AI coding practices, explained from the inside.
We use Avorelo to build Avorelo. These articles come from real runs, real failures, and real patterns we observed. No invented scenarios. No generic AI productivity advice.
Every Field Note comes from real Avorelo development work. No invented scenarios. No generic AI productivity advice.
Field Notes
Observed patterns from real Avorelo sessionsRepeated context becomes hidden cost
Across long AI coding sessions, the same task context kept getting reintroduced. Avorelo surfaced the pattern and preserved the useful parts for the next run.
You will understand: how context repetition accumulates as invisible token overhead. Read note → ScopeBroad access is the unsafe default
AI coding tools often start with more access than the task requires. Avorelo helped turn task-only access into the default instead of relying on manual cleanup.
You will understand: why broad access is the real risk pattern and how task-scoped access works. Read note → ProofUncaptured proof cannot guide the next run
When proof lives only in a past session, every review becomes reconstruction. Avorelo keeps evidence attached to the work so it can be reused later.
You will understand: what a session receipt is and why proof must be structured to be reusable. Read note → RoutingMost model choices are made too late
Wrong routing often starts as a small context mismatch, not an obvious model mistake. Avorelo uses task signals before more AI spend happens.
You will understand: how early routing signals cut wasted spend before a run starts. Read note → AccessAccess should expire with the work
When a run ends, broad access should not remain behind. Avorelo scopes access to the task and removes it when the work is done.
You will understand: why idle standing access is the unsafe default and how task-expiry works. Read note →AI Waste & Cost
Where AI spend goes without producing valueWhat is AI waste in software development?
AI waste is the gap between what AI coding tools consume and what they produce. Wasted tokens, repeated context, scope drift, and unsupported findings all add up.
You will understand: the four categories of AI waste and which produce the highest overhead. Read article → CostWhy AI coding costs grow faster than teams expect
Token costs compound through repeated context, long windows, wrong model selection, and calls that could have been avoided. Here is where the growth actually comes from.
You will understand: how token cost compounds and which compounding factors are preventable. Read article → MeasurementHow to measure AI coding productivity without fooling yourself
Activity metrics are easy to game and easy to misread. Measure validated outcomes, rework avoided, review load reduced, and proof coverage instead.
You will understand: why outcome metrics with confidence labels beat activity counts. Read article →Scope, Safety & Access
Why bounded agents are safer agentsWhy AI agent approval fatigue makes teams less safe
When every AI action triggers an approval prompt, teams learn to dismiss them quickly. Too many approvals produce a safety system that is technically present but practically ignored.
You will understand: how approval volume inverts safety and what scope-first patterns replace it. Read article → CapabilityWhy AI agents should not see every tool by default
Giving AI agents access to every available tool increases risk surface and token cost. Scoped tool exposure matches what is available to what the task actually needs.
You will understand: how task-scoped capability exposure reduces both risk and token overhead. Read article → ScopeHow to detect AI agent scope drift before it becomes review work
Scope drift turns a small task into a large review. Catch it by comparing planned actions to the request and narrowing before the run continues.
You will understand: how to detect drift from the plan and respond with narrow, block, or escalate. Read article → SafetyHow to let AI coding agents work safely without babysitting them
Safe autonomy comes from scoped access, proof receipts, and cheap rollback, not from supervising every action an agent takes.
You will understand: the safe autonomy model that replaces constant supervision. Read article → SafetySkill governance for AI agents
As AI agents gain skills and tools, governing which ones each task can use becomes a safety question. What skill governance is and why task-scoped access matters.
You will understand: why capability should be granted per task and revoked when the work ends. Read article → SafetyAI skill supply chain security
Third-party AI agent skills and plugins are a supply chain. Why unvetted skills are a security risk and how scoping and provenance reduce the exposure.
You will understand: why a skill is an active dependency and how scope contains the risk. Read article →Proof & Evidence
What makes AI output trustworthyWhy AI security findings need evidence
AI tools surface potential security issues. But plausible findings without evidence create work, not safety. How to tell the difference and what to require before acting.
You will understand: how to distinguish actionable findings from plausible-but-unsupported alerts. Read article → EvidenceHow AI-generated bug reports waste engineering time
AI-generated bug reports often lack reproduction, impact, and a proof path. Why that creates false alarms and how evidence-backed reporting fixes it.
You will understand: what a bug report needs before it earns engineering attention. Read article → EvidenceWhat is AI slop in software security?
AI slop in security is plausible-but-unsupported output: noisy findings, vague risk language, and missing evidence. How to define it and require proof contracts instead.
You will understand: how to name AI slop and replace it with a proof contract. Read article → EvidenceHow to reduce noise from AI code reviews
AI code reviews produce many suggestions but few decisions. How to separate suggestions from evidence-backed findings, filter unsupported claims, and route low-risk work.
You will understand: how to turn a wall of AI suggestions into a short list of decisions. Read article → EvidenceWhy evidence-backed AI coding matters
Evidence-backed AI coding attaches tests, receipts, diffs, and validation to outputs. Why a proof layer reduces review load and makes future sessions faster.
You will understand: what evidence to attach so AI output can be trusted without re-checking. Read article → ProofWhat is an AI coding proof receipt?
A proof receipt records what an AI coding run changed, what it validated, what evidence exists, and what remains uncertain. Why it is reusable context, not a log.
You will understand: what a proof receipt contains and why it beats a commit log. Read article →Context & Continuity
Keeping sessions informed without re-explaining everythingWhat is context engineering in AI coding?
Context engineering is selecting, filtering, and structuring the right information for an AI coding session. It is not about bigger windows or more context. It is about the right context.
