Claude Code Goes Mechanical: Deny Rules, Dumb Tests, Pinned Subagents, and Strategic Forgetting

Today's findings

1. #1 Turn transient Claude Code rules into mechanical deny entriestechnique

   Put mid-session constraints into committed .claude/settings.json deny rules so they survive context loss and chat resets.

   Claude Code guardrails

   Turn transient chat rules into committed deny entries

   Said in chat

   • Rule lives in chat
   • “Don’t touch migrations”
   • “No merge before review”
   • Forgotten after reset

   → Enforcement survives context loss and chat resets

   Committed as deny

   • Rule lives in settings.json
   • Edit(migrations/**)
   • Bash(gh pr merge *)
   • Checked before allow

   Deny is checked before allow; expiry stays human-owned.

   Why it matters: Agents forget conversational instructions, but deny entries are checked before allow and persist with the repository.

   How to apply: Convert path-shaped rules like “don’t touch migrations” to Edit(migrations/**), action-based rules like “don’t merge before review” to Bash(gh pr merge *), commit the settings, and keep expiry or removal human-owned.

   claudeagentsguardrails

   Read more: Temporary rules for Claude Code: make the block mechanical and the expiry human

2. #2 Encode old architectural decisions as dumb teststechnique

   Write tiny tests for deliberate constraints so fresh Claude Code sessions cannot “clean up” decisions they do not understand.

   Why it matters: Long-lived repositories contain non-obvious invariants, and tests act as durable memory that agents must respect.

   How to apply: For each historical decision, add a minimal test such as a concurrency cap or no-extrapolation check, include a comment linking the rationale, and run it in CI or the agent loop.

   claudetestingagents

   Read more: What finally stopped Claude Code from undoing my old decisions: turning them into dumb tests

3. #3 Pin subagent models to keep Claude Code costs predictabletip

   Pin every custom subagent to Sonnet or Haiku in frontmatter and choose Opus only per task with a written rule.

   Claude Code cost control

   Pin subagents to stop silent Opus spend

   Pinned subagent

   • model: sonnet or haiku in frontmatter
   • Routine work stays on cheap models
   • Cost per agent stays predictable

   vs

   Inherited model

   • Subagent copies the main-loop model
   • Routine calls quietly run Opus
   • Bills spike with no signal

   Pin every subagent to Sonnet or Haiku; let Opus in only via a written per-task rule.

   Apply: add model: sonnet or model: haiku to each subagent, then audit cost per agent.

   Why it matters: Inherited main-loop models can silently turn routine subagent work into expensive Opus calls.

   How to apply: Add model: sonnet or model: haiku to each subagent definition, keep a short decision rule for when Opus is allowed, and audit cost per agent.

   claudeagentscost

   Read more: How I configure subagents so the cost stays predictable

4. #4 Forget history to explore better: reset agents from the best candidatepaper

   Berkeley’s “Scaling Discovery Requires Forgetting” reports that agents periodically discarding session and search history while keeping the best candidate beat parallel agents by up to 15%.

   Why it matters: Accumulated context can entrench bad paths and waste tokens; disciplined resets can improve discovery quality.

   How to apply: Store the best artifact plus its score, start fresh context from that candidate when progress plateaus, and benchmark the reset policy against a parallel-agent baseline.

   agentscontextpaper

   Read more: Berkeley paper: agents that dropped their own history beat parallel agents by up to 15%. What do you keep when you reset context

5. #5 Check whether your agent security layer actually blocks flagged contenttechnique

   StackOne’s open-source Defender shows detection can flag risky content without blocking unless blockHighRisk is enabled and enforced by the app.

   Why it matters: A dashboard saying “prompt injection detected” creates false assurance if tool results still reach the agent.

   How to apply: Audit guard configuration, force high-risk findings into deny verdicts, test application enforcement rather than logging alone, and define fallback behavior when an optional LLM reviewer refuses.

   agentssecuritymcp

   Read more: How to check whether your AI agent's security guard actually blocks what it flags

6. #6 shura.cpp adds a VRAM governor to llama.cpprepo

   A llama.cpp fork targets sharing one 16 GB GPU between a 35B MoE model and normal desktop use.

   shura.cpp — a llama.cpp fork

   A 35B MoE can share your 16 GB GPU with the desktop — only inside hard limits

   bounded capability

   Run Tiel-Coder-35B-A3B beside everyday desktop apps

   limit Hard 16 GB VRAM cap

   scope Reserved desktop headroom

   monitor Governor enforces cap

   Caps prevent OOM failures and desktop stutter; decode speed vs swapping still needs a head-to-head.

   Why it matters: Local agents often need to coexist with workstation apps, and hard VRAM caps can prevent out-of-memory failures and desktop stutter.

