Who this is for, and what they are trying to get done

Every claim below carries a mark. The mark says how it is known, never how much we like it.

Two rules that stop the marks inflating. A * never becomes a ✓ by repetition — five people’s memory is five people’s memory, and only an instrument promotes it. And a re-runnable query proves that something was said, not that it is true: an issue body, a forum comment and a tool’s own README are all people describing their own behaviour, so the ✓ covers the utterance and the behaviour underneath stays self-report.

The evidence base, in one paragraph. Five public issue trackers, the largest of them 89,923 issues. 2,036 Hacker News comments downloaded and read in full, including one 274-comment thread on exactly this subject. A GitHub repository search, and four real agent-material repositories counted file by file through the tree API. About 140 captured frames of other people’s products. And one thirty-minute conversation with one practitioner. That is everything. Reddit, named across community surveys as where this population actually talks, blocks our crawler and is missing entirely.

Why there are no names, ages or photographs. The trap this stage most easily falls into is inventing a persona from the specification and then citing it back at the specification as evidence. The audience sentence in our own spec — design engineers, 25+, visually literate, living in Linear, Vercel, Raycast, Figma — was written before a single user was observed, so it is a claim under test, not a source. Each persona below is named by what it does, and the passport is left blank because there is nothing honest to put in it.

The blanket label is gone and every mark stayed exactly where it was. This page used to carry one word — provisional — over all of it, lifting on five practitioner conversations, of which one happened and four became unavailable. On 15 Sep 2026 the register dropped the word rather than meeting its trigger, on the rule the evidence system is built on: the mark is per claim, not per document. One word both over-warned about the primary persona, which carries 20 confirmed against 5 unknown, and under-warned about the empty-handed one, which carries 2 against 6 and is the persona two shipped surfaces already serve. Nothing was promoted: every one practitioner said so is still that, every unknown is still unknown, and the four conversations are still owed and still unavailable. Read each card’s own what this card is allowed to settle line instead.

What was read before any of this was written, and the seven things it turned up.

Phase 01 established what vendors sell and what breaks. None of it established who the person is. This phase reads the public record with new instruments and then, for the first time, asks somebody. It is unfinished: four more conversations are owed.

What was read, and what could not be

The Claude Code issue tracker — 89,923 issues, which is the tracker of the product this audience actually uses and which stage 3 never opened; the Codex and Gemini CLI trackers; 1,762 Hacker News comments downloaded from eight threads and read in full; Stack Overflow; GitHub repository search. Reddit blocks our crawler, so the largest venue in the space is missing by technical accident rather than by judgement.

Seven findings

Then every claim in the personas and the jobs was audited one by one — 238 of them, 163 confirmed, 48 hypothesis, 27 invented — and the five questions that audit raised were taken back to the public record. That second round counted four real agent-material repositories file by file, which is the first instrument here that measures somebody’s collection instead of asking them how big it is. It moved five marks and changed one persona’s character.

Split on behaviour we can actually see, never on demographics.

Two axes separate them, and both have evidence at both ends. Does anyone else ever open this material — 182 ranked issues across two trackers contain not one about a team, against 6,592 reactions on a request whose stated motive is collaborating with developers on other tools. And how many tools the same material has to serve — 245 reactions for bring it to the editor I already use, against 6,592 for one source for all of them. A third axis, author against consumer, produces the third persona and carries a warning printed on it.

The contest we expected did not happen, and the reason is worth more than the answer.

The plan expected a fight between the person whose pain is sighted — an archive that lands and does not run — and the collector the specification is positioned for. The one practitioner asked says they are the same person:

“Already had the collection, and that’s the point. When it was one file I knew what was in it. At forty files with overlapping instructions I have no working model of what’s active on a given run. The collection created the problem.” *

That is one sentence from one person and it is the only causal claim anybody has made. If it holds, the collector is the breaker, and P1 is not a choice between two candidates but the merge of them.

What would overturn this — stated in advance, so it can happen

• Four more answers that separate the collection from the breakage. If practitioners two to five say the breakage came from a single server on a new laptop and had nothing to do with the size of their collection, the merge collapses, P1 splits back into two personas, and the primary becomes whichever the remaining evidence favours.
• An answer that says adoption is driven by getting material, not by keeping one’s own. Then P3 becomes primary, and the library becomes the centre of the product rather than a browse screen.
• Four more people answering the genie question with show me what ran. Then the primary survives but the core-value sentence in the specification does not.

The bias, printed on the page. Every public source is a person who chose to write in public; the one practitioner interviewed is also a filer, so a single conversation cannot correct it. Reddit is missing. Every number spoken in the interview is a recollection — forty-something files, one in three fresh environments, about half of it — and the respondent flagged his own unreliability twice, unprompted.

A job survives a change of product. A feature does not.

The specific way this goes wrong here is not inventing jobs but laundering the specification into job form. Our spec already names four surfaces and all the mechanics, so “when my set is complete, I want the system to run a validation pass” would be the spec with a when in front of it, and the whole exercise would confirm the spec by construction. Every motivation clause below was machine-checked against a list of our own feature names; twenty-two formulations, zero hits.

Where this differs from the specification, said once. Our spec words the value as assembly with validation. The main job is a transfer job: assembly is what you do in order to move something, and the check is how you find out before the move rather than after. The second contains the first, and it puts the weight on the far side of the handover.

