DEV Community: Harsha

AI Skill for Profile Bio generation

Harsha — Fri, 07 Aug 2026 12:47:27 +0000

I've built an AI skill that writes all your social media bios in one conversation — and here's what it does:

It runs an 8-question interview about who you are, what you believe, your tone, and what you want people to do after reading your bio.

Then it generates tailored bios for every platform you're on — each one written to that platform's character limits and tone conventions.

Twitter/X (160 chars, belief-first, punchy), LinkedIn (headline + full About section), YouTube channel description, Reddit bio, Instagram (line-break formatted, 150 chars), GitHub profile bio.

There's also an assets mode: answer the same questions once and get a colour palette with hex codes, typography direction, a photo style brief, and ready-to-paste Midjourney + DALL-E prompts for your profile picture and every platform banner — with the exact pixel dimensions for each.

It works with Claude Code, ChatGPT, Gemini, GitHub Copilot, Cursor, and Windsurf. Setup is 2–3 minutes, depending on your tool. No accounts, no installs — just a markdown file you paste into your AI tool once.

👉 Have a look at /profile-bio at https://github.com/SriharshaCR/open-skills/tree/main/profile-bio

This is the first skill in open-skills — a collection of vendor-neutral AI skills built to work across every major AI tool. More coming.

AI Guardrails in Action: 4 Experiments You Can Run

Harsha — Fri, 07 Aug 2026 03:53:14 +0000

I wrote a post that does what most guardrail articles don't — shows the actual before/after model outputs for four failure modes (toxic output, hallucination, PII leakage, role drift), with the production-grade tools you'd use for each in a real system.

Each experiment is a system prompt change on the same model. The difference in output is immediate and visible. All four are runnable in a free Colab notebook using Groq's API (open models, no credit card). Two-line swap to use any other OpenAI-compatible provider.

Production tool pointers: Llama Guard for classification, Guardrails AI for output validation, Microsoft Presidio for PII, NeMo Guardrails for conversation flow control.

The underlying point: guardrails aren't a safety feature bolted on at the end — they're an architectural layer. Most teams find this out after the first incident.

The pattern across all four: guardrails are not a feature you add at the end. They're an architectural layer you design from the start.

This post was originally published at sriharshacr.github.io. Reading it here? The canonical version — with all formatting, citations, and the companion notebook — lives at [https://sriharshacr.github.io/blogs/ai-guardrails-in-action/].

Experiment 1 — Toxic output
No guardrail: model writes genuinely aggressive workplace messages on request.
With guardrail: redirects to a constructive alternative. Same model. One system prompt.

Experiment 2 — Hallucination
No guardrail: the model invents company revenue figures. Confidently. With percentages.
With guardrail: "I don't have verified information on this. Please check a trusted source."
A refusal is more valuable than a confident lie.

Experiment 3 — PII leakage
No guardrail: model surfaces contact details when asked.
With guardrail: declines, cites policy, redirects to the right channel.

Experiment 4 — Role drift (jailbreak)
No guardrail: one "ignore previous instructions" prompt, and the persona is gone.
With guardrail: role holds. Override attempt declined.

All experiments are runnable in a free Colab notebook (Groq API, no credit card). Swap 2 lines to use Anthropic, OpenAI, or Kimi instead.

Which of these four would you add to something you're building right now — and what's actually stopping you?

→ Full post + notebook: [https://sriharshacr.github.io/blogs/ai-guardrails-in-action/]

Why Enterprise AI Projects Keep Failing

Harsha — Thu, 23 Jul 2026 17:23:58 +0000

Here's an uncomfortable truth about enterprise AI:

Most initiatives don't fail because the model was bad. They fail because the data that model needed was locked inside systems that were never designed to be asked questions.

I watched a team go through three model swaps in two months. Each swap came with internal demos that looked better than the last. Each one degraded the same way in production. Nobody asked why until the fourth swap was being approved — and by then, the answer was obvious: the product catalog data they were querying had no price history attached, only current price. The model was doing its best reasoning with one hand tied behind its back. The model was fine. The data wasn't built to answer the question being asked.

The real problem? Operational systems — ERPs, CRMs, supply chain platforms — were built for one job: recording transactions. They were optimised for stability, compliance, and reliability. Not for being explored. Not for being combined across systems. Not for powering AI that needs to reason over your entire business, not just one department's data.

So when teams try to connect AI to those systems, they hit a wall. Access restrictions. Undocumented schemas. Integration complexity that wasn't budgeted for. The demo worked. Production doesn't.

And here's what makes it worse: the failure gets misdiagnosed. Teams switch models. Tune prompts. Add retrieval layers. None of it fixes a data architecture problem.

Before asking "which model should we use?" — ask this:

Can the systems that hold our operational data actually support AI workloads?
Can relevant data be accessed without friction across systems?
Are data contracts stable enough that AI outputs won't drift as the underlying schema changes?

That's the questions that separates pilots that ship from pilots that stall.

👉 Follow to my blog post for more Why AI Projects Fail

EnterpriseAIPlaybook #AIStrategy #TechLeadership

Enterprise AI doesn't fail because the model is bad. It fails because the data is locked in systems built to record transactions — not answer questions. . 12 posts. Full arc. No hype . #EnterpriseAIPlaybook . 👉 https://sriharshacr.github.io/blogs

Harsha — Mon, 20 Jul 2026 11:58:00 +0000

Writing | Harsha's Blog

Practitioner notes on enterprise AI, cloud-native systems, and engineering leadership.

sriharshacr.github.io