bashout - List of b-logging posts

Prompt-engineering and AI workflows

2026-06-22T00:00:00+00:00

Introduction

I haven't yet entirely come around on the ethics of the tools we are rapidly being exposed to in our day to day, though I certianly feel ass though it's necessary to ride the wave or jump on the bandwagon as it were - to ensure my future is secure.

Prompts

An efficient prompt is one that maximizes the AI’s ability to generate accurate, useful, and contextually relevant responses while minimizing unnecessary token consumption.

Key characteristics of effective prompts include¹²³:

Clarity and Specificity: Prompts should be unambiguous, detailing exact inputs, expected outputs, and constraints. Vague or overly broad prompts lead to irrelevant or generalized responses, increasing token waste and requiring more user intervention.
Structured and Contextual: Breaking complex tasks into smaller, logical steps and providing relevant context (e.g., referencing specific files, previous code snippets, or architectural constraints) improves AI comprehension and response quality.

For example, instead of asking “How do I implement authentication?”, a well-crafted prompt would specify:

Implement OAuth2 authentication for Google and GitHub providers in a Node.js app. Use Passport.js with the following constraints:

- No use of `eval()`; only parameterized queries.
- Follow the conventional commits format for commit messages.
- Add environment variables for OAuth credentials in a `.env` file.

N.B. There's debate surronding the future of prompt engineering⁴ and how necessary it'll be into the future, however right now it has a large impact of the quota usage you'll be subject of.

When to Save Prompts

Prompts should be saved into a prompt library⁵ when they meet the following criteria:

Reusability: The prompt addresses a recurring task or workflow that multiple team members or projects will encounter.
Proven Effectiveness: The prompt has been tested and refined to reliably produce high-quality outputs.
Contextual Relevance: The prompt includes or references specific context (e.g., project files, build commands) that is stable and reusable across sessions.

Skills

Encapsulate a workflow which requires consistent quality and adherence to constraints. Enforce best practices and reduce reliance on individual prompt writers, enabling team-wide reuse and onboarding.

Custom Agents

Custom Agents are suited for tasks that require a separate workspace to avoid cluttering the main chat, such as planning, implementation, or security reviews. They enable parallel execution and are useful when tasks are large, complex, or require independent handling.

Prompts vs. Skills vs. Custom Agents

Feature	Prompts	Skills	Custom Agents
Reusability	Low; typically one-off	High; designed for reuse	Moderate; reusable but task-specific
Structure	Flexible, free-form	Structured, defined inputs/outputs	Structured, role-specific
Execution Context	Runs in current session	Runs within main agent context	Flexible
Use Cases	Exploration, debugging, one-off tasks	Repeated workflows, onboarding	Complex, independent tasks
Maintenance Overhead	Minimal	Requires versioning and updates	Moderate; needs context management
Token Efficiency	Can be inefficient if vague	Optimized for reuse, reduces token load	Can reduce token load by isolating tasks

Strategies for Optimizing Token Usage⁶ ⁷

Token optimization is critical for cost-effective AI usage, especially under usage-based billing models. Key strategies include:

Model Selection: Use lighter models for simple tasks and reserve advanced models for complex reasoning and refactoring.
Separate Planning and Implementation: Plan first with a reasoning model, then implement with a cheaper/faster/lower reasoning model to reduce token consumption.
Context Management: Start new chat sessions for unrelated tasks, use an equivalent context compaction command to summarize long conversations, and exclude irrelevant files from context.
Disable Unused Tools: Prevent unnecessary tool calls that consume tokens without adding value.
Use Tools Over Thought: Try to ensure you're spending tokens on the 'thinking' side of the process or the creative implementation activity. Any repetitive 'deterministic' tasks should be offloaded to programmed solutions/plugins and wrapped in skills the LLM can make use of.
Custom Skills for Preprocessing: Use skills to preprocess prompts, stripping noise and redundancy before sending to the model. There are also tools that may aid in automatically processing prompts which could be implemented.
Monitoring and Debugging: Regularly review token usage via Agent Debug Logs and the Copilot status dashboard to identify and correct inefficient workflows.

Instruction Files

Strong instruction files are critical for guiding AI assistants within coding projects. They act as a centralized source of truth for project-specific information, significantly reducing the need for the AI to hunt through codebases or repeatedly ask for clarification. They provide the following benefits:

Context Preservation: Instruction files provide persistent context across sessions, eliminating the need to re-explain project conventions, coding standards, or architectural decisions.
Token Efficiency: By pre-loading the AI with relevant instructions, you prevent token waste from repetitive explanations or context-seeking questions.
Consistency: Ensures all team members and AI interactions adhere to the same standards without manual reinforcement.
Onboarding Acceleration: New developers (human or AI) can quickly understand project-specific requirements.

Instruction files should be:

Comprehensive: Cover all project-specific conventions (coding style, testing requirements, deployment processes).
Structured: Use clear sections and bullet points for easy parsing by both humans and AI.
Maintainable: Version-controlled and updated alongside the project.

