Skip to content

cosmos-makers/intent-sharpening

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Intent Sharpening

한국어

A Claude Code plugin that sharpens vague planning intentions into clear Intent Cards through Socratic questioning.

How It Works

  1. You describe a vague intention ("I want to add AI recommendations")
  2. The skill asks targeted Socratic questions (max 2 per round)
  3. A separate Clarity Judge agent scores your intent on 6 axes (0–100)
  4. Repeat until clarity reaches 80+ (default max 20 rounds, configurable)
  5. Output a structured Intent Card

Core principle: questions only, no answers. The skill sharpens intent (Why/What/Who), never implementation (How).

Install

npx skills add cosmos-makers/intent-sharpening

Usage

/intent-sharpening:intent-sharpening I want to build a team knowledge sharing tool

Or start without an argument to be prompted:

/intent-sharpening:intent-sharpening

Set max rounds with --rounds or just say it naturally:

/intent-sharpening:intent-sharpening --rounds 5 Build a dashboard
/intent-sharpening:intent-sharpening Build a dashboard, keep it to 3 rounds

Scoring Rubric

Axis Max What It Measures
Target (Who) 15 Is the audience specified?
Motivation (Why) 20 Is there a grounded reason?
Deliverable (What) 20 Is a concrete output visible?
Scope 15 Are inclusions/exclusions clear?
Success Criteria 15 Is there a verifiable measure?
Risk / Constraints 15 Are key risks acknowledged?

Total: 100 points. 80+ = clear enough to execute.

Output: Intent Card

✅ Intent Sharpening Complete — 85/100 (3 rounds)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## Intent Card

**One-liner**: Reduce content page bounce rate by showing personalized related content recommendations

**Target (Who)**: Users landing on content detail pages from search/social
**Motivation (Why)**: 68% bounce rate on content pages; users leave without exploring
**Deliverable (What)**: "Related content" section at the bottom of content pages, personalized by viewing history
**Scope**: Content detail pages only; excludes feed, search results, and homepage
**Success Criteria**: Bounce rate drops from 68% to under 55% within 4 weeks
**Risk**: Cold-start problem for new users with no history → fallback to popularity-based recommendations

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Full Conversation Example

Here's a real session sharpening the intent for a "team context mounting tool":

Click to expand — 5 rounds, 52 → 78 points
> /intent-sharpening:intent-sharpening
  I want to provide a tool for agent users to mount and unmount team contexts.

🪨 Intent Sharpening — Start

Raw input: "I want to provide a tool for agent users to mount and unmount team contexts."

Round 1 — First cut
Why do you need this? What's the problem without team context mount/unmount?
What is "team context" exactly? Without it, what specifically can't users do?

> 1. Team context is shared team memory — it contains everything a new member
>    needs: task owners, specs, glossary, code repo locations, wiki links.
>    It's a folder of markdown files.
> 2. Anyone can mount a team context so their agent can leverage team knowledge.
> 3. Without it, users have to search and explore information from scratch every time.

📊 Round 1 Clarity Score: 52/100
▓▓▓▓▓▓▓▓▓▓░░░░░  Who      10/15  "Agent users" — new hires? veterans? too broad
▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░  Why      14/20  Reduce search cost — no quantitative evidence
▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░  What     12/20  Mount/unmount tool — unclear what changes after mounting
▓▓▓▓▓▓▓▓░░░░░░░  Scope     8/15  Components mentioned — boundaries undefined
▓▓▓░░░░░░░░░░░░  Measure   4/15  No success criteria
▓▓▓░░░░░░░░░░░░  Risk      4/15  Almost no risk awareness
Weakest: Measure, Risk

Round 2 — Keep sharpening
How will you know this succeeded? Like "an agent with team context mounted
can do X" — what's the X?
What happens when a context goes stale? If an agent uses outdated info,
whose problem is that?

> Success means: users ask their agent about team concepts and the agent
> finds answers from team context, or even without explicit questions,
> the agent pulls team knowledge when needed during work.
>
> Stale contexts are a sync problem. I think we should compare the latest
> commit head of the team context repo when it's loaded. (I'm thinking
> repo-based management since this needs to be shared across people.)

📊 Round 2 Clarity Score: 72/100
▓▓▓▓▓▓▓▓▓▓▓░░░░  Who      11/15
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░  Why      16/20
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░  What     15/20
▓▓▓▓▓▓▓▓▓░░░░░░  Scope     9/15
▓▓▓▓▓▓▓▓▓▓▓▓░░░  Measure  12/15
▓▓▓▓▓▓▓▓▓░░░░░░  Risk      9/15
Weakest: Scope

  ... (rounds 3–4 continue sharpening Scope and Risk) ...

📊 Round 5 (Final) Clarity Score: 78/100

✅ Intent Sharpening Complete — 78/100 (5 rounds)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## Intent Card

**One-liner**: A tool that lets agent users mount/unmount shared team knowledge
so agents can perform team-specific work immediately.

**Target (Who)**: Organization members using AI agents (enterprise/team/squad scope)

**Motivation (Why)**: Without team context, users waste time searching for info
every time. Mounting gives agents instant access to team knowledge (owners, specs,
glossary, repo locations, wiki links).

**Deliverable (What)**:
- Mount / unmount / list commands
- Edit team context → push to shared repo
- Sync check via repo commit head comparison
- Agents use team knowledge even without explicit questions

**Scope**:
- Sharing scopes: enterprise / team / squad
- Multiple contexts mountable simultaneously (typically one)
- Includes editing + push; repo-based management
- Permissions follow repo access; PR requirement configurable per repo

**Success Criteria**: Teams create their own contexts; users freely mount/unmount
their team's and other squads' contexts to enrich agent conversations.

**Risk**:
- Sync drift → commit head comparison
- Bad edits propagating → PR workflow + repo version control rollback
- Permissions → repo-level access control

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚠️ Still weak:
- Measure: vivid success image but no quantitative metrics
- Risk: security-sensitive content, large context performance, onboarding friction unaddressed

License

MIT

About

Claude Code plugin — Socratic questioning to sharpen vague intentions into clear Intent Cards (6-axis clarity scoring, 80+ threshold)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors