Skip to main content
skill CX Codex

weak-model-protocol — Claude Code skill: features and install

The weak-model-protocol Claude Code skill sets rules for an optimizer driven by a weaker or cheaper language model. It activates when actions are malformed, context runs short, or the provider returns errors.

Apache-2.0 ★ 99 updated 2026-11-12

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add TokenRhythm/InfraRSI --skill weak-model-protocol -a claude-code
Open on GitHub

What weak-model-protocol does

The skill limits the model to one JSON action per turn, chosen from an explicit menu. It puts configuration proposals ahead of patches and validated templates ahead of free-form patches; free patches may change no more than 80 lines. L1 stays locked until enough L0 screens have run or progress stalls, and L2 remains closed.

Only two relevant skill sections and a compact state receipt enter the context. A verifier checks each response against a checklist and can reject it. Rejected replies get at most three retries per turn, and every attempt counts toward the cost.

For APIs, the protocol uses JSON response formatting; vLLM and SGLang use guided decoding. Context is capped at 32K tokens. The served model ID is checked on every reply, with no silent switch to a stronger model.

What it helps with

  • constrain a weaker model to an explicit action menu
  • reject malformed JSON actions before execution
  • keep injected context focused on relevant state
  • track costs for failed calls and retries

Example prompts for the agent

  • Set up weak-model-protocol for a weaker model driving the optimizer.
  • Check that optimizer actions follow the JSON schema.
  • Add bounded retries and account for every model call.

How to install weak-model-protocol

1Skills CLI (Vercel)

npx skills add TokenRhythm/InfraRSI --skill weak-model-protocol -a claude-code

2Manually: put the files in place

~/.claude/skills/weak-model-protocol/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/weak-model-protocol/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about weak-model-protocol

How do I install weak-model-protocol in Claude Code?

Run npx skills add TokenRhythm/InfraRSI --skill weak-model-protocol -a claude-code. The skill file is at ~/.claude/skills/weak-model-protocol/SKILL.md.

What does weak-model-protocol do?

It defines how an optimizer accepts actions from a weaker or cheaper model. A verifier checks replies, while retries and call costs remain visible.

What does weak-model-protocol need to work?

API use relies on JSON response formatting; vLLM and SGLang use guided decoding. The protocol checks the served model ID on each reply.

Source text

SKILL.md · Apache-2.0
---
name: weak-model-protocol
description: Protocol that lets a cheaper LLM drive the optimizer safely - JSON-schema actions, one action per turn from an explicit menu, configs before patches, templates before free patches, small injected context, verifier checklist and bounded retries.
license: Apache-2.0
when_to_use: The driver is a weaker or cheaper model (API or self-hosted); invalid or malformed actions; context overflow; provider errors; 'weak model', 'DeepSeek API driver', 'json schema', 'retries', 'menu'.
metadata:
  author: InfraRSI A0 (human-authored)
  source: human-authored (A0)
  version: 1
  tags: [weak-model, protocol, json, actions, retries, context, verifier]
---

# Weak-model protocol

The same harness guarantees validity for every driver; the protocol only makes a weaker model productive and cheap.
Numerics, budgets and applicability are checked by programs; the model chooses and explains.

## Trigger conditions
- Weak mode (policy.modes.weak), or any driver producing malformed actions, oversize patches or repeated invalid configs.

## Required observations (receipt fields)
- Receipt digests (<= 1.5K tokens) instead of raw receipts; the STATE block (budget, best, L1 gate, templates).

## Rules (FINAL_PLAN 7.2)
1. Actions are JSON constrained by a schema (response_format json_object for APIs, guided decoding for vLLM/SGLang), temperature 0.
2. One action per turn from an explicit menu: inspect / propose_config / apply_patch_template / propose_patch / run_trial / write_memory / finish.
3. Configs first: L1 stays locked until enough L0 screens ran (default 6) or L0 stalls (3 screens without a new best). Validated templates before any free patch; free patches <= 80 changed lines. L2 is closed.
4. Inject only the top-2 relevant skill sections (<= 4K tokens) and a <= 1.5K-token receipt digest.
5. The verifier re-checks with the checklist and may veto.
6. Context <= 32K tokens with a rolling summary; structured state (candidates table) is persistent and re-rendered each turn.
7. At most 3 retries per turn after a rejected reply, each with the validation error fed back; every attempt is charged.
8. Never escalate silently to a stronger model; the served model id is checked on every reply.

