A Claude Agent Skill that thinks like a senior consultant, not a text rewriter.
npx skills add VivekParmar-18/skillroomYou type a quick prompt:
build login API
…and the model guesses. It picks a stack you didn't want, skips rate-limiting, forgets validation, invents constraints silently, and you spend three follow-ups fixing it.
Good prompts are a skill. Most people don't have time to write a 300-word, role-assigned, constraint-complete prompt for every task — so they don't, and they get mediocre output.
Prompt Intelligence Engine sits between your rough idea and the model. It detects what you're trying to do, adopts the right expert role, reuses context from your conversation, fills the gaps with assumptions it shows you, scores the result, and hands back a prompt that's actually good — then offers to run it.
🧠 Prompt Intelligence — coding · medium
Role: Senior Backend Engineer
Assumptions: [ASSUMED: Spring Boot 3 / Java 17] · [ASSUMED: MySQL] · [ASSUMED: stateless JWT]
Quality: 94/100 (clarity ✓ · role ✓ · constraints ✓ · format ✓ · success-criteria ✓)
─── Optimized Prompt ─────────────────────────
Act as a senior backend engineer. Implement a login API in Spring Boot 3 (Java 17)
backed by MySQL. POST /auth/login → signed JWT. Validate input; BCrypt compare; never
log credentials; rate-limit attempts; consistent error responses; structured logging.
Deliver controller + service + DTOs, security config, and unit tests for success /
bad-password / unknown-user / rate-limit paths. Success: compiles, tests pass, no
plaintext secrets, edge cases handled.
──────────────────────────────────────────────
▸ Run this now? (yes / edit / show your work)
Most prompt improvers just reword your text. This one runs an adaptive pipeline gated by task difficulty, so a tiny ask stays fast and a big one gets the full treatment.
| Typical prompt rewriter | 🧠 Prompt Intelligence Engine | |
|---|---|---|
| Detects intent & picks an expert role | sometimes | ✅ always, automatic |
| Reuses context already in your chat | ❌ | ✅ infers stack/versions/decisions |
| Fills missing constraints | silently guesses | ✅ surfaced as [ASSUMED: …] you can override |
| Quality control | none | ✅ 8-point scorecard, auto-improves below 90/100 |
| Scales to task size | one-size-fits-all | ✅ simple → enterprise (phased decomposition) |
| After optimizing | just text | ✅ offers to run it |
- Intent classification — coding, debugging, research, SQL, AWS, QA, docs, marketing, resume, LinkedIn, career, education, brainstorming.
- Automatic expert role — you never write "act as…".
- Context inference — reuses what you already said in the conversation.
- Hidden-requirement expansion — a login API silently implies auth, validation, rate-limiting, logging, tests.
- Surfaced assumptions — every gap is tagged
[ASSUMED: …], never hidden. - Quality scorecard — transparent 0–100 score; auto-improves weak prompts.
- Difficulty-tiered depth — adds decomposition, self-critique, and compression only when the task is big.
- Optimize → offer to execute — copy it, or run it on the spot.
Via the skills.sh CLI (recommended):
npx skills add VivekParmar-18/skillroomManually (Claude Code / claude.ai):
git clone https://github.com/VivekParmar-18/skillroom.git
cp -r skillroom/skills/prompt-intelligence-engine ~/.claude/skills/Restart Claude, then invoke it (below).
Just ask Claude to optimize a prompt:
- "Optimize this prompt: write a blog about AI"
- "Improve my prompt before running it: fix this SQL query"
- "Make this prompt better: build a login API"
Then reply:
yes→ it runs the optimized promptedit→ correct an assumption, it re-emitsshow your work→ full diagnostics (intent reasoning, prompt AST, critique, token counts)
| Before | After |
|---|---|
write tests |
A senior-QA-framed prompt scoped to the target file, covering happy/edge/failure paths, with deterministic arrange-act-assert and explicit pass criteria. |
make a landing page |
A frontend-engineer prompt with audience, sections, responsive + a11y requirements, framework [ASSUMED], and a clear definition of done. |
build amazon |
An enterprise-tier phased plan (Requirements → Architecture → Data → Backend → Frontend → Testing → Deployment), each phase runnable on its own. |
A tight SKILL.md plus reference files loaded only when the difficulty tier needs them (progressive disclosure — keeps it fast):
skills/prompt-intelligence-engine/
├── SKILL.md entry: triggers, difficulty gate, pipeline, output contract
├── references/
│ ├── pipeline.md full stage instructions
│ ├── domain-playbooks.md intent → role + enhancers + hidden requirements
│ └── scoring.md 8-point lint scorecard, ≥90 gate, compression
└── examples/transformations.md before→after calibration examples
Pipeline order: difficulty → intent → context → role → skill-level → hidden-reqs → (tiered stages) → format → scorecard.
Does it work outside Claude Code?
It's a standard Agent Skill, so it works anywhere the skills.sh CLI installs (Claude Code, Cursor, Codex CLI, and more). The optimized prompt itself is portable to any model.Will it slow down simple prompts?
No — difficulty estimation runs first. "Fix this CSS" gets a light pass; only complex/enterprise prompts trigger decomposition and the critique loop.Does it secretly change what I asked for?
No. It preserves your intent and fills only the gaps — and every gap it fills is shown as an[ASSUMED: …] tag you can override with edit.
Is my data sent anywhere?
No. The skill is plain instructions Claude reads locally — there's no server, no telemetry, no external calls of its own.- More domain playbooks (data science, mobile, game dev, legal)
- Optional "ask me 1 question" mode for high-stakes prompts
- Saved prompt templates / favorites
- Community-contributed playbooks
Have an idea? Open an issue.
PRs welcome — see CONTRIBUTING.md. Good first contributions: new domain playbooks, before/after examples, and FAQ entries.
- 🐛 Bugs / ideas → GitHub Issues
- ⭐ Like it? Star the repo — stars drive discovery on skills.sh and GitHub search.
MIT © Vivek Parmar
If this saves you a few prompt rewrites, give it a ⭐ — it genuinely helps others find it.