You will understand: the five-step context compiler and why selection beats stuffing. Read article → TeamsStop re-explaining your project to AI
A lot of AI waste comes from rediscovery. How to build reusable context so future sessions start faster instead of from zero.
You will understand: how proof receipts carry context forward so you stop starting from scratch. Read article → ContextRAG vs memory vs project knowledge in AI coding
Retrieval, memory, and a project knowledge layer solve different problems. How they differ, where each fits, and why AI coding needs structured project knowledge most.
You will understand: which context mechanism fits which problem, and why they are not interchangeable. Read article → ContextThe risk of stale AI memory
AI memory records what was true when written, but code changes. How stale memory creates confident-but-wrong output and why verification matters more than recall.
You will understand: why a confident-but-stale memory is worse than none, and how to keep it honest. Read article → ContextWhy AI coding needs a project knowledge layer
A project knowledge layer is the validated, reusable account of how a codebase works. Why AI sessions need it and how it differs from docs, memory, and retrieval.
You will understand: why a kept-true knowledge layer compounds while a bigger context window does not. Read article →Teams & Reporting
Measuring AI value without tracking peopleHow to report AI coding value without tracking people
Measuring AI coding outcomes for a team does not require tracking individual developers. Aggregate by workflow, repo, tool, and capability to get useful signals without surveillance.
You will understand: which aggregation levels give signal without exposing individual contributors. Read article → TeamsFrom AI spend to AI proof: how to know what your AI budget actually produced
Prompt counts and token bills do not tell you what AI work actually produced. Here is how to connect spend to evidence-backed outcomes.
You will understand: how to read the AI Spend + Proof report and what validated workflow means. Read article → TeamsAI activity is not productivity: measuring what actually got done
High agent activity and high output are not the same thing. How to separate AI runs that produced value from runs that produced noise.
You will understand: how the What Got Done report separates signal from AI churn. Read article → TeamsScope repair before approval: how to stop babysitting AI agents
Review fatigue happens when every small drift becomes an approval. Avorelo narrows scope first and asks only when real authority is needed.
You will understand: how scope repair reduces approval load without reducing oversight. Read article → TeamsWhat a team AI value report should actually show
Most AI reporting shows spend and activity. What a team AI value report should show instead: validated outcomes, scope safety, and proof, without tracking individuals.
You will understand: why a value report should show outcomes, not effort, aggregated by work not by person. Read article → TeamsHow AI review load gets reduced
AI coding can flood reviewers with changes and findings. How review load actually drops: declared scope, evidence-backed findings, and proof captured before review.
You will understand: why moving scope and proof earlier shrinks the review queue. Read article → TeamsAI team reports for engineering leaders
Engineering leaders need to know what AI coding is producing across a team. Which reports answer that, what each one shows, and how to read them without surveilling people.
You will understand: the five reports that answer a leader's real questions without per-person scoreboards. Read article → TeamsWhat an AI-native engineering team looks like
An AI-native team is not one that uses AI the most. It is one that has built scope, proof, and context discipline around AI so the work stays trustworthy at volume.
You will understand: why AI-native means disciplined, not maximal, and which habits define it. Read article →Workflow Health
Keeping AI workflows clean as they accumulateWhy AI workflows get messy and how to keep them clean
Accumulated instructions, overlapping tools, and stale context make AI work slower and riskier over time. How to track and fix it.
You will understand: the five workflow health signals and which ones predict future drift. Read article → WorkflowAI workflow hygiene: keeping agent setups clean
AI coding setups accumulate instructions, tools, and stale context over time. What workflow hygiene means, the signals that predict drift, and how to keep it clean.
You will understand: why good hygiene is subtraction, and which signals to trim first. Read article → WorkflowWhy AI agents repeat the same mistakes
AI agents make the same mistakes across sessions because corrections are not captured. Why feedback evaporates and how durable receipts stop the repetition.
You will understand: why capturing the why, not just the what, ends the repetition loop. Read article → WorkflowThe hidden cost of AI code accumulation
AI makes adding code cheap, so codebases accumulate faster than they are pruned. Why volume becomes a cost and how to keep AI-assisted growth healthy.
You will understand: why cheap writing makes scope discipline more important, not less. Read article → WorkflowWhen to delete AI-generated code
Not all AI-generated code is worth keeping. Signals that code should be deleted rather than maintained, and why pruning is part of healthy AI-assisted work.
You will understand: the signals that say delete, and why keeping code is a decision, not a default. Read article →Routing & Capability Control
Matching tasks to the right model and capability setModel routing vs workflow routing: what AI coding teams actually need
Model routing selects which model handles a task. Workflow routing decides the full task path: context, tools, policies, proof requirements, and only then, which model.
You will understand: why model selection is the last routing decision, not the first. Read article → RoutingWhy cheaper AI models are not enough to cut coding costs
Switching to a cheaper AI model lowers the per-token price but not the waste. Why routing, scope, and reuse cut cost more than model price alone.
You will understand: why model price is one lever, not the lever, for AI coding cost. Read article → RoutingAI gateway vs workflow router: what is the difference?
An AI gateway routes API traffic by cost and availability. A workflow router decides the full task path: context, tools, policy, proof, and model. How they differ.
You will understand: why a gateway optimizes the call while a workflow router optimizes the work. Read article → RoutingHow to route AI coding workflows safely
Routing an AI coding task is more than picking a model. How to route by scope, risk, and prior receipts so low-risk work flows and high-risk work gets depth.
You will understand: why safe routing is proportional to scope, boundary, and prior receipts. Read article →No articles in this category yet.