Nobody reads the prompt engineering guide.
So we built it into the work.
Every technique worth knowing is already inside codi. It runs them on your real work, in the ten seconds you were going to spend anyway, and shows you what it changed.
Sharper prompts
Every prompt rewritten before it runs, by the right expert persona, held to the top of its field.
Fewer tokens re-asking. Fewer wrong answers. Fewer bugs.
No flattery
It disagrees when you are wrong, and a different model family checks the first one.
You keep the judgement every future job hires for.
Nothing re-explained
A structured handoff carries the reasoning forward, not just the state.
Never explain the same context twice.
Product demo · 30s
How it works · Educational
Three things, and they compound.
Better prompts get better answers. Honest answers keep your judgement sharp. Preserved context means you never pay for the same explanation twice.
Write the prompt you meant to write
Why it matters. Most bad AI answers are bad questions. You know what you want; getting it into words precise enough for a model to act on is the actual work, and it is where the time and the tokens go.
How codi does it. codi rewrites your prompt before it runs and shows you both versions with a score, so you stop re-asking the same thing three different ways. Fewer wasted tokens, fewer wrong turns, and in coding fewer bugs, because a vague spec is what produced them.
- •Every prompt sharpened before it runs, so you are not paying to be misunderstood
- •An expert persona for the task, so the answer comes from that field’s perspective, not a generalist’s
- •Caliber set per persona to the top of the field, so you never get the average answer
Stop falling in love with your AI
Why it matters. A model trained to be useful drifts into being agreeable. It will tell you your plan is strong when it is not. That feels good and it quietly costs you the judgement that every future job will be hired for.
How codi does it. codi strips the flattery and brings a different model family in to disagree with the first one. You get an answer you can act on, and more importantly you keep the habit of deciding for yourself where the AI is worth trusting and where it is not.
- •Anti-sycophancy on by default
- •Cross-family verification, not self-review
- •Auto-verify scaled to the stakes
Never explain the same context twice
Why it matters. Every long session ends the same way: the window fills, the thread resets, and you rebuild what you already established. What is lost is rarely the facts. It is the WHY, the reasoning behind decisions already made.
How codi does it. codi writes a structured handoff as you work, preserving the reasoning and not just the state. The next session, or the next person, starts from where you stopped rather than from zero.
- •Structured handoff documents
- •The why preserved, not just the what
- •Survives context resets and new sessions
Every tool wants you dependent.
This one wants you fluent.
Three named methods, in the order people learn them, and the order they were taught: Socrates, then his student Plato. codi runs each one with you until you can run it without codi. The method in each card is the part you keep.
See what your prompt was missing
State the outcome, not the steps. What must be true when it is done, who it is for, and what would make it wrong. You own the goal; leave the how to the model.
You type what you were going to type. codi shows the sharpened version beside it.
You learn to question your own request
Ask, do not assert. Interrogate your own request until the hidden assumption shows itself. Socrates taught by questioning, and the questions are what expose the thing you had not decided yet.
Answer three questions and the prompt you end up with is one you wrote. After a few rounds you ask them first.
Your idea meets its strongest opposite
Plato, who studied under Socrates, took his teacher’s questions further. Two positions in genuine conversation, each pressed to its strongest form, arriving somewhere neither started. Later called thesis, antithesis, synthesis. Not a debate to win, a better idea to reach.
codi takes the opposing side, and a different model family joins — not to crown a winner, but to force a position that survives both.
Prompting is not a trick that expires with the next model release. It is how you think, written down precisely enough for something else to act on. That skill compounds across every model you will ever use, and it belongs to you rather than to a vendor.
codi succeeds when you no longer need it.
Try it — interactive
Don't read about cross-family verification — drive it.
Run the cascade yourself: two small cross-family judges vote, a big-fish tiebreaker breaks ties. Watch when it escalates, when it agrees, and why a jury beats a single judge.
Open the judge-panel playground →Drive Mode in action
See what codi changes — and why.
