  ├── ③ Background capture:
  │     ├── Worth writing? ✓
  │     ├── Calls knowledge-write
  │     ├── Updates decisions/ai-startup-ideas-2026.md
  │     │  (appends a trade-off table for four new directions)
  │     ├── Creates cases/ai-wrapper-graveyard.md (anti-patterns)
  │     ├── Wires "contrasts / warns" edges
  │     ├── Backlinks idea.md → ai-startup-ideas-2026
  │     ├── Writes journal/2026-05-09.md
  │     └── STATUS.md counter +2
  │
  └── ④ Tail step: auto-sync to Feishu (default on; skipped if nothing changed or auto-sync off):
        ├── Reads FEISHU_MAP.md frontmatter → target = wiki space "R&D"
        ├── Incrementally pushes the 2 changed .md files
        ├── decisions/...     (update) → https://xxx.feishu.cn/wiki/...
        ├── cases/ai-wrapper-graveyard.md (new) → https://xxx.feishu.cn/wiki/...
        └── Appends the links to the final reply

The interesting part is the next turn. You ask: "Our proxy plan vs Cursor's early pricing strategy — which fits cold-start better?" Claude reads those notes plus cases/cursor-deep-dive.md, answers with real data and your strategic context. It already remembers your project and your thinking — and your teammates see the latest in Feishu.

---

The real point — better suggestions, not just notes

Notes by themselves are nearly worthless. The point is to make AI use your notes to give you a sharper next move.

A vanilla ChatGPT chat starts from zero every time — it doesn't know what you've researched, what your project's about, what you already ruled out. So the advice is generic.

AskNow turns your knowledge base into Claude's long-term working memory. Next time you ask "based on the X, Y, Z I've already studied, give me an ABC plan", it actually reads those files and folds in your research, your judgment, your project stage. The advice comes out only valid for your project.

Three patterns that pay off

1. Cross-note synthesis — distill new plans from your own research

"Combining Harvey, Cursor, and Manus that I've already studied, plus the 'traffic / data / token' framework in idea.md, design me a cold-start path for an agentic-commerce product. Avoid the failure modes in cases/ai-wrapper-graveyard.md."

Claude actually reads those five files, returns a plan that fits your specific context — not a stock template.

2. Decision support — push prior calls into next actions

"We already decided not to die on agentic commerce (see idea.md). Of the four directions in decisions/ai-startup-ideas-2026, which has the best ROI for the next 3 months? Give me an execution checklist."

Claude picks up where your strategic call left off and pushes one level into action.

3. Reflection & blind-spot scan — let AI mirror your own past

"Take every postmortem under cases/ from the last month and extract the failure-pattern signature. Could I be running into one of those right now?"

Claude uses your own counterexamples to check you. Plain chat can't do this.

A metaphor

Treat the knowledge base as the AI's working memory:

	Without it	With it
Start of every chat	Re-onboard an amnesiac	AI has read your past research
Quality of advice	Generic, googleable	Fits your project, judgment, stage
Your role	Endlessly re-feed context	Jump straight to high-level decisions

Capture is the means. Letting every future AI conversation stand on the shoulders of your past thinking is the end.

---

Who this is for

Strongly recommended for:

• Engineers tracking a domain over time (distributed systems, AI, databases, cryptography…)
• Writers — bloggers, newsletter authors, researchers — who need to compound daily thinking
• Consultants / analysts processing high information throughput
• Tech leads turning team Q&A into a Feishu/Confluence-ready wiki
• Students / self-learners stacking up problem solutions, paper notes, interview prep

Probably not for:

• "I just want AI to write code, not accumulate knowledge" → vanilla Claude Code is fine
• Highly sensitive content that must never hit local disk → disable capture
• Tiny note volume (< 5 entries / week) → any note app is enough; automation is overkill

---

Advanced

With Obsidian

Point Obsidian at ./. Claude writes, Obsidian reads:

• Obsidian Graph view = visualization of GRAPH.md
• Obsidian backlinks = the related: field Claude maintains
• Obsidian full-text search + Claude semantic retrieval = dual engine

With Git

git init && git add . && git commit -m "init"

End of each day:

git add . && git commit -m "$(date +%F)"

Now you have a personal Wikipedia with full version history.

With a team Feishu space

First-time setup (one-shot, kicks in when you say "sync to Feishu"):

Claude: "Sync to a Feishu wiki (Wiki space, nested categories) or docs (Drive folder, flat list)?"

You: "wiki"

Claude: "Paste a Feishu link (space or node). Example: https://example.feishu.cn/wiki/space/.... Reply list to list available spaces."

You: paste a space URL

Claude: parses → checks permissions → writes FEISHU_MAP.md frontmatter → locks it as the default target

After that, every turn auto-syncs (default on, dialog-controllable):

When a turn changes .md files, Claude appends a "synced N file(s)" block with Feishu links to its final reply.

Take control any time:

"turn off auto-sync" · "skip the sync this turn" · "switch the sync target to https://xxx.feishu.cn/drive/folder/..." · "push all concepts/ notes tagged importance: high now"

Full command list (all natural-language) is in the "In-conversation commands" section above. No slash prefix needed.

---

FAQ

Q: How is this different from ChatGPT's Memory? A: ChatGPT Memory is opaque, unreadable, non-portable, vendor-decided. AskNow's "memory" is local Markdown — readable, editable, portable, shareable, version-controllable.

Q: Won't Claude pollute the base with random files? A: The asknow / knowledge-write skills define categorization rules, file granularity, depth floors, and dedup logic. The knowledge-maintenance skill cleans up periodically. Worst case: a few extra files — delete them. It's all plain Markdown.

Q: Will it create files for every chat? A: No. The asknow skill explicitly excludes small talk, raw commands, trivial content. Multi-turn conversations append to the same file. Below the depth floor → merge into an existing file rather than create a thin one.

Q: Can I use it with GPT or Gemini? A: The rules in SKILL.md files can port, but the skill system itself is Claude Code-specific. Effect on other AIs is degraded.

Q: What about long-term volume? A: Journals auto-archive quarterly after 90 days. knowledge-maintenance merges duplicates. STATUS.md tracks global state. It scales.

Q: Multi-machine sync? A: Use Git / iCloud / Dropbox / Syncthing — sync the whole directory. Or rely on the Feishu mirror as the shared source of truth.

---

One closing line

AI isn't a tool that thinks for you. It's a tool that amplifies your thinking — but only if your thinking compounds.

You talk to AI every day. Every conversation ends, AI keeps nothing, you keep nothing. That is the most quietly wasteful pattern of this era.

AskNow makes every question lay a brick, and makes every next question stand on the wall you've already built. Three months in, you don't just have a knowledge base — you have an AI collaborator that actually understands your project.

cd into the project, run claude, ask the first question of today.

---

License

MIT