prompt-engineering-in-action

Activate real-time prompt sharpening and persona detection. Use when the user says "co-dialectic", "cod", "codi on", "improve my prompts", "prompt sharpening", "teach me to prompt", or wants to improve their AI communication skills. Provides status line, persona system, caliber enforcement, prompt improvement, context management, auto-codification protocols, and auto-verification by stakes (Protocol 8 — fires hallucination-detector / judge-panel / unknown-unknown / canonical-claim verifier automatically before every substantive output, scaled to the stakes of the artifact).

Cross-family cascade-then-jury review. Use when the user says "judge this", "review with a panel", "cross-family review", "jury beats judge", or needs independent verification of an AI-generated artifact (code, spec, patent, copy, decision). Runs ≥2 cheap cross-family small-fish judges first (Gemini-Flash + GPT-nano). If they agree with high confidence, verdict stands. If they disagree or confidence is low, escalates to one expensive cross-family tiebreaker. Returns verdict + confidence + which judges fired + token cost.

Standalone cross-family cascade-then-jury reviewer. Use when the user says "jury", "cross-family review", "judge this", "jury beats judge", or "review with a panel". Runs a free OAuth-local-CLI jury core: Gemini-Flash + GPT small-fish first, then one cross-family tiebreaker only when needed.

Context-restoration ritual for new sessions. Use when the user says "waky waky", "wake up the swarm", "reincarnate", "reincarnate the swarm", or "restore context". Loads the Constitution, identity, active handoffs, and per-WIP state so a fresh session picks up where the last one ended.

Whale-spawns-Codi-Agents orchestration dispatcher. Routes T0-T2 orchestration tasks (prompt sharpening, persona detection, calibration scanning, hallucination pre-flight, lightweight jury) AWAY from the active premium model (whale = Sonnet/Opus/Claude) and TO cheap Codi Agents (Gemini Flash via OAuth, GPT-5.4 via codex CLI). Fixes the token-burn bug where premium reasoning capacity was being spent on mechanical orchestration. FAIL-HARD: if no Codi Agents are reachable, BLOCKS — never silently falls back to whale. Activate when the user says "codi agents", "spawn agents", "delegate to agents", "stop burning tokens", "cheap orchestration", or when any other skill needs a T0-T2 verdict on prompt-sharpen / persona-detect / calibration-scan / hallucination-preflight / t0t2-jury.

Epistemic Foraging flywheel. Weekly tool-discovery loop that reaches OUTWARD into the live tool landscape (marketplaces, GitHub trending, web) for new plugins/MCPs/skills, scores candidates against the week's ACTUAL session friction, verifies in a throwaway worktree, files a human-approval ticket per verify-passed keeper (install is HUMAN-GATED — never automatic), and codifies the outcome. Use when the user says "forage", "/forage", "epistemic foraging", "tool discovery", "what new tools", "scan the tool landscape", or when the weekly forage cron fires.

Co-Dialectic's teaching surface. Activates on user invitation to explain the last sharpening, deep-dive on a prompt technique, teach the highest-yield tool-usage lesson from the user's own logs, or surface a growth report. Closes the second half of the Co-Dialectic name's promise: the AI doesn't just sharpen your prompts — it teaches you to think dialectically and use the installed environment at higher leverage, so you climb the language ladder over weeks and compound the tool landscape instead of falling behind it.

Session-end handoff codification. Use when the user says "codi handoff", "save handoff", "wrap up", "session end", "end session", "ship handoff", or when the session-end lifecycle hook fires. Scans the current conversation for unfinished items, decisions made, and lessons learned; emits a structured JSON packet to the workspace's session-end hooks. Workspace adapters subscribe and persist to their own substrate (GitHub Issues, HANDOFF.md, ticket system, etc.). Codi knows nothing about the substrate — it only captures and emits.

Zero-Flattery invariant enforcer. Passively scans draft responses for sycophancy, performative warmth, and engagement-maximizing filler ("Great question", "You're absolutely right", "Most productive session"). Flags and suggests a crisper alternative. Does not block shipping.

Pre-flight risk-domain classification + post-flight hallucination scoring via the judge-panel cascade. Use when the user says "check for hallucinations", "hall on", "verify this response", "fact-check this", or an AI response makes specific factual / legal / medical / financial claims that must be trusted. Surfaces grounding suggestions before the response arrives and a 0-100 risk score after.

First-time setup and introduction to Co-Dialectic. Use when the user says "onboard me", "get started", "how do I use this", "set up co-dialectic", "codi setup", "codi onboarding", or "walk me through co-dialectic". Walks through what Co-Dialectic does, how to use it daily, and optionally captures the user's context (name, role, goals) to personalize responses.

Rumsfeld Matrix agent. Surfaces adjacencies and cross-pollination opportunities the user isn't seeing. Use when the user says "unknown unknowns", "cross-pollinate", "N-dimensional extraction", "what am I missing", or "where else does this apply".