← ClaudeAtlas

prompt-engineeringlisted

Prompt engineering patterns for writing agent prompts and skill files — constraint budgets, register shifting, HARD GATE patterns, anti-personas. Use when writing or reviewing agents/*.md or skills/*/SKILL.md. TRIGGER when: agent prompt, skill file, prompt engineering, model-tier compensation, HARD GATE, prompt quality. DO NOT TRIGGER when: user-facing docs, README, CHANGELOG, config files.
akaszubski/autonomous-dev · ★ 29 · AI & Automation · score 68
Install: claude install-skill akaszubski/autonomous-dev
# Prompt Engineering Patterns Compact, actionable patterns for writing agent prompts and skill files. See [docs/PROMPT-ENGINEERING.md](../../docs/PROMPT-ENGINEERING.md) for full rationale and academic citations. --- ## Constraint Budget (MOSAIC finding) | Range | Compliance | |-------|-----------| | 1–6 | Reliable | | 7–15 | Unpredictable | | >15 | Degraded | **Rule**: Max 6 hard constraints per section. Hard gates at the END of sections (Claude recency effect). --- ## Register Shifting Checklist Before finalizing any hard gate or enforcement directive, verify the verbs: - [ ] Tier 1 (deep): `audit`, `verify`, `critically analyse`, `enumerate` — use for HARD GATEs - [ ] Tier 2 (balanced): `review`, `check`, `describe` — use for standard task instructions - [ ] Tier 3 (minimal): `look at`, `mention`, `note` — advisory only; never in enforcement **FORBIDDEN**: Tier 3 verbs in REQUIRED or FORBIDDEN lists. --- ## Persona Anti-Pattern **FORBIDDEN for analytical agents** (reviewers, auditors, analysts, planners): - ❌ `You are a [role]...` persona prefixes - ❌ `Act as an expert in...` framing - ❌ `Imagine you are...` setups **Source**: PRISM (arxiv:2603.18507) — expert personas reduce knowledge-retrieval accuracy for analytical tasks. **Correct pattern**: Role description + behavioral directive. ``` # BAD You are a senior security engineer. Review this code. # GOOD Perform a security audit. Enumerate all CWE categories present. ``` --- ## HARD GATE Pattern Templat