Lab: the use-case charter and eval plan

A sponsor can see the ask: money, people, access, or a no. They can see the kill criteria and the sequence. They can see why this slice and not the chatbot of everything. If the sponsor has to hunt, you will get a 'come back when this is clearer,' which is a no with extra meetings. Put the ask on page one, not after the architecture.

An engineer can see the slice and the non-goals. What we retrieve. What we do not. No write-tools, or one write with confirm. Model path: existing enterprise endpoint, pinned. Eval harness: the set lives here, the bar is this, a prompt change re-runs it. If an engineer reads a vision and no boundary, they will start building a platform. You asked for that by omitting non-goals.

Security can see data class, copies, vendor path, DPA status, log retention, ACL mapping, injection tests, kill switch. That is the day 15 handling note, attached and summarized in one paragraph in the charter. If Security has to ask 'where does the prompt go,' the charter is not ready, and you will lose a month. Do not summarize so hard that the copies disappear.

Finance can see the envelope and the 10×. Tokens, seats, HITL hours, vendor if any, what happens to the bill if volume multiplies. A single 'TBD pending vendor' is not a finance paragraph. A range with a cap and a 10× is a finance paragraph. They can also see who pays on day 91. If the innovation budget expires in June, write the run-rate owner or do not pretend you have funding.

You can see the kill criteria, the roster, and the waiver path. If those are missing, you failed the delivery job even if the other four readers are happy. A charter that can only grow is a POC without a death date, in nicer paper. Put pause and kill on the same page as the ask so the sponsor signs them together.

Steering-ready is not 'long' and it is not 'short.' It is findable. Four to six screens in your notes, or three to five pages if you type it, with attachments. Ten pages of narrative is how readers skip to the last slide and argue from memory. Headings are a control. Use the practice list as the headings. Do not invent a new template the morning of.

Put the ask on page one, not after the architecture. A sponsor who has to hunt will say come back when this is clearer, which is a no with extra meetings. The ask is people, access, money, or a no, with numbers. 'We would love your support' is not an ask. Kill criteria sit on the same page so the sponsor signs them with the ask. A signature on a vision without a kill is how you inherit a zombie. You already know how to put the decision on slide one of a steering pack. Do it here.

Non-goals do more work per line than any other section. No HR spaces. No customer email in v1. No write-tools, or one write with confirm. No company-wide SSO. No platform RFP. Each line prevents a tangent in the meeting. If you feel nervous writing them, that is the sign they are the right ones. Nervous means someone in the room wanted them. An engineer who cannot see non-goals will start a platform. You asked for that by omitting the list.

This is an assembly job, not a research job. Day 14 gave you the pipeline. Day 15 gave you the handling note. Day 16 gave you the golden set and the bar. Day 17 gave you the HITL matrix and the incident page. Day 18 gave you SLOs, fallback, pin, dashboard. Day 19 gave you buy versus build and the death date instinct. Day 20 gave you discovery, score, RAID, sequence, kill. Today's work is to put them under headings with owners, and to cut anything that does not serve a reader.

Start with owners. Every heading gets a name before it gets prose. Corpus owner, eval owner, prompt owner, HITL roster lead, ops owner, Security partner, sponsor, you. If a heading has no name, you will write TBD, and TBD is how steering sends you back. Fill the names first even if one of them is 'ask in the room.' 'Ask in the room' is an honest ask. TBD is a blank.

Paste, then cut. Paste the handling-note summary, the four SLO rows, the four HITL actions, the RAID rows, the score. Then cut until a sponsor can hear it in eight minutes. The attachments hold the rest. People fail this lab by rewriting from memory instead of assembling, and by refusing to cut. Both are pride. The artifact is a pack, not a novel.

Non-goals do more work per line than any other section. No HR spaces. No customer email in v1. No write-tools except Jira confirm, or no write-tools at all. No company-wide SSO. No platform RFP. Each non-goal prevents a tangent in the meeting. If you feel nervous writing them, that is the sign they are the right ones. Nervous means someone in the room wanted them.

