@khamsouk

Khamsouk Souvanlasy
Fractional CTO · Applied AI

Guides · Slide deck

Working with AI: managing a brilliant new hire

It reads faster than anyone you have hired, and it forgets yesterday overnight. Fifteen habits for managing it, each with the reason behind it and one thing to try this week.

17 slides · No code required · ← → to move · N for notes · Esc to exit

Speaker notes

This deck is for people who manage AI work. Each slide is one habit from the Claude Code guides (Claude Code is Anthropic’s AI assistant that works in your files): the reason for it, the evidence and one thing to try. Arrow keys move; each slide builds itself, then waits.

01 · The new hire

Write down what it needs to know tomorrow

Each new session forgets yesterday’s conversation. Some tools, Claude Code among them, save a few notes on their own, but they choose what to keep. Anything you agreed and will need again goes in a briefing file you control.

Try this: At the end of a session, ask it to write the decisions to a file. Start a new session, tell it to read only that file, and ask it to list them.

Read the guide: The memory hierarchy

Speaker notes

The model keeps no conversation between sessions. Claude Code’s auto-memory saves selected corrections and preferences, which helps, but it decides what to save and loads only the first 200 lines (or 25KB) of its main notes file. Treat it like a sharp new hire on day one, every day: the briefing you leave is most of the continuity it gets.

  1. MondayLearns your project, your naming, your decisions
  2. OvernightThe session ends; the conversation is gone
  3. TuesdayStarts with the briefing file and whatever notes it chose to save

02 · The new hire

Check what kind of job it is before using AI

AI does well where several answers are acceptable: drafts, summaries, first ideas. In one study, consultants using GPT-4 did more work, faster, on tasks it suited. On one task it did not suit, they were wrong more often than consultants without it.

Try this: Before you hand over a task, ask: would two different answers be a bug? If yes, keep AI out of the live job. If several answers are right, let it draft.

Read the guide: When to use AI, and when to build the system that doesn't need it

Speaker notes

The Jagged Frontier study followed 758 BCG consultants, some given GPT-4 and some not. Inside the frontier they did more work, faster. On one task outside it, they were more likely to get it wrong than without AI.

  • 12.2%More tasksOn consulting tasks suited to AI
  • 25.1%FasterOn the same tasks
  • 19%Less likely to be rightOn one task outside what GPT-4 did well

Dell'Acqua, McFowland, Mollick et al., Organization Science, 2026

03 · The desk

Give it only what the task needs

AI works only from its working memory, called the context window. Picture a desk: every message, file and reply stays on it, and each message you send carries the whole desk back to the model. So cost grows as you go, and a crowded desk makes it miss instructions.

Try this: Start your next request with a two-line brief: the task and the files or documents it needs. Ask it to name both back before it begins.

Read the guide: Session craft

Speaker notes

The context window is the desk. Its own instructions and your briefing file are on it before you type. Each turn adds your message, the files it opens and its reply. In my sessions, a typical file it opens adds 500 to 2,000 tokens. Every turn re-reads the whole conversation; content it has seen recently is re-read from a cache at 90% to 97.5% off, which softens the cost while the desk stays just as crowded (token economics guide). Claude Code now gives every current model a 1M-token desk by default. My measurements come from 200K desks; the bigger desk fills the same way.

  1. setupIts instructions and your briefing file
  2. turn 1You ask a question
  3. It reads six files, then answers
  4. turn 2A follow-up: more files, a longer answer
  5. turn 3Past 65%: instructions start to slip

04 · The desk

Start a fresh session at each new task, with a handoff

Quality drops well before the desk is full: it misses instructions and contradicts earlier decisions. A fresh session fixes that, and a short handoff note carries the decisions across. In my sessions on a 200,000-token desk, the drop came at about two-thirds full.

Try this: Before you clear, ask it for a handoff: decisions made, what is left, the next step. Read it, then start the fresh session from it.

