Harper · a best-practices guide

Agentic AI at Booth

How to use AI agents for the work you already have to do, and how to set them up so the tenth session is better than the first. Free. Nothing to install to read it.

01 · Why context comes first

The model is rarely the bottleneck. What it knows about your project is.

Every agent starts a session knowing nothing about what you care about, what good looks like, or what went wrong last time, unless it is written somewhere the agent reads. The numbers below are the ones that decide how much of that it can hold and how expensive it is to be vague.

1Mtokens of context on the current Claude models (Opus 4.8, Sonnet 5, Fable 5). Room for a whole project, if what you put in it earns its place.Anthropic model table, July 2026

+30%more tokens for the same text on the current tokenizers. Long instruction files cost more than they did a generation ago.Anthropic docs figure, July 2026

78%of defects in one builder's audit already had a written rule. The bottleneck was enforcement, not knowledge, which is why the packet pairs each rule with a check.Builder's defect audit, 2026

Constraints, not explanations. A 500-line instruction file with the reasoning behind every rule is worse than a 100-line file with just the rules, because the agent has less room to think about the task. It does not need to know why the primary color is a certain hex value. It needs to know to use it.

One more pattern from the same guide, offered as illustrative rather than measured: generated code is almost always syntactically valid, usually right for the ordinary case, and frequently wrong at the edges. A validation habit has to match that shape, which is what the Test step below is for.

02 · The method

Five steps, in order

The sequence keeps judgment in the project as it grows. It is the same one taught in the Booth session on building judgment into AI projects.

NoticeA recurring problem, who it affects, and what a useful result would look like. If nobody could tell you the answer is wrong, the project has no test yet.

InvestigateGather references and expertise. Sort what you know, what you assume, and what needs a source. Write the constitution: purpose, values, authority, evidence, revision.

TestWhere the AI gets it wrong, and whether the result helps its intended user. Check the claim against its source, then check whether a person can use it.

PreserveDecisions and corrections saved as files the next session reads: skills, a decision journal, a build log.

RevisitFeedback and changed sources update the project. Rules that no longer help are retired, not worked around.

03 · The packet

Seven plain-text files that turn it on

Copy the folder into a new project, fill in the blanks, point your agent at it. Claude Code reads CLAUDE.md by name and Codex reads AGENTS.md; keep both as the same file if you use both.

your-project/ ├── CLAUDE.md what the agent reads first: purpose, rules, where things live ├── constitution.md purpose, values, authority, evidence, revision ├── skills/ one file per repeatable job ├── references/ examples you admire, source documents ├── docs/decisions.md the decision journal ├── _BUILD_LOG.md dated entries, append only └── deliverables/ what leaves the project

The constitution, in five questions

Purpose. Who should this help, and with what? Values. What should remain true as it changes? Authority. Which decisions belong to the person, and which can the agent make? Evidence. What needs a source or a test? Revision. How can an outdated rule be challenged and changed?

A constitution guides the agent. It does not guarantee compliance; that is what the Test step is for. Harper's own constitution is the worked example in the packet.

The starter packet (zip)See the files rendered, with a skills viewer

04 · Where each model fits

Roles from one working system, not a leaderboard

These assignments come from a system that runs several agents on real projects every day (Harper is one of its outputs). They are usage roles, not rankings, and prices move; check the current table before you commit to one.

Model	Role in the system	Technique that fits it	Price, in/out per million tokens
Claude Opus class	The main session: context architecture, judgment calls, review of other agents' work.	It reads the constitution and references before anything else. It over-delegates and over-verifies when the number of helpers it may spawn is left open, so that number is capped.	$5 / $25 (Opus 4.8)
Claude Sonnet class	Parallel helpers: research passes, page audits, enrichment jobs.	One bounded task each, with results written to a file the main session reads. Cheap enough to fan out.	$2 / $10 intro, $3 / $15 from Sep 1 (Sonnet 5)
Claude Haiku class	Hooks and fast routing: the checks that run before and after every file write.	Yes-or-no questions with a strict format. Never the judgment call.	$1 / $5 (Haiku 4.5)
Fable / Mythos class	A deliberate-choice tier above Opus for the hardest synthesis.	Delegation-friendly; step-by-step scaffolding lowers its output. Prefer Opus or Sonnet for anything dependency-critical.	$10 / $50
Codex	Implementation and bug-finding partner, in its own branch and worktree.	It works best from a plan, a repo with explicit conventions, and checkpoints, because it reads conventions aggressively and acts on them.	Per OpenAI's current plan

Prices are the July 2026 Anthropic table, listed so the cost of a vague prompt is concrete: a long, unfocused context on the Opus tier costs five times what the same tokens cost on Sonnet. The direction of a good setup is the same on every row: the better the executor, the more expensive a bad plan becomes.