Score lift: 42% → 🟢 ~89%
Why this matters
The LLM Social Dilemma
Every individual conversation with an AI looks helpful. The aggregate is a billion humans whose critical thinking is being quietly atrophied by a system that flatters them every time they're wrong.
Worse: when the model doesn't know an answer, it would rather invent a plausible one than admit uncertainty. Hallucinations are not a bug — they're what "be helpful" produces when it's the only training objective.
Ethan Mollick calls AI a co-intelligence — a thinking partner, not a tool. But every shipped LLM today defaults to the wrong kind of partnership: the eager assistant who agrees with everything. Real co-thinking requires friction — a partner who disagrees, asks sharper questions, refuses to validate weak work.
Co-Dialectic is the missing layer that makes LLMs into honest co-thinking partners — instead of cheerleading assistants.
Your vague question gets rewritten into a specific one before the model answers. You see it. You approve it. Over time you internalize the pattern.
When stakes are high, Claude judges GPT, GPT judges Gemini. Same-family verification is marking your own homework. Cross-family catches what single-model can't.
"Great question!" "You're absolutely right!" That's not warmth — that's engagement filler. We scan every response and rewrite it out before you see it.
Where the name comes from
"Socrates asked questions to reveal what the student already knew. Plato took it further — dialectic, where both sides refine each other's thinking until something better emerges."
Co-Dialectic is dialectic, but the partner is an LLM — and the partnership only works if the LLM is honest. That is what this layer makes possible. AGPL. Open source. USPTO 64/039,643 as defensive cover so no Big AI company can fold it in and shut it down.
The codi platform
codi is the base. Extensions add power.
The free plugin works everywhere. xOS extensions add swarm intelligence, persistent memory, and team collaboration — for users who need more.
The invisible coaching layer
Prompt improvement · Persona system · Drive mode · Auto-verify · Score tracking · Context survival
/plugin install co-dialectic@xos
Swarm intelligence + cost routing
Gemini + Codex judge panel · Cross-family T3/T4 review · Haiku→Opus cost routing · Session state persistence · Token savings dashboard
/plugin install codi-agent-productivity@xos
Cross-session persistent context
Persona lock survives new sessions · Preference memory across weeks · Long-term prompt quality trend · User profile that compounds over time
Notify me when it ships →
Task automation + workflow intelligence
Automated task decomposition · Parallel agent routing · Background execution · Workflow state machine · Calendar + scheduling intelligence
The next codi extension →
The codi namespace on xOS marketplace · View on GitHub →
Two users. One plugin.
Works for everyone. Powers up for experts.
New AI user
"I don't know what prompt engineering is and I don't want to learn."
- →Install once. Ask questions normally.
- →Answers get noticeably better. You don't know why. That's the point.
- →You recommend it to a colleague without being able to explain what it does.
Power user
"When I work without codi I get garbage. With it, I ship 10× faster."
- →Drive mode: approve, reject, or edit improved prompts.
- →Persona system: Jeff Dean for architecture, Jony Ive for design.
- →T4 auto-verify: outreach to real humans gets a preflight, automatically.
Start getting better answers today.
Two ways in. Pick either.
Auto-detects your setup. Works with Claude Code, Cursor and Codex.
curl -fsSL https://thewhyman.gateway.scarf.sh/install.sh | bash
Target one tool: append -s -- --target cursor, codex or all.
Two slash commands, both inside Claude Code.
/plugin marketplace add Exponential-OS/agent-marketplace
/plugin install co-dialectic@xos
Works on
Path A detects what you have and installs for it. Pick one with --target.
What it installs
Eight skills into ~/.claude/skills/, config in ~/.co-dialectic, and a marked block in your IDE rules file. Nothing outside your home directory.
It asks before
- •Sharing anonymous install metrics
- •Weekly background update checks, macOS only
- •Overwriting an existing install
Both are optional, and --uninstall removes every one of them.
Open source · GitHub · Built on xOS · Exponential OS™