← Ben Lai

Start with the problem

The archive is chronological. Your week is not. Pick the thing that is eating time, money, or attention; each path is three essays and one honest skip condition.

  1. 01

    Burning money on tokens?

    Your AI bill is getting weird

    Start here if the model bill is growing faster than the value, or if every agent answer is longer than the decision it supports.

    For you if

    You own an LLM workflow, agent, or support/tooling loop and need to cut cost without pretending usage will go down.

    Skip if

    You are still pre-production and have no real token distribution, logs, or provider bill to inspect.

    1. 1

      The output token tax

      Shows where output verbosity quietly became the expensive side of the invoice.

    2. 2

      Your prompt isn't cache-shaped

      Turns prompt structure into a cacheability problem instead of a provider-pricing complaint.

    3. 3

      You're optimizing the wrong axis of agent cost

      Reframes cost away from per-call speed and toward failed loops, retries, and unnecessary work.

  2. 02

    Agents in the repo now?

    Your team just got agents

    Read this path when agents have moved from demo to daily workflow and the operational boundary is still mostly vibes.

    For you if

    You are letting coding agents or MCP-connected assistants touch repos, tools, or production-adjacent workflows.

    Skip if

    Your only agent use is one-off chat with no tool access, no repo writes, and no shared operating process.

    1. 1

      One worktree per agent

      Establishes the working-tree boundary that keeps parallel agent work from trampling itself.

    2. 2

      Your MCP server is a prod dependency

      Treats MCP servers as dependencies with versioning, security, and outage surfaces.

    3. 3

      Vibe-coded code has more bugs. Price it in.

      Prices in the defect tax instead of pretending velocity comes free.

  3. 03

    PRs sitting for days?

    Your week disappears before coding

    Use this path when the code itself is not the bottleneck; context loading, review loops, and deploy memory are.

    For you if

    You lead or work inside a small engineering team that feels busy but cannot point to where the week went.

    Skip if

    The bottleneck is a known missing feature or a single overloaded person; fix that directly before diagnosing process drag.

    1. 1

      The 25% rule

      Separates coding time from the three quarters of engineering work nobody optimizes.

    2. 2

      The deploy log is your changelog. Tag accordingly.

      Converts release memory from hand-written ceremony into a durable operational artifact.

    3. 3

      Observability is a tax you pay before you owe it

      Makes the case for paying observability cost before incidents make it non-negotiable.

  4. 04

    Too much company for the work?

    Your small team has big-company habits

    This path is for founder/operators trying to keep a small team sharp without importing the meetings, spend, and headcount logic of a larger one.

    For you if

    You are choosing between hiring, spending, pricing, or process changes while the company is still small enough for each choice to matter.

    Skip if

    You already have the market, margin, and management layer to absorb big-company defaults without slowing the builders.

    1. 1

      The 3-person team is the new 50-person team

      Sets the baseline: small teams can beat larger ones when meetings are not the product.

    2. 2

      The runway you didn't spend is the option you bought

      Treats unspent cash as strategic optionality rather than unused budget.

    3. 3

      Your first hire is a multiplier or a manager. Pick.

      Frames the first hire as leverage, not status or managerial comfort.