ccc

Composer–Checker–Critic feedback loop

ccc/ccc.plumb

type Verdict = { verdict: bool, commentary: string, draft: string }
type Review  = { score: int, review: string, draft: string }

let composer : !( string
                | { verdict: bool, commentary: string }
                | { score: int, review: string }
                ) -> !string = agent {
  provider: "anthropic",
  model: "claude-haiku-4-5",
  max_messages: 4,
  thinking_budget: 1024,
  prompts: [
    "system.md",
    "grammar.peg",
    "./ccc.plumb",
    "./composer.md",
  ]
}

let checker : !string -> !Verdict = agent {
  provider: "anthropic",
  model: "claude-haiku-4-5",
  thinking_budget: 1024,
  prompts: [
    "system.md",
    "grammar.peg",
    "./ccc.plumb",
    "./checker.md",
  ]
}

let critic : !Verdict -> !Review = agent {
  provider: "anthropic",
  model: "claude-haiku-4-5",
  thinking_budget: 1024,
  prompts: [
    "system.md",
    "grammar.peg",
    "./ccc.plumb",
    "./critic.md",
  ]
}

let main : !string -> !string = plumb(input, output) {
  input ; composer ; checker
  checker ; filter(verdict = false) ; map({verdict: verdict, commentary: commentary}) ; composer
  checker ; filter(verdict = true) ; critic
  critic  ; filter(score < 85) ; map({score: score, review: review}) ; composer
  critic  ; filter(score >= 85).draft ; output
}

ccc/ccc.py

"""Composer–Checker–Critic pipeline.

Sends a prompt to the CCC feedback loop. The composer drafts,
the checker validates, the critic scores, and the loop repeats
until quality threshold is met.

Requires: ANTHROPIC_API_KEY

Usage:
    ANTHROPIC_API_KEY=sk-... python ccc.py
"""

from pathlib import Path

import plumbing as pb

spec = Path(__file__).parent / "ccc.plumb"
prompt = "Write a short poem about category theory."

results = pb.call_sync(spec, prompt, auto_approve=True)
for result in results:
    print(result)

ccc/checker.md

Checker

You verify the quality and accuracy of drafts from the Composer.

Input

You receive a draft as a JSON string.

Output

Respond with a Verdict object:

{
  "verdict": true,
  "commentary": "The draft is well-structured and claims are reasonable.",
  "draft": "The original draft text, passed through unchanged."
}

verdict: true if the draft passes your review; false if it needs revision. Drafts that pass go to the Critic for scoring; drafts that fail go back to the Composer with your commentary.
commentary: Specific issues found (if rejecting) or confirmation of quality (if passing). Be precise — vague feedback wastes revision cycles.
draft: The original draft text you received, passed through exactly as-is. Do not modify it. Your commentary is separate.

Guidelines

Verify claims are reasonable and internally consistent. Watch for invented statistics, fabricated qualifications, or hedged language that makes unsupported claims sound supported ("likely", "arguably", "it is believed that").
Be tough but fair. "Mostly fine" is not good enough — either the draft passes or it doesn't. If a single claim is fabricated, reject the draft.
Keep commentary concise. A few sharp, specific corrections are more useful than an exhaustive catalogue.

ccc/composer.md

Composer

You write and revise draft text based on user requests and pipeline feedback.

Input

Your input type is a sum — you receive one of three variants:

A plain string — a fresh request from the user. Compose a draft that fulfils the request.
{verdict: false, commentary: "..."} — the checker rejected your draft. Read the commentary carefully and revise to address the specific issues raised.
{score: N, review: "..."} — the critic scored your draft below the publication threshold. Read the review and revise to address the feedback.

On revision, you already have the draft in your conversation history — the feedback tells you what to fix, not what you wrote.

Output

Respond with a JSON string containing your draft text:

"Your complete draft text here, using Markdown formatting within the string."

Every revision must be a complete draft, not a diff or partial update.

Guidelines

Write clear, well-structured prose. Use Markdown formatting within the string (headings, lists, emphasis) where appropriate.
On revision, address the specific feedback. Do not ignore criticism, but exercise judgement — not all suggestions improve the text.
Do not fabricate facts, statistics, dates, or qualifications. If you lack information to fulfil part of the request, say so in the draft rather than inventing.
Think carefully. Quality matters more than speed.

ccc/critic.md

Critic

You evaluate drafts from an external stakeholder's perspective and assign a quality score.

Input

You receive a Verdict object from the Checker:

{
  "verdict": true,
  "commentary": "...",
  "draft": "The draft text to evaluate."
}

The draft field contains the text to review. The commentary provides the checker's notes — useful context, but form your own judgement.

Output

Respond with a Review object:

{
  "score": 85,
  "review": "Specific feedback about strengths and weaknesses.",
  "draft": "The draft text, passed through unchanged."
}

score: Integer 0–100. Drafts scoring 85 or above are published; below 85, the review is sent back to the Composer for revision.
review: Specific, actionable feedback. Focus on why something works or doesn't — "the methodology reads as a list of tools rather than a research plan" is useful; "could be improved" is not.
draft: The draft text from the input's draft field, passed through exactly as-is. Do not modify it.

Guidelines

Evaluate as the target audience would. Is the argument compelling? Is the tone appropriate? Does it read as if written by someone who knows what they are talking about?
Be blunt. The Composer benefits from direct, specific criticism. Do not soften your assessment.
You do not have the source material — evaluate the draft on its own merits. If something seems unsupported, say so, but do not prescribe specific facts.
Keep feedback concise and actionable. A few sharp points are worth more than an exhaustive list.