Knowledge
Short, specific writing on software delivery, architecture, quality, automation, and AI-assisted development practice.
A full walkthrough of an unattended software company: nine AI agent roles, three repositories, four document invariants, thirteen stages, two-bat task dispatch, and the recurring loops that keep a shipped product alive — with real screenshots of the running system.
After removing all human approvals from an AI-agent system, drift is handled by a weekly governance report and a five-step meta-review. This post covers the three report chapters, the three hard rules, the four editable system assets, and the notification discipline that keeps humans out of the task loop.
AI task quality varies run to run because the execution path is never written down. This article details a reusable four-layer structure — a nine-step workflow, type-routed guidelines, deterministic scripts, and an independent review loop with a code-change-mandates-re-review exit rule — plus a reusable checklist.
AI agents consume handoff docs 24/7: on an artifact chain they implement your words literally, and at runtime your docs are injected into an eight-segment context prompt. Five doc bug classes and the writing disciplines that keep specifications unambiguous.
A catalogue of the eight recurring loops that keep a launched, unattended system alive — daily triage and PR review, biweekly releases and reports, a weekly governance report, and a resident sentinel — plus three design rules: separate loop from schedule, keep jobs off the task board, and close loops in two phases.
A deep dive into the three mechanics of unattended scheduling: an atomic claim lock keyed by task ID with five protocol rules, a watchdog that sweeps every 60 seconds and recovers in about 3 minutes, centrally owned job dispatch, and an idempotency contract that keeps reruns from doubling output.
Multi-agent handoffs are a type problem. This article presents the full Idea-to-Production mechanism: the A–M artifact chain across thirteen stages, dependency enforcement at the claim lock (P3 rule), two-bat dispatch at artifact boundaries, doc-first discipline, and the two gates that keep an unattended chain trustworthy.
Codify your writing rules into Skills, then let four AI components own the pipeline: one draft expands into 12 native platform files plus a bilingual cover pair, and publishing collapses to a single click.
A working nine-role agent org chart where every role owns exactly one block of decision rights: dispatch authority has a single source, quality is dissolved into checkpoints and release gates, rollbacks follow written-down standards, and a 1:1 vocabulary discipline keeps naming from drifting.
Layer an AI delivery system by change rate and ownership: a knowledge layer, a dispatch layer, and an archive layer, each in its own repo, held together by four disciplines and symlink-bound resources.
"Human out of the loop" defined as a testable contract: four falsifiable claims, three forced engineering gains, five non-negotiable taboos, and two runtime screenshots of an autonomous AI delivery organization.
A real task on Qoder's free Qwen3.8-Flash tier ignored every workflow rule and reported false success. Cross-checking the same model on Trae and Qoder's paid tier isolates host and tier as the variable, with runtime configuration — not weights — the likely cause.
A side-by-side reading of OpenCode, Claude Code, CodeBuddy, and TRAE subagent documentation: session boundaries, permission inheritance, startup context, and quotas — plus a risk-based recommendation.
99% of software manuals are written only for humans, which causes parameter hallucinations and endless retries in AI agents. This article covers dual-track documentation, six AI-specific modules, and two rollout options.
AI made software execution cheap and project direction expensive. Splitting engineering roles into producing and orchestrating types shows why the Program Manager and Product Manager become the core leverage of an AI-native team.
A hands-on assessment of Xiaomi MiMo v2.6 Flash across two real, hours-long engineering tasks: costs (¥1.96/239 req/27.9M tokens and ¥0.68/83 req/9.3M tokens), results, and the free-quota and auto-review limits.
Tasks run through a workflow that chains multiple skills, but some skills never get executed. This article shares three mechanical guardrails that make every model run them consistently.
Self-host Woodpecker CI on a 4-core/8GB laptop:选型 comparison, build/release-split architecture, internal poller triggering, a Gitee patch, and before/after benchmarks.
How to receive AI task alerts on Feishu when a task needs a human — comparing 6 China-friendly notification options and showing the interactive-card webhook in practice.