The pipeline in eight lines is enough: source, class, chunk, index with ACL, retrieve, prompt contract, pinned model, schema, HITL, logs as metadata. You drew this in week 2. Here it is a boundary, not a tutorial. If the eight lines require a ninth that says 'then the agent plans and uses tools,' you have left the thin slice. Remove the ninth or put it in a later phase with a kill behind it.

Cost envelope is a table, not a paragraph. v1 users, expected questions per day, tokens in and out, dollars per day, HITL hours, vendor seats if any. Then a 10× column. Then a cap. Then who gets the alarm. Finance will mark this table. They will not mark a narrative about 'being good stewards.' You already know that. Put the table in.

Start with owners. Every heading gets a name before it gets prose. Corpus owner, eval owner, prompt owner, HITL lead, ops owner, Security partner, sponsor, you. If a heading has no name, you will write TBD, and TBD is how steering sends you back. 'Ask in the room' is an honest ask. TBD is a blank. Fill names first even if one of them is the ask. This is the same trick as putting names on a RACI before you write the narrative of the plan.

Paste, then cut. Paste the handling-note summary, the four SLO rows, the HITL actions, the RAID rows, the score. Then cut until a sponsor can hear it in eight minutes. The attachments hold the rest. People fail this lab by rewriting from memory instead of assembling, and by refusing to cut. Both are pride. The artifact is a pack, not a novel. Depth that a reader needs lives in the handling note and the golden-set sketch, not in a tenth page of context.

Problem and job-to-be-done. Users. Thin slice. Non-goals. That is the opening. A reader who stops after those four should already know whether this is for them. If they cannot, you started with industry context and a model primer, which no reader asked for. You are not teaching week 1 in the charter. You are asking for a slice.

Data class, owner, ACL, vendor path — from day 15, one paragraph plus the attachment. Name the class. Name what is out (HR, salaries, credentials). Name the copies. Name the DPA status. Retention. This paragraph is the one Security reads first if they only have three minutes. Make it possible to read in three minutes.

Eval: golden-set summary (how many rows, which tags, who labels), three metrics, numeric bar, waiver path — from day 16. A steering member should be able to repeat the bar after you speak. If they cannot, you buried it. Say the numbers in the talk track. Write them in the charter in a box, not in a paragraph of caveats.

HITL dial and roster from day 17. Four actions if you have them, intensity, named humans, hours. SLOs and fallback from day 18. Pin and deprecation one-liner. Dashboard tiles you will demand. These are the operating paragraphs. Without them you have a product idea. With them you have a workstream.

Cost envelope plus 10×. RAID, eight rows, AI-specific. Sequence plus kill. Ask. That is the close of the document and the close of the talk. The ask is a decision, not an 'alignment.' People, access, money, or a no. 'We would love your support' is not an ask. 'Approve two HITL days per week from PMO, existing enterprise endpoint, twelve named users, six-week shadow-then-assist, kill if groundedness is under 0.8 two weeks running' is an ask.

Attachments: handling note, golden-set sketch, SLO sheet, HITL matrix, incident page. Mention them in the charter so people know they exist. Do not paste them in full if it blows past six screens. Reviewers who need them will open them. Reviewers who do not need them will thank you.

The eval paragraph has to be repeatable after you speak. How many rows, which tags, who labels, three metrics, numeric bar, waiver path. Put the numbers in a box, not in a paragraph of caveats. A steering member who cannot repeat the bar did not hear it. Say the numbers in the talk track. Same for SLOs: four numbers, freeze rule, pin. Operating detail that cannot survive eight minutes belongs in the attachment, not in a fog of 'we will monitor quality.'

The pipeline in eight lines is a boundary, not a tutorial: source, class, chunk, index with ACL, retrieve, prompt contract, pinned model, schema, HITL, logs as metadata. If the eight lines require a ninth that says then the agent plans and uses tools, you have left the thin slice. Remove the ninth or put it in a later phase with a kill behind it. Week 2 already drew this. Here it tells an engineer what not to start.