   How to apply: Try shura.cpp’s governor with Tiel-Coder-35B-A3B on a 16 GB card, set explicit headroom, and compare decode speed and swapping against upstream llama.cpp.

   llama.cpplocal-llmvram

   Read more: shura.cpp: a llama.cpp fork with a VRAM governor, so a 35B MoE can share one 16 GB card with the desktop you keep using

7. #7 Hillock v1.0 swaps probabilistic RAG for local neuro-symbolic memoryrepo

   Hillock v1.0 is a local memory engine that claims deterministic retrieval under 1.2 GB VRAM to avoid hallucinated vector hits on unanswerable questions.

   Why it matters: Vector search always returns nearest chunks, which can push small local models into confident hallucinations when the corpus lacks an answer.

   How to apply: Test Hillock with an unanswerable-query eval set and compare latency, VRAM use, and hallucination rate against your current vector RAG stack.

   ragmemorylocal-llm

   Read more: After 6 months of daily building, I just released Hillock v1.0 stable: a local neuro-symbolic memory engine that runs under 1.2GB VRAM with zero hallucinations · After 6 months of daily building, I just released Hillock v1.0 stable: a local neuro-symbolic memory engine that runs under 1.2GB VRAM with zero hallucinations · Vector search is probabilistic and causes hallucinations on unanswerable queries. After 6 months of testing, Hillock v1.0 is out to fix it. · Vector search is probabilistic and causes hallucinations on unanswerable queries. After 6 months of testing, Hillock v1.0 is out to fix it.

8. #8 FlashMap brings flow-map speculative decoding for local inferencepaper

   Princeton and Polymathic AI released FlashMap, a flow-map speculative decoding approach aimed at faster generation.

   Paper release

   Flow-map speculative decoding for faster local LLMs

   FlashMap

   feature Paper

   Princeton + Polymathic AI

   runRead the paper

   Faster local generation without new hardware — port benchmarks pending.

   Why it matters: Speculative decoding is one of the most direct ways to reduce local LLM latency without upgrading hardware.

   How to apply: Read the paper and watch for llama.cpp or vLLM implementations; benchmark acceptance rate, memory overhead, and throughput against standard draft-model speculative decoding.

   local-llminferencespeculative-decodingpaper

   Read more: Princeton and Polymathic AI release FlashMap: Flow Map Speculative Decoding · Princeton and Polymathic AI release FlashMap: Flow Map Speculative Decoding

9. #9 Invisible Unicode can break LLM eval graderstip

   A single invisible Unicode character flipped Claude Code and other grader verdicts; annotating exact character names and positions fixed judgments.

   Eval pitfall

   One invisible character flipped the grader

   1

   zero-width Unicode character — present in the text, invisible to the eye — was enough to flip Claude Code's verdict

   pass → fail

   0 px

   visible width of the culprit glyph

   1 note

   naming each character and its position restored the correct judgment

   Reveal or normalize invisible characters before an LLM grader scores the text.

   Why it matters: Eval scores can be wrong for invisible formatting reasons, making prompt and model comparisons unreliable.

   How to apply: Normalize or reveal zero-width and invisible characters before grading, append a generated note listing each character and position, and add regression cases to the eval set.

   evalsclaudeunicode

   Read more: I tested 3 LLMs as evaluators on an invisible unicode character · Found a weird issue with Claude Code's native eval grader over an invisible space

10. #10 Version prompts like code with small eval setstechnique

    Treat prompts as programs: git-version them, run a 20–50 example eval set before changes, review eval diffs, and log drift.

    Prompt ops

    Treat prompt edits like code changes

    1

    Version

    Git-tag each prompt

    2

    Eval

    Run 20–50 fixed cases

    3

    Review

    Diff eval outputs

    4

    Merge

    Ship only on pass

    5

    Monitor

    Log output drift

    Drift → rollback

    Provider updates can silently change behavior — the 20–50 case eval gate and one-step rollback keep it pinned.

    Why it matters: Prompt edits and provider updates silently alter behavior; eval-backed versioning gives teams reviewability and rollback.

    How to apply: Store prompts in the repo with version tags, maintain real input and expected-output cases, run evals before merge, log outputs for drift, and keep a one-step rollback path.

    promptsevalsproduction

    Read more: How are you versioning and testing prompts in production?

11. #11 Qwen3.8-27B tuning on a 20 GB AMD cardtip

    A tuned llama.cpp/RDNA setup runs Qwen3.8-27B IQ3_S at 60–70 tok/s short and 45–50 tok/s with long history on a 20 GB RX 7900 XT.

    Why it matters: It shows local 27B-class inference is viable on consumer AMD hardware when offload and cache settings avoid RAM spill.

    How to apply: Use an RDNA-boosted llama.cpp build with the community IQ3_S GGUF, avoid RAM spill, and tune context/cache settings against long-history workloads.

    local-llmllama.cppamd

    Read more: Qwen3.8-27B on a 20GB 7900 XT: 60-70 tok/s after two months of fiddling, here's my config