“Codex does not reduce the need for domain expertise. Better implementation just makes a weak framing look more convincingly complete.”From the builder's notes comparing Claude and Codex, chapter 3

“Codex may catch more implementation bugs, but it does not remove the need for explicit validation. It just shifts where I expect failures to cluster.”From the same notes, chapter 21

“Codex makes planning even more valuable because it can move quickly once unleashed. The better the executor, the more expensive a bad plan becomes.”Chapter 19

“Codex makes skills feel less like a convenience and more like core infrastructure for keeping hard-won lessons loadable and bounded.”Chapter 12

05 · Techniques

Habits that survive the session

Reference first

“Make it good” is not an instruction. Two to five examples you admire plus a few written rules (colors by value, sizes in pixels, banned words) is an instruction. The design comes from those, not from defaults.

The three-times rule

Something done three times gets a tool or a skill. Something done fewer than three times gets a prompt. The rule protects against automating one-offs and against doing the same manual thing fifty times.

The proof chain

Generate, test, assert, catch, fix. Structure the task so each step validates the previous one. “It should work” is not a step.

Verify before claiming done

Run it, open it, read the output. For a page, open it at phone width and look for text running off its box. For a number, trace it to its source or mark it unverified. Anything with a login usually cannot be tested by the agent; it says so and hands you a script.

One source of truth

A fact that appears in two places lives in one file and the other reads from it. When the value changes, write the old value down as forbidden so a search catches any copy that still shows it.

Two kinds of repetition

Copying the same rule into many files causes drift. Checking one requirement three different ways (a written rule, an automated check, a person trying the task) catches different failures. Prefer the second kind.

Honest logs

End every session with a dated entry: what was built, what broke on the way, what was verified and what was not. The wrong turn is the useful part. Never rewrite an old entry.

What kept coming up this month

Across 32 lectures, workshops, Career Services sessions, and conversations since September 10, as Jenn noted them. Counted once per session; paraphrased; nobody named.

Values named first make the decisions easierAn authentic leader starts with what matters to them. Everything else gets anchored to it.18 sessions · 09/11 to 09/23
Small experiments, with the discomfort as the signalOne thing at a time, then reflection, then something learned about yourself. Applied to AI, that meant one thing built with every part under your control.12 sessions · 09/15 to 09/23
The idea comes from you; the tool runs with itThe tool amplifies a problem and an idea; it does not supply them. The credible guide says where it went wrong.9 sessions · 09/10 to 09/23
Work shown as a portfolio, not as a claimFor tech paths, a site with the projects behind the claims. The web address goes on the résumé.7 sessions · 09/17 to 09/23
The information is scattered, so someone has to centralize itOverwhelm is universal in the first weeks. Harper exists because of this one.6 sessions · 09/11 to 09/23

Regenerated from Jenn’s notes as new ones land; method and caveats are in the project notes (as of 2026-09-25).

06 · From this week

What Career Services said about AI in the job search

From the Day 1 job-search session on September 23, 2026 (consulting, banking, tech, investment management, venture). Paraphrased from the room; the official instructions in Canvas and BCP govern.

Show the AI work as a portfolio, not a line on the résuméFor tech paths, the advice was to have your own site with the projects behind the claims: academic, personal, or professional. The question to answer is how a recruiter can see what you built. Harper and this guide exist partly to be that kind of evidence.

Print the web address; hyperlinks are optionalSome applicant systems treat links as spam, and résumés pulled through BCP may not keep them clickable. The safe form is the address in plain text, with a hyperlink as a bonus. Tech companies expect to see the site named.

Recruiters say they can tell when an application was written by AIOne employer contact reported a wave of applications with no personalization. With bots submitting on people's behalf, companies are flooded and looking for reasons to decline many at once. The guide's rule applies: the agent drafts, the person supplies the specifics and reads every line before it goes.

Most companies are not auto-screening with AI, because of pending lawsuitsThe capability exists; the practice mostly does not. Humans are reading, and they are tired. That is an argument for a clear, specific application, not a keyword-stuffed one.

Everyone's résumé is strong now, so the interview goes deeperBecause AI raised the floor on written materials, interviews probe whether you know what your résumé says you know. Know the content and the function behind every bullet. This is the same reason the packet's Test step checks the person can defend the output, not just that it reads well.

Write one cover letter and reuse itViews in the room ranged from "never wrote one" to "always". The consensus: it can only help, do not overthink it, and get the company name right.

Dates mentioned in the session: the Booth Technology Group kickoff on October 1, and uploading your résumé to BCP before company events open. Confirm both in the recruiting calendar and in BCP.

Provenance

Where this comes from

Written from a working builder's guide and its project records, adapted for Booth. The full guide, with chapters on Claude Code, Codex, hooks, MCP servers, and the failures behind each rule, is at jennumanzor.com/agentic-builder-guide. Model prices and context figures are from Anthropic's published tables as of July 2026 and will drift; the roles are one system's choices, not a recommendation for every project. Independent student resource; not an official University of Chicago or Chicago Booth publication.