Read the guide: Session craft

Speaker notes

Past about two-thirds full, it misses instructions, repeats suggestions and contradicts earlier decisions. Starting again costs a few thousand tokens; a written handoff carries the decisions across. I clear at each task boundary or after about 50 exchanges. The phases come from my own tracking, not product guarantees.

  • 0–40%FreshFollows every instruction
  • 40–65%WorkingStill effective; decide whether to clear
  • 65%+DegradingMissed instructions, repeated suggestions

From my own tracking on 200K windows, not product guarantees

05 · The desk

Keep the briefing file short; move detail elsewhere

The briefing file (CLAUDE.md) is read at the start of every session, and each line competes for attention. Anthropic, which makes Claude, recommends about 200 lines or fewer. Add a line when it makes a mistake or the project changes, and move specialist detail to files that load when needed.

Try this: Read your briefing file line by line and ask of each: would removing it cause a mistake? Delete the lines that would not.

Read the guide: CLAUDE.md

Speaker notes

Anthropic’s guidance is about 200 lines or fewer. I cut one file from 280 lines to 85 with no drop in quality, and one task came out better, which I put down to fewer instructions competing for attention. My own site’s file fell from about 740 lines to 77, with the detail moved into rules files that load only for the files they cover.

  • Before 280Instructions for mistakes it never made
  • Ceiling 200Anthropic's recommended maximum
  • After 85No drop in quality

06 · The team

Send each task to the cheapest model that does it well

Models come in sizes, junior to senior, and price rises with size. Send search and extraction to a junior model; pay for a senior one where judgement decides what ships. In my tests, the junior Haiku matched the mid-size Sonnet at code search for a twentieth of the price.

Try this: Pick a search whose right answer you already know. Run it twice on each model; move that job to the cheaper model only if it gets both right.

Read the guide: Model selection

Speaker notes

At October 2026 list prices per million input tokens: Haiku 5.5 $0.10 (prompts up to 100,000 tokens), Sonnet 5.5 $2, Opus 5.5 $4 and Fable 5.1 $10; output costs five times input on every model. Haiku costs a twentieth of Sonnet and a fortieth of Opus. On Terminal-Bench 4.0 Opus 5.5 beats Fable at 40% of the price, in Anthropic’s own launch figures, so I keep Fable for the rare problem Opus cannot solve (model selection guide). Prices and benchmarks change every few months; the job sizes change less.

  • JuniorHaiku 5.5Search, fetch, extract, classify
  • MidSonnet 5.5Standard building and tests
  • SeniorOpus 5.5Leads the session; drafting, review, root causes, security
  • RareFable 5.1Problems Opus cannot solve

Prices are in the speaker notes; they change every few months

07 · The team

Have it ask its questions before it starts

Given a vague request, AI guesses rather than asks, often wrongly. In one coding test, hiding detail from the brief cut success from 68.0% to 40.0%; letting it ask clarifying questions brought it most of the way back, to 61.4%.

Try this: End your next request with: “Ask me any questions whose answers would change the plan, then wait.” Answer every one before it starts.

Read the guide: The tech-lead pattern

Speaker notes

Ambig-SWE rebuilt a coding benchmark with detail withheld. Sonnet 4 solved 68.0% with the full brief, 40.0% with detail hidden, and 61.4% when it could ask up to three rounds of clarifying questions. It is one benchmark, but all six models tested dropped without the detail and recovered when they could ask. So I have it ask its questions first, and bigger work gets a written brief before any code.

  • Full brief 68.0%
  • Detail hidden 40.0%
  • Allowed to ask 61.4%

Ambig-SWE, peer-reviewed at ICLR 2026, an AI research conference

08 · The loop

Let checks, not the AI’s word, decide when work is done

AI can say a job is done when it is not. So AI drafts, automatic tests check the draft the same way every time, and a reviewer reads it. A draft that fails goes round again. A person decides what the tests cannot.