A second candidate survived, and the method’s rule says what that means. Knowing whether any of my material does anything did not fold into the main job — you could close one perfectly and leave the other untouched, and both have ✓ and * behind them. Two surviving main jobs mean two products. It is recorded rather than resolved, because it is the one we cannot build: we run nothing on anyone’s machine.

This is a single-user desktop tool, and the population evidence for a social layer is an absence that was looked for twice. Inventing a social layer to fill the template is the thing the plan told this stage not to do.

Jobs against personas. An importance nobody has told us is ?, never an averaged 2.

1 = it comes up · 2 = it costs them something · 3 = it is a reason they would change how they work. Read the P2 and P3 columns first. No receiver has ever been asked, and after four rounds and five venues none has ever spoken in the first person — every account of a handover is written by the sender. P2 still carries a number in five cells: three of them are second-hand, and two were filled by watching what receivers did in public diffs, because somebody who never posts still leaves one. P3 carries a number in none of the ten rows, and four instruments have now hunted for that person and failed. That is a fact about our instruments rather than about those people, and it is the most useful thing this table produces.

Ten sourced jobs, six hypothesis jobs · wide table, scrolls sideways on a narrow screen

Importance 3 for the primary persona and not closed by the market. After the audit that rule chose one buildable job, not three — and as of 10 Sep 2026 it chooses three again, because two more cells reached a 3.

Updated 10 Sep 2026, and recorded rather than applied. Two more cells reached a 3, so five jobs now score 3 for the keeper — and with one of them a constraint on everything else rather than a thing to build, and one of them the job this product cannot close at all, the rule now selects three buildable jobs by itself. It has stopped needing the judgement the box below was written to make honest. But they are not the same three: RJ-2 enters and RJ-4 is displaced. Changing the shortlist is the owner’s call, not a research round’s, so what follows stands as written.

This section used to say “applied mechanically”, and it no longer can. Three jobs still score 3 for the keeper — the main job, not be quietly overruled, and stop suspecting half of it is dead weight — but the second is a constraint on how everything else behaves rather than a thing to build, and the third is the one job the product cannot close at all. The other two below now score 2 and are kept on stated grounds: the market is open in both, and the specification has already spent a mechanism on each. That is a judgement with the matrix as its input, not a result the matrix produces.

The two other 3s, and neither is a row to build — this matters more than the shortlist

• Not be quietly overruled. It is not a feature to build; it is a constraint on how every other feature behaves, and the specification already encodes it. It belongs in the core as a rule, not as a row.
• Stop suspecting half of it is dead weight. The feature cell is empty because nothing we can build fills it. This is the highest-importance job in the matrix that the product cannot close, and its honest form is a positioning question rather than a backlog item.

What might not be worth building — a list of hypotheses, not a cut list

Read the label before the list. Every feature below closes a job whose every cell is ?. That is not evidence that nobody wants it — it is an absence in instruments that cannot see presence. The trackers are blind to find and reassemble by construction, and nobody in the empty-handed shape has ever been observed. Cutting a feature on this list would be acting on our own blindness. Nothing is removed.

• The scope switch and the curated shelf — closes only H-J4, in the most crowded space in the survey, for a persona who has now been observed for the first time and four of whose five voices refuse, minimise or prefer their own material. The sharpest finding here — and it is about what the shelf holds and how it sorts, not about whether it ships.
• The example project — weaker as a feature, stronger as an argument.
• Duplicating a project — cheap, closes nothing evidenced, and the market closes it in a keystroke. A candidate to postpone, not to cut.
• Promotion of a locally modified item — after the audit it stands on nothing rather than on one aside: the aside cited is about a conflict outside the set, not a wish for a per-project override. It stands or falls with in-project editing.
• Library-wide search and filters — the market’s best-served flow, our thinnest evidence, and priced against a library size nothing supports. Build for fifty, not for three hundred.
• Per-item usage facts — not an orphan. The only specified feature a real person asked for in our own words, unprompted, and then extended.

Every ? above, restated so none of it can be read as a finding.

Read the third column knowing the instrument is gone. Most rows below say four more conversations, and as of 9 Sep 2026 those conversations cannot be run. An unavailable instrument removes an event; it does not lower a bar, and nothing here is promoted because of it. What is reachable without a person was written down as a third round of research and run on 10 Sep 2026: the handover test on our own machine, forks read as handovers — the only way to observe a receiver without asking one — drift measured in git history, whether imported material is ever touched again, and the competing tools installed and pointed at a deliberately broken set. It closed H6, measured the behaviour under H7 and H8 without reaching the person in them, and reached none of the rows about motive or feeling. Where a row below has moved, it says so.

The half-day was spent, and it paid twice. Two rounds of research, six instruments, 19,703 repositories and 2,036 comments could not touch the single most load-bearing claim in the specification — a receiving agent performs the setup from the document alone. On 2026‑09‑10 a third round stopped reading the public record and ran it. The claim holds: three receivers out of three did the work. The half nobody had thought to doubt does not: leaving the faults for the receiver to notice produced one correct diagnosis, one confident wrong one, and one silence. The disclosure the sender gets before export is the disclosure the receiver turns out to need.