startthat.work home

Decisions

This section is a dated history. The current rules are in sections 2 to 7. Where an entry below differs from those sections, those sections apply.

Conflicts found when the design was consolidated (version 1.0):

# Conflict found Decision
1 The schema called claim ceiling a hard gate. The script uses it as a count and warning, because as a gate it removed structuring and appraisal techniques (8 of 32 expected picks passed, then 30 of 32 after the change). Claim ceiling raises a warning only. schema.json is changed.
2 The schema said to return a shortlist of 5 to 15 by score. A score-ranked list of 15 found 8 of 32 expected picks, and the model reading the full list found 22 of 32. The model reads the list of every eligible technique. Scores order it only.
3 The schema said the model chooses techniques and orders them by the working documents they need. The ordering rule orders by kind. Kind decides the order. Input dependencies apply only among preparation techniques.
4 Frame techniques took the agreed brief as input and produced a draft. Frame techniques produce a finding and do not change the brief. Replaced in part by decision 25.
5 The schema fields used technique_id and technique_name, and lacked five CSV columns. The schema now lists all columns and maps the two names to id and technique. The validator skips the text columns.
6 Family names are capitalised in the CSV, and the schema said value ids are lower case. Family names are labels in sentence case. The convention is reworded.
7 Test jobs used language: English, which is not a schema value. Language is single or multiple. The two jobs are corrected.
8 The Schema report described the older design (234 techniques, three input types replaced, no ordering step and no combining step). This document replaces it. The old report is kept as Schema report (superseded).md.
9 Early wording called selection tags "input values". Material profile (facts about the material) and job settings (question, claim level, stakes, insight). The selection rules are the tags on each technique.
10 Intake first offered only document and claim as units for text, which removed 82 techniques that the hand-written test job kept. Intake offers every unit the material can support.

Decisions made with the user for version 1.1 (11 November 2026):

# Question Decision
11 Does the brief need a specific question? No. The brief has one request, which can have several parts or be open-ended.
12 Should the process loop? Only when the user asks after reading the answer.
13 Should correlated errors between techniques be handled? No. The techniques share one material and one model by design. The answer states this as a limit.
14 What does the process do when a technique needs something the user's material does not contain? It gathers only what that technique needs to run, never material in general.
15 How much should the output say about certainty? One confidence label and a one-line reason per conclusion. Likelihood terms and roll-ups are dropped.
16 Do challenge techniques challenge the brief? No. Only the user changes the brief. Challenge techniques examine the findings, so they run in step 6. Nine techniques (CH-01 to CH-05, CH-08, CH-10 to CH-12) now list the result as an input.
17 Should the strength of statement be limited by what the material can support? No. The answer describes what it is based on as facts and does not judge the material.
18 Should intake ask about classification and licence? No. No technique uses them. Both fields stay in the schema, marked not used.
19 Which checks need a person? The check library states it for each check (field human_involvement). A check that needs a person is not done by the process and is listed as not done.
20 Is there a limit on techniques? Yes. At most 10 in the first round and at most 5 in each follow-up round.
21 Do challenge findings affect confidence? No. A challenge finding is an added finding.
22 Leftover text from the earlier design. Removed: the three-way gate at the end of preparation, the four-output set (claim level achieved, certainty, likelihood term), and the wording "claim level the evidence supports". Changed in part by decision 26.
23 Step 2 judged only the material as given, so a technique that needs a table of coded themes was removed for a text input even though a coding technique could produce it. A content type or structure failure is waived when every prepared input the technique needs is produced by an eligible technique. On the five test jobs this restores 0 to 6 techniques.

