fidelity-ai-workspace/agent-memory/workflows/prompt-engineering-lessons.md

type, status, updated, tags

type	status	updated	tags
agent-workflow	active	2026-04-21	prompting quality standup workflow

Prompt Engineering Lessons

Goal

Capture reusable prompting lessons so the agent does not need to re-research the same quality patterns for future prompt and command improvements.

---

Stable Lessons

• Start with the smallest prompt that solves the measured failure mode, then add constraints only where they fix a real recurring error.
• Prefer explicit output contracts over vague quality goals.
• Define completion criteria clearly so the model knows what "done" looks like.
• Separate sections by purpose: source selection, decision rules, output rules, and anti-patterns.
• Put high-priority behavioral instructions in the command or prompt that directly controls the output, not in project-facing notes.
• Use concrete anti-patterns when a recurring failure mode is known; models respond better to explicit "do not do X" guidance than to general advice alone.
• Prefer source-bound selection rules when stale or overly broad context can pollute the answer.
• If a task depends on chronology, state chronology explicitly as an output rule.
• If a task depends on dates, inject an explicit temporal context block near the top of the command: current timestamp, today, calendar yesterday, and the default previous workday. Do not rely on the model to infer this from session history.
• Delegate date arithmetic to code or shell commands and feed the result back to the model as data.
• If concision matters, define how to compress: what should be merged, what should remain split, and what should be omitted.
• If the output is meant to be sent directly, make "copy/paste ready" part of the contract.
• Avoid mixing task logic with human-facing project documentation; reusable prompting logic belongs in prompts, commands, skills, or agent memory.

---

Standup-Specific Lessons

• The model should not infer "worked yesterday" from durable status alone.
• The standup command should anchor today, yesterday, and previous workday with absolute dates before reading logs or Mattermost evidence.
• Yesterday should be tied to previous-workday evidence first, then disambiguated with current memory.
• Only active work items should be expanded by default; avoid loading every ticket note when generating short status output.
• Future-sprint work should be excluded from Today unless it is a real blocker or the user explicitly wants forward-looking planning.
• Chronological ordering inside a Jira item reduces awkward or misleading summaries.
• Closely related events should usually be compressed into one concise sub-bullet when they belong to the same continuous investigation.

---

Quality Loop

• Treat prompt improvement as eval-driven iteration: identify the exact bad output pattern, add the smallest correction, and check whether it fixes the failure without bloating the prompt.
• When a correction is about workspace behavior, update the controlling prompt or command immediately so the next run benefits.
• Keep project memory clean while improving prompt quality; do not store agent heuristics in project-knowledge/.