## Allowed actions
- The menu above; configs as small deltas (<= 6 keys) on a named base candidate.

## Failure signals
- Same invalid config proposed twice (duplicate detector), schema errors after 3 retries, provider errors (HTTP 429/5xx), context near the limit.

## Verification
- Accounting complete for every call (tokens and cost known, model id matches); JSON actions valid; InvalidAccept = 0.

## Success case (real)
- Closest project record: the 32-GPU first RSI pilot (pilot-1) drove the model with "exactly one JSON action, without markdown" over seven actions (read_source, define_analysis, run_analysis, patch_target, submit_candidate, propose_package, finish); every call was a fresh tools-disabled session and every call's cost was accounted. The protocol worked; the search itself found no speedup (B0 / c1 / retest 10.441060 / 10.448436 / 10.490263 s). (rsi_infra_execution_20261018/rsi_pilot_v1/D1_PILOT_COMPLETION_V1.md)

## Failure case (real)
- The second pilot (pilot-2) driver hit 23 consecutive provider 429 responses (calls 11-33); call 34 only reached STARTED, no session end, selection unfrozen, no retest. The known lower bound of spent cost was 132,677 tokens and USD 2.7731617499999994. Lesson: malformed or failed calls cost budget and must be visible; there is no hidden retry, and an exception without a cost receipt closes accounting. Model self-claims also need program checks: a claim that "the AST blocks in-place mutation" was refuted by dict |= and nested-list +=. (rsi_pilot_v1/D2_SEARCH_EXTERNAL_TERMINATION_V1.md; agentic_moe_sft_tech_report_20261018/RSI_TRACK.md 9)

© InfraRSI A0 (human-authored) · Apache-2.0 · TokenRhythm/InfraRSI

Similar skills

SKILL.md skills catalog →

From the same set19

The whole set →
verifier The verifier Claude Code prompt checks one infrastructure change before it moves to screening. It can veto a candidate based on a checklist or the parent result. GP ChatGPT ★ 99repo stars weak_proposer The weak_proposer Claude Code prompt tunes Megatron-Core training for speed. It proposes config changes, runs trials and identifies the fastest valid candidate. GP ChatGPT ★ 99repo stars reflect The reflect Claude Code prompt revises an optimizer package using the trace from a finished task. It turns observed evidence into rules for later tasks instead of carrying over one task’s settings. GP ChatGPT ★ 99repo stars proposer proposer is a Claude Code and ChatGPT prompt that tunes training throughput in InfraRSI by proposing configuration and overlay changes while preserving fixed task settings. GP ChatGPT ★ 99repo stars optimize-data-path The optimize-data-path Claude Code skill optimizes data and state transfers between the host and GPU without changing the data contract. It is triggered by costly loading, host synchronization, or optimizer-state transfers. CX Codex +1 ★ 99repo stars learn-optimization-skills The learn-optimization-skills Claude Code skill guides an optimizer’s learning between development tasks and records reusable lessons. It runs for post-task reflection or when a search turns up a reusable finding. CX Codex +1 ★ 99repo stars profile-whole-update The profile-whole-update Claude Code skill measures full-update time and checks whether a change improved the end-to-end update. It comes into play when reviewing profiles, comparing candidates, or tracing delays between steps. CX Codex +1 ★ 99repo stars validate-training-lifecycle The validate-training-lifecycle Claude Code skill checks a model-training run from checkpoint save through fresh-process resume, HF export, and forward comparison. CX Codex +1 ★ 99repo stars

Project sponsors

Companies that keep this analytics open