Learning Materials

The citations on this page can act as good reference points for building understanding. However the following certifications/official training can also aid in growing confidence:

There are also many useful references for building skills and agents that appear successful here:

Awesome Copilot - A Repo maintaining many agents, instructions, skills and other LLM goodies

Citations

GitHub. "Prompt engineering for GitHub Copilot." GitHub Docs, 2024. https://docs.github.com/en/copilot/concepts/prompting/prompt-engineering

Beam.ai. "Stop Wasting Prompts: 10 AI Techniques That Actually Work". 2025. https://beam.ai/agentic-insights/stop-wasting-prompts-10-ai-techniques-that-actually-work

Google Cloud. Prompt engineering: overview and guide, 2026. https://cloud.google.com/discover/what-is-prompt-engineering

⁴

MIT Sloan Teaching & Learning Technologies. "Effective Prompts for AI: The Essentials" https://web.archive.org/web/20260904185649/https://mitsloanedtech.mit.edu/ai/basics/effective-prompts/

⁵

Ionos. What is a prompt library? Explanation, benefits, and best practicesk, 2026. https://www.ionos.com/digitalguide/websites/web-development/prompt-library

⁶

Visual Studio Code. Optimize AI credit usage in VS Code, 2026. https://code.visualstudio.com/docs/agents/guides/optimize-usage

⁷

Visual Studio Code. Best practices for using AI in VS Code, 2026. https://code.visualstudio.com/docs/agents/best-practices

Democratisation

2026-02-23T00:00:00+00:00

I've long held the sense that given too much power, the systems we've built cannot prevent the contortion that befalls companies in the endless chase of shareholder appreciation. This is in spite of me being a strong proponent of a capitalistic society, we should ensure it's an ethical society.

Let's talk alternatives

Ideally people should vote with their feet, wallets and headspace. Currently a small minority of corporations dominate the tech space. They take this defacto standardisation as a blank cheque to play with our headspaces, time and in the most sinister of interpretations our futures. (i.e. Cambridge Analytica)

Not to suggest this is implicitly negative. The same centralisation has likely enabled rapid progress in a number of areas (EVs, SpaceX, AI etc) thanks to concentrations of money, brains and power. Potentially I'm simply not a pure technocrat but recent events have demonstrated time and time again that one centralised power does not often lead to the most moralistic decision making process...

Unfortunately people appear to be perpetually lazy or at least ignorant of the consequences to which they are not directly victim to or able to see.

Considering this, I make the simple argument, that to be moral and to maintain said same convenience, doesn't come with great cost, simply a slightly more burdensome upfront cost. Below are my recommendations:

Traditional Software	Alternative Softare
Google Photos	Immich
Netflix	Jellyfin
Spotify/Apple Music	Jellyfin
ChatGPT/Claude/Gemini	Mistral
GSuite/Microsoft Office	Libre Office
Windows/MacOS	Your Linux Distro of Choice (Arch btw)
Android/iOS	GrapheneOS/Ubuntu Touch/PostmarketOS

A Small Furball

Some of this comes alongside the expectation of learning to self host, though I think the real end goal of society is true democratisation of technology, to the extent where all people will own the hardware which their data is being processed upon.

We wouldn't let people snoop in our homes without warrant, so why are we consistently allowing them to snoop on our mental state without the same scrutiny?

State of the Estate 2025

2025-03-06T00:00:00+00:00

I would claim that 2024 can be described as the year of getting back on track. In it, I spent the latter half of the year sober, minimised gaming time, focused on my health and on work; in doing so, I feel I reached the closest state to the "zone" I may ever reach.

BJJ

In March 2024 I made a concerted effort to commit to training Brazilian Jiu-Jitsu once again. Historically being on and off with hobbies I was skeptical, but really hopeful that this may be the time for it to stick. Having found a local, supportive gym to restart that journey on, I trained consistently between 2–5 days per week and I'd even argue it's showing results.

My physique, which historically has been left lacking, has begun to fill out as I've learned to improve what was likely a very meagre diet. I've begun the process of understanding my conflict with migraines — the increased exertion coincided with a spike in occurrences (8 since April 2024). Successes are pending on this front.

I've even dipped my toe in recreational competition, taking part in three different competitions throughout the year. My gi record stands at 4-3, with one silver and two bronze medals. The real positive of each competition has been the takeaways — either the immediate technical lessons or the more metaphorical ones.

BJJ can be thanked for my improved workplace motivation, sleep pattern, eating habits and general physical activity. It's not an activity I can see myself ever dropping, so onwards and upwards. Osu!

Finances

I'm rather financially motivated and don't mind posting a general overview.

Starting the year house-poor, a significant goal was to rapidly build more than a simple emergency fund. Thanks to stable work progress (see the Work section), some savvy decisions and a thrifty mindset, I achieved my year goal of maxing the government-allowed Stocks and Shares ISA limit.

Being a firm believer in putting money to work, my primary investment strategy is long-term investing in an all-world tracker fund. I'll likely provide a greater breakdown of my portfolio in a future post.

Work

I started the year in flux, struggling to find a place on my project at the time (the Home Office). In October the department adopted the PFA system for pay awards, which lit a fire under me. That momentum, plus guidance from a supportive lead who carved out key pieces of work for me and the routine from my improved lifestyle, helped me secure a solid pay raise in January 2025 by being considered a "guru" for a mid-level developer.

Riding that momentum, I pushed for a senior role I had already been recommended for on several occasions. After a laborious application process and an interview where I didn't feel I conveyed myself well, the outcome was highly positive: as of 25 February 2025 I secured the title of Senior Software Developer, specialising in Java.

What's next?

Moving into 2025 off the back of a broadly successful year, I still have numerous goals across life, work and health. I believe these are achievable through the learnings from 2024.

Principles for 2025:

7–8 hours of sleep per night
Work only during the social hours
Moderate or intense exercise 4 days per week

Alongside taking adequate rest, inspired by Jon Gjengset and outlined in his blog "About 40 hours", I hope to continue this routine and make 2025 an even greater success.

References

Jon Gjengset — "About 40 hours": https://thesquareplanet.com/blog/about-40-hours/