Skip to content
ObsiKO
English
ํ•œ๊ตญ์–ดEnglishๆ—ฅๆœฌ่ชž็น้ซ”ไธญๆ–‡็ฎ€ไฝ“ไธญๆ–‡
Get the app

Prompt Helper ๐Ÿช„

9 min readUpdated 2026-11-01
On this page

"How do I phrase this so the AI gets it?" Just write the way you normally talk. ObsiKO's Prompt Helper attaches your notes as material and completes the request, and you can copy it or try it right away. There are no prompting techniques to learn.

โ† Home: ObsiKO 101 ยท Parent: 16 AI Agent

Where do I open it? ๐Ÿ”ญ

  • Open the right side panel and pick Prompt Helper (โœจ) in the navigation. โ†’ 02 A Tour of the Screen
  • It has three tabs: Write a request, Repeat work and Wiki. It works on web, desktop and mobile.
  • AI buttons appear only when an AI is connected (API key, desktop login or on-device AI). Without one, writing, attaching material, copying and checking all still work.

โœ๏ธ Write a request tab

Write a request tab โ€” the request box and material buttons
โ–ฒ Write what you want, and attach notes as material if you need to.

Your request

  1. Write it the way you talk. For example: Pull only the decisions out of these meeting notes
  2. Attach material if you have any. With ๏ผ‹ Note as material or ๏ผ‹ Selected text, material goes on top and your request goes last, the layout that tends to work best with long material. The total token count is shown. Notes blocked with ai: deny or .obsiko-ai-ignore cannot be attached.
  3. More detail (optional): why you need it, who will read it, what the result should look like, and an example answer you like. Use these only when needed.
  4. Use it. Copy it into any AI app, Add to note, or Save as helper.

With an AI connected, you get three more.

  • โœจ Refine with AI: makes the request clearer without padding it or inventing facts. Unknown details stay as {{blanks}}, and it shows why it changed things and asks only what it really needs. Nothing changes until you press Use this, and Diff shows what changed.
  • Trim it down: removes repetition while keeping every requirement.
  • Try it: gets an answer right in the panel. If it falls short, write what was not quite right and press Fix the request. Press Use as example to keep an answer you like. This uses your connected AI's quota and does not touch tools or files.

Blanks written like {{name}} get their own input automatically. It will not run while a blank is still empty.

Re-check a changed request on other text ๐Ÿ”

When you change a request, this text may get better while something that worked before breaks. Keep the inputs you tried, and you can run them all again each time you change the request. You can skip this for a one-off request.

Re-checks โ€” a new answer and your call for each text
โ–ฒ Read each new answer and decide for yourself whether it is fine.

  1. Keep. Under a Try it answer, press Keep for re-checking. The material, blank values and answer from that run are kept, with nothing new to write. If you have already written what was not quite right, it is kept as one to fix. You can keep up to 5.
  2. Run. After changing the request, press Run all to run them one at a time with the current request. The button shows how many AI calls it makes, and you can Stop at any time. Nothing runs by itself. This uses your connected AI's quota.
  3. Read. Read each new answer and press Looks fine or Needs a fix. Show last answer lets you compare it with the one before. The call is yours; the AI does not score anything.
  4. Fix. Fix the request from these problems gathers what was wrong and rewrites the request. Nothing changes until you press Use this, and if it got worse, press Go back to the last request.

Under Must-haves (optional), set words to include or exclude, a character or line limit, or a shape (list or table). Answers are then checked automatically and only what is broken gets flagged. This makes no AI call, so it costs nothing. Meeting these does not mean the text is good, so still read the answers.

  • Answers that were cut off or empty, and notes that must not be sent, count as Did not run, and that answer is not recorded.
  • For note material, only which note it is is remembered, and the note is read again at run time. A note changed to ai: deny in the meantime is not sent. Selected text and answers stay in your vault's .obsiko/prompt-checks.json until you press ร—.
  • This runs on the AI connected now, exchanging text only. Results may differ when pasted into another AI or in work that uses tools. AI answers vary a little even for the same request, so do not conclude too much from one better answer.

In AI circles, checking results by running set inputs like this is called an eval. You do not need the word to use it.

Check a prompt you wrote

Reads the prompt in your note (or selection) without AI and flags only clear issues. A prompt written well in plain language gets no complaints.

  • Phrases today's AI does not need (think step by step, promised tips, world's best)
  • Piled-up emphasis (MUST, NEVER, CRITICAL several times)
  • Words with no clear standard (appropriately, somehow)
  • Instructions that are only "don't"s, and long material mixed in without any marking

If a flag does not apply, press Leave it. Words inside code, quotes and material do not count as instructions.

Refine with AI ยท Build it yourself

  • Refine with AI: refines or trims the prompt in your note, then lets you copy, insert, replace the selection, save as a new note or compare. Send to AI panel only fills the input box; you press send.
  • Build it yourself (collapsed): inserts context, instructions, input, examples, output format and success criteria sections at the cursor. More sections adds role, goal, audience, tone and constraints. Use only the sections this task needs. You can choose XML or Markdown style, and Insert example fills in real examples by topic.

๐Ÿ” Repeat work tab

This is where you keep the things you ask for often. The ObsiKO agent picks them up right away.

Repeat work tab โ€” creating a helper
โ–ฒ Save a frequent request as a helper.