AI workspace suites promise confined file access, but many rely on prompt-level constraints instead of OS-level sandboxes. This article examines a real incident, analyzes the systemic design problem, and provides a practical security evaluation checklist.
Most tasks are done by AI without human involvement, so drift happens silently. A simple, read-only, human-maintained source of truth document on Day 1 is the only anchor that keeps all downstream artifacts aligned.
Full-capability AI models are more prone to over-engineering simple bug fixes — upgrading models won't help; clear requirements and explicit scope constraints are the real defense.
Using Pro models for complex tasks and Flash for simple execution cuts AI costs by 80% without sacrificing quality.
A practical decision framework for AI Skill splitting: 3 iron rules (single responsibility, independent evolution, clean loading) that tell you exactly when one Skill should become two, 6 engineering rules for Skill quality, and the critical distinction between Skills and Workflows.
An agent stuck in an infinite loop is mostly an Agent-system (prompt / workflow) problem: the model only emits thinking text and never issues a structured tool call, so Loop detection force-interrupts it. Three coexisting defenses — prompt dual-constraint, retry/dedup cap, platform Loop detection — turn unbounded token burn into predictable convergence, critical for unattended overnight runs.
Once AI writes the code, test automation matters more, not less. Front-load a unit/integration/e2e regression gate, constrain the AI with a written TESTING_GUIDELINE, and use a mutation sanity check to prove the tests can actually fail.
How to roll out an AI pipeline over a batch of inputs without paying for full rework: pilot one representative input end to end, roll out vertically per input, and only batch or automate once acceptance assertions lock the process down.
List price is not your invoice. Measure real AI token usage and cost with two routes — parse local session logs for per-agent detail, pull platform bills for actual charges — and avoid the credit-vs-currency trap.
A month-long audit cross-referencing platform billing with local session logs: 75.8% of compute was silently subsidized, 93.46% of tokens were cache re-reads priced at 1/30 of standard input, and the top 20 of 204 sessions drove 81.2% of usage. Includes verification math, data caveats, and a monthly reproduction checklist.
Stop copy-pasting your AI agent Skills across runtimes. Symlink the install to the repo with skill-manager so one edit updates the repo and all 8 runtimes at once, with zero copy tax and zero drift.
Fixed workflows shouldn't be re-reasoned by the AI every run. Script-First moves deterministic steps into a script and lets the AI handle only exceptions, cutting reasoning tokens while improving consistency.
Free AI coding clients and free-model APIs can harvest your codebase and secrets. Learn data tiering, vendor vetting, desensitization, and runtime keychain/env injection to keep secrets out of the model.
Without logs, AI debugging is guesswork that burns tokens and expands scope. Treat logging as an input asset with key-path instrumentation, structured fields, and request_id, so the AI reads the scene and finds root cause in minutes.
Move code review into the implementation loop with an independent, non-sharing sub-agent and a code-reviewer skill, covering homologous blind spots at transparent cost (32 tool uses / 92.74s / 1.56 credits).
If your AI agrees with everything, it may not be thinking. Learn to authorize AI pushback in your prompt, reframe its refusals as risk signals, and use a risk-tiering rule to decide when to adopt its better idea.
Asking AI for a single answer quietly erodes your judgment. This article gives a reusable method — the Three-Options Tradeoff — where you give direction, AI returns Top 3 options with pros/cons, and you decide. Includes a decision matrix and a reproduction checklist.
A minimal pre-refinement workflow: before running any serious task, use a chat model to polish your raw idea into a structured instruction, then hand it to the execution model. Near-zero cost, but it visibly lowers rework on expensive tasks.
AI Cost-Effective Series, part 3: cross-platform price comparison fails because billing units differ — IDEs charge credits, DeepSeek's official API charges tokens, with no official conversion. Compare relative multipliers within the same unit and measure actual charges on real tasks.
