惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Hugging Face - Blog
Hugging Face - Blog
量子位
H
Help Net Security
Microsoft Azure Blog
Microsoft Azure Blog
MongoDB | Blog
MongoDB | Blog
小众软件
小众软件
爱范儿
爱范儿
博客园 - 【当耐特】
Vercel News
Vercel News
S
SegmentFault 最新的问题
M
MIT News - Artificial intelligence
F
Fortinet All Blogs
Apple Machine Learning Research
Apple Machine Learning Research
GbyAI
GbyAI
博客园 - 叶小钗
博客园_首页
V
Visual Studio Blog
宝玉的分享
宝玉的分享
B
Blog
MyScale Blog
MyScale Blog
C
Check Point Blog
博客园 - 三生石上(FineUI控件)
L
LangChain Blog
V
V2EX

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
Quick Win Card #03 — Your sub-agent never read your MEMOR...
Michel Faure · 2026-06-15 · via DEV Community

Quick Win Card #03 strip — Michel writes a brief to his sub-agent, sub-agent codes, Niran reads git log: commit on main, Michel adds a Load-bearing feedbacks section at the top of the brief.

The hook

QW-01 said: your status file lies. QW-02 said: your counter lies. Here's the third layer: your sub-agent doesn't know what you know. Not out of malice — by construction. The context it receives is your brief, full stop. Not your MEMORY.md, not your user-scope feedbacks, not the doctrine you've accumulated over 70 days. You talk to it like a colleague who's read the wiki, when it has just walked in and never opened the page.

What broke

May 18, late afternoon. I delegate six files to my ERP sub-agent — first-contact autosend, ADR-0072. Three-paragraph brief: context, deliverables, phase-0 grep audit before any INSERT. Short, clean, what I'd give a human who knows the project.

Forty-five minutes later, the report lands. I grep my git log out of reflex: commit 3756e63 went straight to main, six files, seven hundred thirty insertions, no feature branch, no pull request. Thirty minutes of cherry-pick to recover. And the rule that should have prevented this — feedback_git_branch_check_avant_commit — was indeed in ~/.claude/agent-memory/, user scope, indexed in my MEMORY.md. I consult it mechanically before every non-trivial commit since the May 14 incident. It didn't betray me. It just never reached the delegate.

The lesson is more uncomfortable than the first two cards. You can build the most rigorous doctrine in the world, index it, version it, backport it across projects — if the brief you send to the sub-agent doesn't cite it explicitly, it is effectively absent. User-scope memory doesn't transmit by inheritance. The sub-agent isn't an heir, it's a subcontractor who hasn't read your wiki.

Fair objection here: GitHub branch protection at the repo level would also catch the commit-on-main case. True for that one example. But the doctrine has to hold even in repos where you haven't wired the protection — and more importantly, the other two feedbacks in the example below (DB probe before DELETE, inscriptions.source audit) have no equivalent server-side guardrail. That's precisely the category the brief needs to cover, because no other layer will.

The brief header that closes the class

Paste this at the top of any sub-agent brief that touches application code, runs longer than thirty minutes, or delegates a decision that could break a silent rule. Three to five load-bearing feedbacks (in Counterpart vocabulary, a feedback = a codified rule extracted from a past incident, stored in MEMORY.md, named with a stable slug).

# Sub-agent brief — <task name>

## Load-bearing feedbacks (apply BEFORE any code)
- `feedback_git_branch_check_avant_commit``git branch --show-current`
  before every non-trivial commit; feature branch + PR for any new feature.
- `feedback_audit_db_pre_flight_canonique``SELECT … GROUP BY` probe
  before any bulk DELETE/UPDATE.
- `feedback_source_audit_inscriptions` — never touch `inscriptions.source`
  rows where source ∉ {NULL, 'migration_notes'}.

## Context
<the brief itself>

Three named feedbacks at the top of the brief. Not the content — just the slug plus one operative sentence. If the sub-agent wants the detail, it opens the file. But the explicit invocation is enough to bring the rule back into its active context. A rule cited by name is a rule considered. A rule that lives in your MEMORY.md and nowhere else is a rule that doesn't exist for your delegate.

ROI

Not time saved on the delegation itself — the brief grows by a few lines, which is neutral. The gain is downstream. Thirty minutes of cherry-pick avoided per commit-on-main mishap. Half a day of DB audit avoided per bulk-DELETE-without-probe. An entire category of silent fail-modes drops out of incident scope because it's named in the brief before code is written. The header doesn't replace the doctrine — it makes it operative for the delegate who has no direct read access to it.

Apply now

Open the next brief you're about to send to a sub-agent. At the top of the file, before the context, paste the ## Load-bearing feedbacks block with three feedbacks relevant to the task. Picking the three is the exercise — it's what makes visible which part of your memory is load-bearing for this specific delegation. If you can't decide which three to cite, your brief is probably too vague to be delegated. In which case the pattern's benefit lands before the sub-agent even starts working: it forced you to re-clarify your intent.

Your quick win takes five minutes — the time to identify the three feedbacks to inline. The doctrine doesn't follow you into delegation on its own. You have to carry it.


Quick Win Card series, episode 03. Counterpart Toolkit v0.7, amendment R9 promoted 20/05/2026. Doctrine repo: github.com/michelfaure/doctrine-counterpart. Triptych QW-01 → QW-02 → QW-03: status file lies, counter lies, delegate's context lies. Three layers of observability that lie by default, three hooks that force them to speak.