Try this: Before a draft with figures goes out, have a fresh session mark each price, date and statistic as checked against a named source, or flag it for you.

Read the guide: Quality gates

Speaker notes

The model can give different answers to the same input, so it works inside checks that always give the same answer: automatic tests and style checkers. A draft that fails goes back round the loop. Nothing leaves the loop on the model’s word alone.

  1. PlanA short brief: what done looks like
  2. BuildThe model drafts the change
  3. CheckAutomatic tests and style checks: the same answer every time
  4. ReviewA separate reviewer reads the change

Fails a check? Back to the start.

09 · The loop

Run several agents at once when the pieces don’t overlap

An agent is one AI worker with its own desk. Independent pieces of a job can run side by side, which cuts the wait but raises the total cost. So split only larger jobs with three or more separate parts. In my client work, one feature took about 20 minutes this way instead of three hours.

Try this: Before you split a job, list its parts. Run parts side by side only if each touches different files and needs nothing from the others.

Read the guide: Multi-agent orchestration

Speaker notes

On one client’s onboarding feature, the work ran sequentially in about three hours. Split across agents, it took about 20 minutes: three building agents worked side by side, each on its own copy of the code. Research, planning and review still ran one after another. Total token spend went up. I split work that would take more than about 45 minutes and has at least three separate concerns.

  • One agent ~3 hEach step waits for the last
  • Parallel agents ~20 minMore tokens in total

10 · Trust

Have someone else mark its homework

Like a person, AI misses some of its own mistakes. So a fresh agent reviews most changes; it never saw the work being written. A person signs off on facts, money and anything public.

Try this: Save the work, clear the conversation, then ask the fresh session to review the change. It sees only what was written, never the reasoning.

Read the guide: Quality gates

Speaker notes

A model reviewing its own work misses its own mistakes, as people do. Separately, in CodeRabbit’s sample of 470 open-source code changes, those classified as AI co-authored had about 1.7 times as many review findings. Plausible output still needs checking.

  1. WriterDrafts the change
  2. ReviewerA separate agent with a clear desk
  3. YouApprove facts, money and anything public

Code changes classified as AI co-authored had about 1.7 times as many review findings (CodeRabbit, 470 open-source changes)

11 · Trust

Turn repeated mistakes into automatic rules

A note asking it to do better is a suggestion it can ignore. A rule is a small automatic check that runs before each action and either fixes a simple slip or stops the action. Each rule is tested on one case it must catch and one it must allow.

Try this: Next time it repeats a mistake, write down exactly what you would see if it happened again, such as a phrase or a file name. That becomes the rule.

Read the guide: Claude Code hook examples

Speaker notes

A written lesson is a suggestion; a rule is enforced. Each rule ships with a test pair: one input it must catch, one it must let through. A rule 60 days old that has never fired is listed for deletion; safety rules stay.

  1. MistakeIt does something wrong
  2. LessonWritten down while it is fresh
  3. Rule + testOne input to catch, one to pass
  4. CaughtNext time, it is fixed or stopped before it runs

12 · Trust

Use AI to build the system, then run it without AI

AI can give two different answers to the same question. Anything that must be right every time, such as tax, currency conversion or compliance, should run as ordinary software. AI is often the fastest way to write that software.

Try this: For a job that must be right every time, ask it for a script and tests with known answers. Check the results yourself, then run the script on its own.

Read the guide: When to use AI, and when to build the system that doesn't need it

Speaker notes

Tax, compliance, currency and queries against live data never run inside a model: two answers to one input is a bug. The model writes the calculator; the calculator runs on its own, tests check it on every change, and a person reviews the exceptions instead of every line.

  1. AI draftsThe code, the tests, the edge cases
  2. You checkAgainst known answers; tests rerun on every change
  3. Code runsSame input, same output, no model

13 · Trust

Decide what it may see before it sees it

On a personal plan, the AI company may keep your sessions for up to five years and train on them unless you switch that off. Shared chat links have turned up on Google. Run client work on a business login, and strip personal details before anything reaches the model.