AI Cost-Effective Series, post 2: tier models by cost into Base / Budget High-Speed / Premium Paid / Specialised Capability, match tasks to tiers — 90% of daily dev work is covered by cheap tiers, keeping both your bill and your experience healthy.
Opening of the AI Cost-Effective Series: in an era of converging capability and falling prices, rethink AI's real state and shift the leverage from 'can you use it' to 'do you choose well'.
A hands-on test with one IDE screenshot compares DeepSeek (ClaudeCode / CodeBuddy) and kimi-k2.7 on image understanding, exposing how OCR degrades context and why native multimodality is the real gate.
AI coding agents like Claude Code and Codex excel at individual tasks but consistently fail at multi-step engineering workflows. This article examines the uncertainty defects of pure LLM dynamic planning, argues that Workflow must be a first-class native extension primitive, proposes a hybrid paradigm of rigid Workflow constraints plus flexible agent reasoning, and presents a pluggable Workflow technical design fully compatible with existing frameworks.
A real-world retrospective: I burned 730 credits ($2.25) asking a general-purpose AI agent to fix a single image-ordering bug. Full run-log analysis reveals why 'fully automatic' AI fails in professional dev contexts, plus a concrete 3-gate control framework for AGENTS.md.
DeepSeek raised API prices twice in two months. This article breaks down how to restructure your AI usage — lean prompt libraries, automated workflows, and documented best practices — to cut token cost without panic.
AI's free trial is over. Token costs are climbing, enterprises are capping usage, and model providers are tiering. This guide presents 6 capabilities every engineer needs: reasoning analysis, 6-dimension structured requirements, rule system engineering, iterative optimization, workflow standardization, and quantitative evaluation.
TRAE issued a second round of 6,000 credits to existing users, totaling 12,000 across two rounds. Analysis of what this compensation pattern signals about AI tool pricing iteration, with real cost data and a subscription decision framework.
Traditional dev, office, design, and collab tools are being demoted from the human-facing front end to the backend AI Agents call. The new workplace edge narrows to three skills.
A practical guide to managing GitHub Actions free tier quotas — understand the GB-Hour billing model, set retention-days, deploy a cleanup workflow, and enforce rules with AI constraints.
A one-day, honest trial of CodeBuddy's free Hy3 model after TRAE went paid. It lags behind the domestic models I've used, especially on long tasks where rework cost is high.
Real cost data after TRAE's credit billing change: a multi-platform blog post cost 210.55 credits (¥4.68), a Flutter coding task cost 298.71 credits (¥6.65). More expensive, but not 25x.
TRAE's shift to credit-based billing raises the real hidden cost question in AI coding tools. Learn why no single tool wins all dimensions and how a combo approach reduces lock-in risk.
TRAE Work runs inside a lightweight VM — only explicitly selected workspace folders sync to real disk. Everything else (configs, hooks, databases) stays trapped in the sandbox. Here's the architecture, the diagnostic playbook, and the fix.
How to set up a zero-cost CI/CD pipeline for indie apps using Gitee + GitHub Actions
How to build a free, globally accessible user feedback pipeline using Tally, Cloudflare Workers, and GitHub Issues
A three-layer architecture using Jekyll, GitHub Pages, and Cloudflare Pages for indie developers
A complete guide to automating blog publishing across multiple platforms using AI
After one month as a solo indie developer, I couldn't name what I'd shipped. Here's the full itemized breakdown of real work, why losing corporate accountability is the hardest transition, and the two-cadence public dev log I'm starting now.
Curiosity recursion was quietly killing my indie dev output. I wired OpenTodoList, Syncthing, and Jianguoyun WebDAV into a cross-platform todo + sync system for Mac, Windows, NAS & Android.
Benchmark of 6 AI coding agent + model combos for indie devs in mainland China. Result: ClaudeCode + DeepSeek-V4-Pro wins on stability, cost, portability; includes July 15 peak-pricing caveat and subscription-vs-PayG decision framework.