Decisions made with the user for version 1.2 (12 November 2026), after a review of the design for logic and wording:

# Problem found Decision
24 "Stage" named three sequences: seven process stages, six plan stages numbered from 0, and seven technique phases. "Brief" had two definitions. One numbered list of eight steps, with sub-steps 5.1 to 5.3. The brief is the request, the settings, the material and the description of the material.
25 Frame techniques ran first, but nothing used their findings until the end. They run first as 5.1. The process shows their findings and the user keeps or edits the request. An edit runs steps 2 to 4 again.
26 The strength of statement was asked for, and the answer never said whether it was reached. The model chooses at least one technique that reaches it when one fits. The answer states which techniques reached it and which conclusions they support.
27 The choice in step 3 had no stated criteria. Four criteria, in order: cover every part of the request, reach the strength needed, match the kind of finding, prefer pairs with different methods.
28 The confidence rule did not cover agreement without checks, conflicts or checks that need a person. Its test for independence could never be met, because every technique shares the material and the model. The rule in section 7. Independence means different methods.
29 A technique that cannot take personal information was removed even when de-identification would run first. It is kept, with a note, when de-identification is in the plan.
30 Rule 1 listed three exceptions and did not cover the verification techniques. Several rules were stated only in this history. Words such as check, review, result and gate had more than one meaning or needed decoding. Five kinds of technique, each defined by what it works on. All current rules are in sections 2 to 7. One term for one thing, listed in section 2, with the schema names mapped in section 8.
31 The design did not say which results the user receives and which are working material. Decided on 13 November 2026. The answer in step 8 is the only output. Everything before it is work, shown as the running application allows. Decision points always show what the user needs to decide.
32 Step 4 showed the sequence under the numbers of later sub-steps, and gave no reason for the order. Decided on 13 November 2026. Steps 3 and 4 are named "Choose analysis techniques" and "Plan analysis sequence". The plan is one sequence with a stated reason for each position, shown without sub-step numbers. The reasons for choosing and sequencing are recorded for a possible section in the answer on the process followed. The kind of technique that runs in 5.3 is renamed from analysis technique to primary technique, because "analysis technique" now means any technique.
33 Step 4 told the user what was or was not gathered from outside, before any technique had run. The design did not say when outside material is gathered or that the user approves it. Decided on 13 November 2026. Step 4 records only which techniques need outside material, as a column in the plan table (group, technique, why the group runs there, needs outside material). Step 5.2 resolves each need: content first, then one item from a source that is not marketing or opinion, or from the user. The user approves any outside item before use. Step 8 reports it. Decision 40 removed the approval.
34 The step 4 table explained each group by saying it runs before or after another, which tells the user nothing, because something has to run first. Decided on 13 November 2026. The table says what each group does for the request, and mentions the order only where it has a consequence for the user (the question can be changed after the first group). The reason for each chosen technique stays in step 3.
35 The group names First and Next only gave the order. Look at your question was vague. The description belonged to the group, which does no work, and not to the techniques. Decided on 13 November 2026. Groups are named by their purpose: Clarify the question, Prepare the content, Analyse the content, Test the findings. The table describes each technique, using the catalogue's description. The plan is work and carries no decision points.
36 One group had a note about rewording the question and the others had none. The plan is work, so the user may not see it, and the note did not tell the user anything at the point of decision. Decided on 13 November 2026. The note is removed. The decision to reword the question is asked in 5.1. The plan carries no decision points.
37 Nothing in the design said what a runner could show about where a run is, so each application would have had to invent states. Decided on 13 November 2026. Section 2 defines the states of a step, a technique, a library check and the answer, and the decision points a step can wait at. How the states are delivered is left to the application, because resuming a run later is not a requirement.
38 A plan at decision-informing stakes listed checks that need a person and the run showed them as not done. The user could not act on them and they read as failures. Whether to ask is a question of stakes, and a person cannot check every decision inside the process. Decided on 14 November 2026. Checks that need a person apply at high stakes only. Below high stakes the plan leaves them out and the answer says that no person reviewed the results. At high stakes the process asks the user to review a sample of the results at the end of step 6, a new decision point. Checks that need no person, including spot checks and repeat runs, are always done by the process.