Try this: Check which login each computer uses for AI. On any personal plan, open the privacy settings and switch model training off.

Read the guide: Claude Code data privacy

Speaker notes

Claude Code on a Pro or Max login runs under Anthropic’s consumer terms: 30 days with model training off, up to five years with it on, and rating a response keeps that whole conversation for up to five years. Team, Enterprise and API logins are not trained on (Team and Enterprise except feedback and bug reports), and Claude Code keeps their data 30 days. Zero data retention needs a developer (API) account approved by Anthropic, Enterprise’s separate Claude Code offering, or Claude on Amazon Bedrock set to keep nothing; Fable, the dearest model, still needs 30 days in most cases. In July 2026 hundreds of shared Claude chats and artifacts, some holding API keys and CRM exports, turned up on Google (reported by 404 Media and TechCrunch). My two CRM tools drop personal fields before they return anything. Terms checked 21 October 2026 (data privacy guide).

  • 5 yearsPersonal plan, training onOr any chat you rate
  • 30 daysTraining off, or a business loginTeam or Enterprise: no training except feedback; API: none
  • ZeroBy agreement, or on AWSApproved API or Enterprise organisations, or Amazon Bedrock set to keep nothing

Anthropic's published terms, checked 21 October 2026

14 · Trust

Retest on known answers whenever the model changes

A model update can break a prompt that worked yesterday, and a better average can hide cases that now fail. Keep cases with known answers, check them the same way every time and rerun them before you switch models.

Try this: When it gets a case wrong, save the case and the right answer as a test before you fix it. Rerun every saved test before you switch models.

Read the guide: Evals and drift

Speaker notes

One compliance assistant I worked on had to quote only approved prices. Its test asks the fee question: the answer must contain the approved $5,000 and never the proposed $6,500 (both figures illustrative). A text check like that is cheap and gives the same result every time. Keep an AI judge for what simpler checks cannot grade, and test the judge first: on some tasks, older AI judges caught only 30% to 60% of defects. Pin the exact model version, and treat a swap as a migration with a tested way back (evals and drift guide).

  1. MissIt gets a real case wrong
  2. TestThe case and its right answer join the set
  3. FixChange the prompt or the code until it passes
  4. RerunEvery test, on every change and every new model

15 · The bill

Cut the bill: route jobs, clear old context

Most of the bill pays for what AI reads: files, logs and the conversation so far. In my own work, a month across six client projects cost over $900; the next month, with the same work, cost $340. In my tracking, three habits did most of it: routing jobs to cheaper models, clearing old context and keeping the briefing file short.

Try this: Note this month’s AI bill or usage before you change anything. Without a starting figure, you cannot tell what a change saved.

Read the guide: Token economics

Speaker notes

The start and end are my real API bills, a month apart, with the same workload and output quality. Making Sonnet the default at the time cut my monthly spend by about 35%. In my tracking, routing, clearing between unrelated tasks and a lean briefing file account for most of the saving. I did not measure each habit’s share, so the chart shows only the two bills. On a flat-price plan the same habits buy more work for the same allowance, though there I keep Opus in the main session and turn down its effort setting (how much it thinks before answering) instead.

Month one $900+Month two, same work $340

  1. Same work, three habitsRoute each job, clear old context, keep the briefing file short

My real API bills, a month apart; each habit's share was not measured

16 · Limits

Where these habits fall short

These habits come from my own client work and the studies each guide cites. They are a starting point, and they change when the models change.

Discuss a related problemAll 17 guides

Speaker notes

Name the limits: prices and benchmarks move every few months; my thresholds come from my own projects; parallel agents and reviewers cost more tokens. The guides carry the detail and the sources.

  • October 2026Prices and modelsThey change every few months
  • MineThresholdsSet from my own projects
  • More tokensChecks costReviewers and parallel agents add spend