Dream Claw

Dream Claw is a creative proof loop for solo operators.

It turns a rough paid creative job into a launch pack, finds the weak part, routes the fix to an agent or a human, and saves proof the operator can use to ship the work and win the next client.

Agents do the repeatable work. Humans make the trust calls. AskClaw saves the proof.

Why this matters

AI can make fast drafts. That is not the hard part anymore.

The hard part is:

• knowing what is weak
• knowing what an agent can safely fix
• knowing when a human must step in
• saving proof that the work is solid enough to deliver

Dream Claw is built for that gap.

What this demo proves

This repo shows a small file-based workflow for one paid-work style run:

1. a rough brief comes in
2. an agent makes a first draft
3. a critic scores the draft
4. the weakest area is found
5. the system routes the fix to an agent or a human
6. the result is rescored
7. proof files are saved

The point is not just output. The point is proof of improvement.

Demo in one minute

paid creative brief
  ↓
first draft
  ↓
scorecard
  ↓
weak point: proof
  ↓
agent repair
  ↓
still weak
  ↓
human trust/taste fix
  ↓
rescore
  ↓
proof pack + proof ledger + submission copy

Run the file demo

python3 dream_claw_demo.py

This refreshes the sample proof ledger and verifies the sample artifact set.

Then open these files:

demo/first_draft.md
demo/scorecard_before_after.json
demo/proof_pack.md
demo/proof_ledger.jsonl
demo/submission.md
demo/post_thread.md

What changed in the sample run

In the sample run, the first draft looked useful but the proof was weak.

proof: 5.6 / 10

The agent tried a safe repair, but the score barely moved. Dream Claw then routed the problem to a human because the final promise needed trust and taste judgment.

After the human fix:

proof: 8.4 / 10

That is the core claim of Dream Claw:

It knows when to ask a human, and it saves the proof.

Why this is Hermes-native

This is not one model doing one trick. It is a workflow of:

• roles
• checks
• routing
• artifacts
• proof

That fits Hermes well.

What is real vs simulated

Real in this repo now

• local file-based demo run
• saved first-draft artifact
• saved before/after scorecard
• proof ledger artifact
• proof pack artifact
• human task artifact
• submission copy
• post thread copy
• explicit scoring rules

Not claimed as real yet

• live browser control room
• real multi-user product flow
• automatic client delivery
• production-grade review network
• audit-grade proof system

This repo is a hackathon MVP, not a finished platform. No bullshit.

Repo map

program.md
  the loop, rules, scorecard, and output shape

docs/
  autoresearch_shape_without_gpu.md
  askclaw_hackathon_positioning.md

demo/
  brief.md
  first_draft.md
  scorecard_before_after.json
  workflow_record.md
  proof_pack.md
  proof_ledger.jsonl
  submission.md
  post_thread.md
  human_tasks/
    proof_fix.md

Design shape

Dream Claw is inspired by the loop design of karpathy/autoresearch, adapted for CPU-safe creative workflow proof rather than ML training.

brief
→ draft
→ score
→ critique weak point
→ repair
→ rescore
→ keep or discard
→ log proof

See:

program.md
docs/autoresearch_shape_without_gpu.md

Scorecard

Dream Claw uses simple 0 to 10 scores:

{
  "hook": 8.0,
  "clarity": 8.4,
  "proof": 8.4,
  "look": 7.8,
  "pace": 7.6,
  "readme": 8.5
}

Meaning:

• hook: does it catch attention fast?
• clarity: is the idea easy to understand?
• proof: does it show real work and real value?
• look: is it clean and easy to read?
• pace: does it move well?
• readme: can a judge understand it in under 1 minute?

AskClaw framing

AskClaw is the workflow and proof layer. Dream Claw is one concrete workflow running on it.

The winning frame is:

paid problem
→ workflow
→ agent help
→ human trust check
→ proof pack
→ next sale

Build focus

Do not turn this into a giant platform during the hackathon.

Cut first:

• auth/accounts
• payments
• marketplace ideas
• complex video generation
• GPU-dependent features
• fancy mascot work that does not help the story

A clear proof workflow beats a broken ambitious demo.

Public message

Dream Claw turns a rough paid creative job into a launch pack, finds the weak part, routes the fix to an agent or a human, and saves proof the operator can use to ship and sell.

Short version:

Creative agents need proof, not just output.