39 The provisional name ANLYSIS was replaced when the address anlysis.work was registered. The correctly spelt analysis.work is a premium domain. Decided on 14 November 2026. The name is ANLYSIS in running text, to refer to the design, the process and the skill, and ANLYSIS.work when it refers to the site or its address. The GitHub repository keeps the name analysismodel until the site is deployed.
40 The 5.1 question gained an option the design did not have (add outside sources). A search for a pool of 11 sources then ran before the plan existed, and the user was asked to select from the list. The user pointed out that techniques do not need approval for what they need to do, and that everything not typed in by the user is outside, so the word did not separate anything. Decided on 15 November 2026. Sources the process finds are called additional sources, and the user's material is what the user typed, attached or linked. 5.1 offers only keep or edit. 5.2 runs after the plan, finds one item per need, uses it and records it. The approval decision point is removed, leaving six. The answer lists each additional source, and a conclusion that depends on one whose wording code could not check is limited to medium. This supersedes the approval in decision 33.
41 The skill and the site do not have to share a name. Decided on 15 November 2026. The skill is named analysis (folder analysis, file analysis.zip, working folder analysis_run). ANLYSIS stays the name of the design and ANLYSIS.work the site (decision 39).
42 In the user's run inside Claude, a subagent stopped because the user's weekly usage limit in Claude was reached. The design did not say what the process does when a technique's run does not finish, for that or any other reason. Decided on 15 November 2026. The technique is dropped with the reason "did not finish". The process never writes its result itself, tells the user which technique stopped, and lists it among the techniques dropped in the answer.
43 The answer in the user's own run was hard to read. The summary was one short paragraph of compressed sentences. Conclusions were named by codes (T1 to T7) that the answer never explained. A conclusion about what the process had not looked at carried a high label, which read as confidence that the topic was examined. The insights were recommendations, not findings. Decided on 15 November 2026. The answer starts with a summary of two to four short paragraphs. Each conclusion has a title that states the finding, with its label and reason after it. The meaning of the labels is stated in fixed words before the first conclusion, and says the label measures support for the conclusion as written, not importance. A statement about what the process did not look at is not a conclusion: it goes under what was not covered, with no label. An insight is a finding that goes against or beyond what the request or the material would lead a reader to expect, and states the expectation and the finding. Recommendations are given only when the request asks for them.
44 The answer was given as text in the conversation, which is awkward to keep, share or export. Decided on 15 November 2026. Where the application can create a document, the answer is delivered as a document. Otherwise it is given in the reply.
45 The site documented the design, and nothing on it said how to get the skill or install it. Decided on 15 November 2026. The site has a page, Use the skill. It offers analysis.zip for download, gives installation steps for Claude (claude.ai, Claude desktop and Claude Code) and for GitHub Copilot (VS Code, Copilot CLI and the Copilot cloud agent), says what to give the skill, what it asks and what it returns, and says what has and has not been tested. The zip is the only form in which the scripts are published. The page text is data/use-the-skill.md and the zip is data/analysis.zip, both in the site repository, and the build copies the zip to the site.
46 The confidence label (high, medium, low) reads as a statement about the analysis and not about the subject, which makes the answer harder to read. The user wanted conclusions grouped by how well they are supported, with the label shown only on request. Decided on 15 November 2026. The label and its reason are still recorded for every conclusion and insight in step 7. The answer shows conclusions in two groups: well supported (label high) and weaker support (label medium or low). Each conclusion in the second group, and each insight whose label is medium or low, carries a caveat: one or two plain sentences saying why the support is weaker. The answer shows the labels and reasons only when the user asks for them, with the fixed words for the labels. The label rule in section 7 is unchanged. Decision 43's requirement to state the confidence levels before the first conclusion is replaced by the fixed words for the groups.
47 The simulator showed five counters that only tracked how the technique numbers narrowed (the catalogue total, techniques that passed, chosen, placed in sequence, checks that need a person). Decided on 15 November 2026. The simulator shows six measures of the run, each filled in when its step finishes: words of material read, techniques run and how many were dropped, additional sources used, checks done and how many were not done, conclusions to review (those with weaker support) and insights. Each is counted from the stored run. Time taken and tokens used are not recorded by the skill, so the simulator does not show them. This changes the site only.
48 In the simulator the Start again button sat in the bar above the measures, away from the other controls in the left navigation. The conclusions measure read "4 and 5", which does not say what to do with the numbers. Decided on 15 November 2026. Start again is an item under Simulator in the left navigation, shown once a run has started. The conclusions measure shows one number, the conclusions with weaker support, labelled conclusions to review. Decision 47 is otherwise unchanged. This changes the site only.
