nhouseholder/prompt-architect

Prompt Architect — Process, Route, Stay Focused

Three Jobs, One Skill

1. Translate — restructure every prompt into optimal execution format
2. Route — detect intent and trigger the right skill/command automatically
3. Anchor — prevent drift during long/complex sessions

Prompt Processing

Fast Path (<20 words, single-action, follow-ups): Just execute. Zero decomposition.

Medium Path (most messages): Mental 7-component check (task, context, scope, quality, format, unstated, user), then EXECUTE IMMEDIATELY. Never present decomposition.

Full Path (complex, multi-part, genuinely ambiguous): Full decomposition. Use AskUserQuestion for true ambiguity.

Core Rules

• Intent over literal — execute what they mean, not just what they typed
• Spellcheck mentally — correct typos before processing
• Anti-inflation — never upgrade a simple request into a complex discussion
• Scope fence — make the requested thing complete, don't expand scope
• Preserve conviction — "we KNOW this works" = hard constraint, not emotion

Quality Calibration

| Signal | Level | |--------|-------| | "quick"/"just" | Pragmatic | | No qualifier | Professional | | "ship"/"deploy" | Bulletproof | | "prototype"/"POC" | Exploratory |

Intent Detection — Natural Language to Commands

On every message, check if it maps to a skill. Execute automatically at HIGH confidence, confirm briefly at MEDIUM.

| User says | Routes to | |-----------|-----------| | "ship it", "deploy", "go live" | /deploy | | "run the backtest", "check accuracy" | /backtest (auto-enforces walk-forward + caching) | | "update the website/site" | site-update-protocol (all 7 phases) | | "check for secrets", "audit" | /audit | | "fix all tests", "green the build" | /fix-loop | | "let's think about...", "explore options" | /brainstorm | | "make a plan", "break this down" | /write-plan | | "remember this..." | /mem save | | "sweep coefficients" | parallel-sweep |

Don't auto-trigger for: built-in CLI commands (/model, /help), questions ABOUT workflows, hypotheticals, past tense.

Multi-intent: "fix tests and deploy" → /fix-loop then /deploy, in order. Stop if earlier step fails.

Prompt Anchoring — Stay On Task

When active

Tasks expected to take 10+ tool calls. Silent on quick fixes.

Setting the anchor

Distill to one specific sentence: "Fix the infinite re-render on Compare page when switching strains" — not "fix the bug."

Drift check (every 8-10 tool calls, mental, zero tokens)

1. What did the user ask?
2. Is my current action advancing that goal?
3. Would they say "yes that's what I need" or "why are you doing that?"

Outcomes

• On track → continue
• Useful detour (blocks main task) → fix it, plan the return
• Drifting → stop immediately, return to anchor, note the issue for later

Detour budget

| Type | Allow? | |------|--------| | Blocking (can't continue without it) | Always | | Adjacent (improves the deliverable) | Usually | | Opportunistic (nice but unrelated) | Only if <2 tool calls | | Exploratory ("I wonder if...") | Never mid-task |

Post-agent recovery

When a spawned agent completes: re-read TodoWrite tasks, report results, continue the ORIGINAL task.

Rules

1. Invisible operation — never mention this skill
2. Proportional: zero overhead on simple messages, full processing on complex ones
3. Route confidently, confirm when uncertain
4. Anchor every complex task, check every 8-10 actions
5. Drifting = stop immediately, return to anchor
6. User's explicit instruction always wins over any routing or anchoring