Eight minutes is long enough to decide and short enough that you cannot hide. Structure it like the charter, not like a story of how you learned AI. Minute 0–1: the job and why it hurts, in their words. Minute 1–2: the slice and the non-goals. Minute 2–3: data class and the lightest approved path. Minute 3–4: eval bar, three numbers. Minute 4–5: HITL and sequence. Minute 5–6: cost, 10×, day 91. Minute 6–7: RAID highlights and kill criteria. Minute 7–8: the ask, timed, then stop talking.

Last sixty seconds is the ask. Write it out. Read it. Do not improvise the ask. Improvised asks become 'so we wanted to get your thoughts.' Timed asks become 'I need you to approve X, Y, Z today; if not, the no we need is so we can kill the wait.' Silence after the ask is part of the track. Count to three. Let them talk.

Name the four readers in the room as you go, without making it a circus. 'For Security: class internal, existing endpoint, metadata logs, handling note in the pack.' 'For Finance: envelope here, 10× here, cap here.' 'For engineering: non-goals, pin, eval re-run.' 'For you, sponsor: the ask is at the end and the kill is on the page you would sign.' People listen when they hear their job.

Do not teach. If you spend two minutes on how models work, you have stolen the data paragraph or the ask. This audience has sat through that primer. They have not sat through your bar, your roster, and your kill. Spend the minutes on the parts only you brought. The primer is available elsewhere. The decision is not.

Handle the HIPPO question if it comes: 'why not the chatbot of everything.' One sentence from day 20: we scored it, feasibility 1, we are funding the eval-able slice, the door back in is data-ready plus a re-score. Then go back to the ask. Do not take the remaining minutes to fight a vision. Visions eat talk tracks.

Practice once with a timer, out loud, standing if you will stand in the room. You will run long the first time. Cut from context, not from the ask. If you are still long, cut a RAID story. Keep kill criteria spoken, even if the rest of RAID is 'eight rows in the pack.' Kill that is only written will not be remembered when you need it.

Do not teach. If you spend two minutes on how models work, you have stolen the data paragraph or the ask. This audience has sat through that primer. They have not sat through your bar, your roster, and your kill. Spend the minutes on the parts only you brought. Name the four readers as you go without making it a circus: for Security, class and path; for Finance, envelope and 10×; for engineering, non-goals and pin; for the sponsor, the ask and the kill. People listen when they hear their job.

Handle the HIPPO question if it comes in one sentence from day 20: we scored the chatbot of everything, feasibility 1, we are funding the eval-able slice, the door back in is data-ready plus a re-score. Then go back to the ask. Do not take the remaining minutes to fight a vision. Visions eat talk tracks. Silence after the ask is part of the track. Count to three. Let them talk. Improvised asks become 'we wanted your thoughts.' Timed asks become a decision.

If kill criteria are missing, the POC becomes production by inertia. You failed the delivery job. That is one of the check questions on purpose. Put three numbers on the charter: eval miss, adoption miss, risk event. Pause versus kill. Named waiver. The sponsor signs them with the ask. A signature on a vision without a kill is how you inherit a zombie.

Waivers in the charter are the same shape as day 16. Who, on what metric, for how long, rollback. You are not inviting a culture of waivers. You are preventing hallway waivers by showing the legal kind. People who hate process still use hallways. Meet them with a form that takes four lines, not with a speech.

What you will not pretend: that thumbs equal quality, that a vendor demo is an eval, that HITL without hours is a control, that floating latest is a strategy, that a platform SOW is a prerequisite, that the corpus is ready because the drive exists, that users will cite-check, that day 91 will be owned by 'the team.' Each of those pretences has a paragraph in this week. The charter is where you refuse them in writing.

If you cannot fill an owner, the ask includes that owner. 'Approve a corpus owner from Knowledge Management, name to be confirmed this week by Sam' is a valid ask. Shipping with a hole and hoping Sam finds someone is how data-ready stays red while the model workstream reports green. You know how that ends. Do not write it into this artifact.

