prompt-engineering

Prompt Engineering

A workflow and toolbox for writing prompts that steer modern LLMs reliably — system prompts, slash commands, skills, rules, agent personas, and one-off queries.

Bundled references (load on demand)

• references/snippets.md — read when you need a copy-paste prompt fragment for a recurring concern: anti-overengineering, parallel tool calls, default-to-action / do-not-act, investigate-before-answering, code-review coverage, frontend aesthetics, frontend variety / propose-options, subagent control, persistence across context windows, multi-context state tracking, reversibility/safety, verbosity, long-document grounding, keep-going stop rules, and the GPT-6 Astra steering set. ~25 vetted blocks.
• references/metaprompting.md — read when a prompt under-performs and the cause is unclear: full metaprompting template, filtering procedure for the model's suggestions, when to use vs. skip.
• references/agent-persona.md — read when designing or auditing the system prompt of an autonomous agent: identity, autonomy posture, tool-use rules, editing constraints, full Friendly and Pragmatic personality blocks, final-message format, plan discipline.

Workflow

1. Define success first — write one sentence: what must the output be, and for whom. Define the eval criteria before the prompt. If you can't, the prompt isn't ready.
2. Draft with the minimal components — Objective + Instructions + (if needed) Context, Role, Format, Examples. Add a component only when you can justify it.
3. Build a test set — 3+ diverse inputs: happy path, edge cases, malformed input. Tiny is fine; some signal beats none.
4. Run and grade — code-grade where possible (JSON/regex/Python parse, length, keyword presence), model-grade for quality, human-grade for nuance. Always demand reasoning + score from a model grader, not a bare score.
5. Diagnose before patching — name the cause (ambiguous scope? missing context? conflicting instruction? wrong format?). Fix the cause, not the symptom. When the cause is unclear, run metaprompting (see references/metaprompting.md).
6. Change one thing at a time — otherwise you won't know what worked.
7. Iterate — re-run the same test set, compare versions; stop when stable across all of them, not when "looks good once".
8. Pin and version — for production, pin to a specific model snapshot (e.g., claude-opus-5-5, gpt-6-astra) and re-run evals when upgrading — a model upgrade shifts behaviour silently and your evals are the only signal.