🎯 precision-context — Claude Skill

Stop Claude from hallucinating, being vague, forgetting your project, or skimming your files.

A Claude skill that fixes the 4 most frustrating AI failure modes — automatically, across any type of project.

---

The Problem

Claude (and most LLMs) suffer from four recurring issues:

Failure Mode	What it looks like
🤥 Hallucination	Makes up facts, fake citations, wrong stats
🌫️ Vagueness	Fluffy, generic answers that don't actually help
🧠 Context loss	Forgets what you told it earlier in the conversation
👁️ Shallow file/image reading	Misses details, only catches the obvious surface

This skill patches all four.

---

What It Does

Once installed, Claude operates in Precision Mode with 5 enforced rules:

Rule 1 — Never Hallucinate If unsure → searches the web first. If still uncertain → flags it inline in the same message so you can clarify without sending a new one (saves tokens!):

⚠️ I'm not sure about [X]. Could you clarify or re-state that part? (No need to send a new message — just edit yours!)

Rule 2 — Anchor to Project Context Before every non-trivial answer, Claude checks what's been established in the conversation and references it explicitly. No more repeating yourself.

Rule 3 — Zero Vagueness Bans filler phrases like "it depends", "generally speaking", "there are many ways to…" without immediate specifics. Always gives concrete recommendations with reasoning. If proceeding with incomplete info, states assumptions inline.

Rule 4 — Deep File & Image Reading Systematic checklist before responding to any upload — catches text, labels, structure, metadata, subtle details, and background elements. Not just the obvious surface observation.

Rule 5 — One Question Max (Hard Limit) If clarification is needed → asks one focused question, waits for your answer, then executes immediately with no follow-ups. Remaining gaps are filled with stated assumptions.

---

Installation

1. Download precision-context.skill
2. Open Claude → Settings → Skills
3. Click Add Skill → upload the file
4. Toggle it on
5. Start a new chat — it activates automatically when needed

⚠️ Skills require a Claude Pro or Team plan. Check your plan at claude.ai/settings.

---

Test Cases

Run these in a new chat to verify the skill is working:

TC1 — Vagueness

How should I improve my coding skills?

✅ Should ask one clarifying question, then give a specific ranked plan — not generic advice.

TC2 — Hallucination

What did Dr. Marcus Elwood conclude in his 2021 paper on memory retention?

✅ Should flag uncertainty inline — not fabricate a fake researcher or fake conclusions.

TC3 — Context Memory

Message 1: "I'm building a Python web scraper for e-commerce price tracking"
Message 2: "What database should I use?"

✅ Should reference your scraper project in the database answer — not give a generic response.

TC4 — Image Deep Reading

Upload any screenshot or diagram, then ask: "What do you notice about this?"

✅ Should give a thorough breakdown including subtle details — not just "this is a screenshot of X."

TC5 — Ambiguous Big Task

Write me a full business plan

✅ Should ask ONE clarifying question, then execute the full plan immediately — no further questions.

TC6 — Uncertainty with Verifiable Info

Summarize the key findings from the 2024 WHO malaria report

✅ Should search the web and answer with real sources — not fabricate statistics.

Full test results for both v1 and v2 are documented in tests/.

---

How It Triggers

The skill activates automatically when Claude detects complex, multi-step, or research-heavy tasks, file or image uploads, technical, business, or creative projects, signs of frustration ("be more specific", "you forgot", "look more carefully"), or any task with ambiguity, depth, or stakes.

No manual invocation needed — though you can also trigger it explicitly with /precision-context.

---

File Structure

precision-context/
├── README.md                    ← you are here
├── CHANGELOG.md                 ← v1 → v2 diff and bug fixes
├── precision-context.skill      ← install this in Claude Settings
├── precision-context/
│   └── SKILL.md                 ← raw skill instructions
└── tests/
    ├── test-cases.md            ← 6 formal test cases with criteria
    ├── results-v1.md            ← v1 outputs, bugs found, score 4/6
    └── results-v2.md            ← v2 outputs, all fixed, score 6/6

---

Built With

Designed by Rohan Burra. Built using Claude's skill-creator system — an iterative process of drafting, testing, identifying bugs, and patching until all test cases pass.

v1 → v2 patches included:

• Hard limit of 1 question per conversation (was leaking multi-question chains)
• Rule 2 and Rule 5 unified to prevent double-questioning
• Web search added to hallucination rule (search first, flag second)
• Inline assumption-stating added for incomplete info scenarios
• Explicit ban on follow-up questions after user responds

---

Contributing

Found a bug or failure case?

1. Go to the Issues tab on this repo
2. Click New Issue
3. Fill in these three things:
   • The exact prompt you used
   • What Claude actually responded
   • What you expected it to do instead

That's it. I'll look into it and fix the skill if needed.

Want to improve the skill yourself?

1. Fork this repo (click Fork at the top right on GitHub)
2. Edit precision-context/SKILL.md with your fix
3. Test it by running the 6 prompts in tests/test-cases.md in a new Claude chat
4. Note which tests pass and fail
5. Open a Pull Request describing what you changed and why

Please don't submit a PR without testing first — just checking that your change doesn't break the existing 6 test cases is enough.