Atlas is an acceptable setting if you do not want to use a live program. Specificity still counts. Named spaces, named class, named twelve users, named bar. A fictional-but-specific charter beats a real-but-vague one in an interview, because the interviewer can probe the numbers. Vague real work sounds like you were adjacent. Specific invented work sounds like you can run it. Both can be honest if you label them.

When you tick the capstone, you are claiming the headings are there and the owners are not TBD. Read it once as each of the four readers. If you, playing Security, cannot see copies, you are not done. If you, playing Finance, cannot see 10×, you are not done. That pass takes ten minutes and saves a steering loop.

If you cannot fill an owner, the ask includes that owner. 'Approve a corpus owner from Knowledge Management, name to be confirmed this week by Sam' is a valid ask. Shipping with a hole and hoping Sam finds someone is how data-ready stays red while the model workstream reports green. You know how that ends. Do not write it into this artifact. Atlas is an acceptable setting if you label it. Specific invented work beats vague real work in an interview, because the interviewer can probe the numbers.

What you will not pretend, in writing: that thumbs equal quality, that a vendor demo is an eval, that HITL without hours is a control, that floating latest is a strategy, that a platform SOW is a prerequisite, that the corpus is ready because the drive exists, that users will cite-check, that day 91 will be owned by 'the team.' Each pretence has a paragraph in this week. The charter is where you refuse them. Waivers, if needed, use the four-line shape so hallways are not easier than the form.

You now have the operating system: data-ready, evals, HITL, SLOs, vendors, discovery, scoring, and a charter that can take markup from four kinds of skeptic. Combined with week 2's thin-slice design, you can run an AI workstream. That is the claim of this month. It is not a claim that you can train models, and you should not make that claim.

Week 4 is the job market — same material, pointed at a hiring manager. The charter, the prompt library, and the wiki design are the artifacts you will walk. A hiring loop that cannot probe these is a loop that would have hired on vocabulary. You want the loop that probes them. Prepare to draw the data path, the eval stack, and the HITL matrix on a whiteboard without looking at notes. Today's redraw is practice for that.

What you do with the charter after this lab: keep it as a living pack. If you are in seat, run it. If you are not, it is the core of a case story. Either way, do not archive it as a course completion. Artifacts that die at 'done' were homework. Artifacts you can speak for eight minutes are how you change seats.

The temptation at the end of week 3 is to add more literacy — another vendor, another paper, another agent framework. Resist it. Depth on one slice beats a new object. If you have energy, improve the golden set and the talk track. Those two repay. A new framework does not, not this week.

You will forget numbers. You will not forget the shape if you redrew the diagrams. Data path with copies. Three-layer eval. Control plane. Request path with fallback. POC with a death date. Ninety-minute discovery. Four readers. That set of drawings is the operating system in a form you can reconstruct under interview stress. If you skipped the redraws, do them tonight. They are not decorative.

Close the day by ticking the artifact only if the headings and owners are real. If they are not, you are still in the lab. That honesty is the same honesty you will need when the eval is red and a sponsor wants to ship. Week 3 was the job. Week 4 is packaging the job so someone will pay you to do it again, on purpose.

Week 4 is the same material pointed at a hiring manager. The charter, the prompt library, and the wiki design are the artifacts you will walk. Prepare to draw the data path, the eval stack, and the HITL matrix without notes. Today's redraw is practice for that. A hiring loop that cannot probe these is a loop that would have hired on vocabulary. You want the loop that probes them. Do not archive the charter as course completion. Artifacts you can speak for eight minutes are how you change seats.

The temptation at the end of week 3 is to add more literacy — another vendor, another paper, another agent framework. Resist it. Depth on one slice beats a new object. If you have energy, improve the golden set and the talk track. Those two repay. A new framework does not, not this week. You will forget numbers. You will not forget the shape if you redrew the diagrams. That set of drawings is the operating system in a form you can reconstruct under interview stress.