  • Create a helper (/command): with a name, description and body, you can call it as /name from the input box. What you add after it goes into $ARGUMENTS. You can also set an argument hint, allowed tools and model. Save as helper on the first tab fills this form: a single blank becomes $ARGUMENTS, and attached notes become reference notes. Nothing is saved until you press Create command. โ†’ 16-1 Basic Usage
    • Reference notes ๐Ÿง‘: pick character, world or style notes and they join the conversation when you call the helper from the AI panel's Helpers button. โ†’ 16 AI Agent
    • Desktop saves to .claude/commands/, while phone and web save to _ObsiKO/commands/ (the folder phone and web chat read). To use a desktop command on your phone too, keep the same content at _ObsiKO/commands/name.md. โ†’ 16-8 Chat-only Mode (Mobile & Web)
  • Always-on rules (AGENTS.md etc.): creates AGENTS.md, CLAUDE.md, GEMINI.md, a system prompt, and the LLM wiki files (soul, user, memory) in your vault. If one already exists, it opens. โ†’ 16-4 Building an LLM Wiki
  • Fill in blanks {{ }}: save a template such as Write {{person}} an email about {{topic}} and change only the values.
  • Create a how-to (skill) ๐Ÿงฉ: creates .claude/skills/<name>/SKILL.md. You call a command yourself, while the agent picks up a skill on its own when the situation fits, so the description must say specifically when to use it (up to 1,024 characters). To use it in other AI tools too, name it with lowercase English letters, numbers and hyphens. Under 500 lines of body is recommended. โœจ Finish with AI only fills the form; you save with Create SKILL.md. In the scaffold's Gotchas section, keep only facts the AI would get wrong without being told.
  • Create subagent ๐Ÿค–: creates .claude/agents/<name>.md (Claude Code compatible). The body is that subagent's system prompt, and the description tells the main agent when to delegate. These files are for external tools that read this folder, such as Claude Code (the agent inside ObsiKO does not hand work to subagents).
  • Agent instruction pieces: inserts scaffolds for tool definitions, output schemas (JSON), guardrails, example pairs, variables, tool use policy, error handling, stop conditions and data blocks at the cursor.

๐Ÿ“š Wiki tab

This is where you use an LLM wiki without memorizing any commands. It only places a request in the AI panel input, and you press send yourself. Nothing costs AI usage until you do.

Wiki tab โ€” status and organizing material
โ–ฒ The app counts your files for the status, and adds a request to organize the material you pick.

  • A vault that is not a wiki: write what you will collect, then press Quick start or Start with questions. To get a wiki someone made, use Import a pack file or Subscribe by link. Starting a new wiki needs a connected AI, and Connect an AI opens the connection options. If the vault already has notes, use Open a new vault and pick an empty folder.
  • Your wiki
    • Status: the app counts 00. Inbox, sources, wiki pages and answers, and shows recently changed pages and recent entries from log.md. Press a number to open that folder in the file list, or a page to open it (recently changed pages are not available on the web version). If a wiki page has broken links, they are listed here, and pressing one opens that spot. Ask to fix requests a fix for just those. None of this uses AI until you send the request.
    • Add and organize material: use Add files to pick files, or paste text or a link and press Add text or link. Files and text are saved to 00. Inbox, and an existing name gets a number instead of being overwritten. A link is not saved; it goes straight into the organize request. You can collect material without a connected AI. Audio, video and large files get a note, but they are not blocked.
    • Pick what you collected and press Ask to organize. Files already marked as ingested are left out, and ai: deny notes cannot be picked. Pages saved with the web clipper show up here too.
    • Ask: write your question and press Add question, or press Ask for a check to have the wiki looked over. Summary, FAQ, Comparison and Timeline make the request in one press. Anything you typed in the box becomes the topic.
    • Rules: press AGENTS.md, Soul.md, user.md or memory.md to open it. The AI reads these every time it works, so the token count is shown too.
    • Share this whole wiki: Send as a file, Create subscription link.
  • A wiki you subscribed to: it cannot be changed, so there is only Add question and Manage subscription.
  • A wiki that stopped midway: if AGENTS.md, index.md or one of the three folders (00. Inbox, 10. Raw Sources, 20. Wiki) is missing, the top of the tab says so. Ask to finish it adds a request to create only what is missing.

Chat AI (on phones, the web and so on) can read and write the wiki only when Allow note editing is on. If it is off, the top of the tab says so and opens that option. It stays on once enabled, so the tab does not turn it on for you. On-device AI cannot read or write files, so the wiki needs another AI.

When to use it ๐ŸŽฏ

  • You do not know how to begin โ†’ write in plain language under Your request, then โœจ Refine with AI
  • You want the AI to answer from your notes โ†’ ๏ผ‹ Note as material, then Copy or Try it
  • The answer falls short โ†’ after Try it, write what was missing and press Fix the request
  • You worry a changed request broke what worked before โ†’ Keep for re-checking, then Run all
  • Your prompt keeps growing โ†’ Check, then Trim it down
  • You ask for the same thing repeatedly โ†’ Save as helper
  • You want consistent rules for the agent โ†’ Always-on rules โ†’ 16-4 Building an LLM Wiki
  • You want to start or run an LLM wiki โ†’ Wiki tab

โ† Previous: Recommended Prompts ยท Home ยท Next: Chat-only Mode (Mobile & Web) โ†’