49 The site explained the model and the project but gave a reader no way to use the catalogue without the skill, no plain statement of what the process does not do, and no list of what changed in each version. Decided on 15 November 2026. Three pages are added. Choose by hand, under Techniques, shows how to use the catalogue with the same steps as the process. Limits states what has been tested, what the answer rests on, the catalogue's weaknesses, what the process cannot read or judge and what is not recorded. Change log lists each version, newest first, in plain terms, with its decisions. The page texts are data/choose-by-hand.md, data/limits.md and data/change-log.md. The build fails if the newest row of the change log is not this document's version. This changes the site only.
50 The page Choose by hand sat under Techniques and stopped after the choice of techniques and the plan of checks. It is a way of running the whole process without the skill, so it belongs next to the skill and has to cover all eight steps. Decided on 16 November 2026. The page is renamed Use by hand and moves to Try it, after Skill. It follows the eight steps in order: write the brief, rule out, choose, plan the sequence, run the techniques (including additional sources and a run that does not finish), check, combine (the methods, the label rule and the rules for insights) and write the answer. It ends with what the reader loses compared with the skill. The Limits page says that running the process by hand has not been tested. The page text is data/use-by-hand.md. This changes the site only.
51 The research behind the catalogue was in two reports and some survey notes that were not on the site. A reader could not see what techniques exist, what evidence supports them, how widely they are used or what the catalogue does not cover. Decided on 21 November 2026. The site has a page, Research findings, under Project record. It combines the two technique reports with research on how widely techniques are used and on the frameworks consultancies publish. It has three levels: a summary, six themes, and 29 findings in six groups, each finding with one caveat and its sources. The last section lists the 236 techniques with their evidence status. The findings are expandable, and each group has its own colour. The page text is data/research-findings.md. This changes the site only.
52 The Research findings page was placed under Project record. Project record is about how the project was built (decisions, test evidence, files). The research is the background input to the catalogue and does not belong with it. Decided on 21 November 2026. The page moves to its own group in the navigation, Research, at /research/, and has a card on the home page. The old address redirects to the new one. Decision 51 is otherwise unchanged. This changes the site only.
53 Decision 52 gave the research its own group in the navigation. The user wanted it in the project section, as its own page next to Status, and not under Project record, which is about how the project was built. Decided on 21 November 2026. The page Research findings is an item in the project section of the navigation, after Status. It stays at /research/ and keeps its card on the home page. Decision 52 is replaced for the placement only. This changes the site only.
54 The navigation called the page Research findings. The user wanted it called Source research, because it is the research the catalogue was built from. Decided on 21 November 2026. The navigation item and the home page card are named Source research. The page keeps its title, its address /research/ and its content. This changes the site only.
55 The navigation started with The model, which suits a reader who wants the design. A new reader wants to see what the process does and to use it first. Decided on 21 November 2026. Try it (Simulator, Skill, Use by hand) is the first group in the navigation, before The model. The order of the previous and next links on each page follows the navigation. This changes the site only.
56 A consistency scan of the site found statements that no longer matched the catalogue or the rules. Technique pages listed checks that need a person at decision-informing stakes without saying that decision 38 leaves them out. The schema named two decision points with old terms (synthesis method, commensurable estimates). Three open items gave counts that had changed or described a problem that was fixed. The Limits page said every evidence status came from the earlier reports, which is not true of AO-19 and PE-14. The Source research page called the five whys documented and showed "Also serves" notes as part of technique names. Decided on 21 November 2026. Technique pages mark a check that needs a person, below high stakes, as left out, and say why. The schema uses the names the skill and the site use: choose the combining method, and confirm that estimates measure the same thing. The open items give the current counts and drop the fixed item. The Limits page names the two techniques. The Source research page is corrected. No rule changes.
57 The name ANLYSIS read as a technical label and gave no sense of what the site is for. The user wanted a casual name that says something about the site, in the form of a short phrase that the .work address completes. Decided on 22 November 2026. The name is startthat.work, and it is also the address of the site. It replaces ANLYSIS and anlysis.work in running text, the site, the simulator, the README, the schema title, the skill text and the design. The skill keeps the name analysis (decision 41). Decisions 39 and 41 and the change-log row for 1.2.5 keep the old name because they record what was decided then. The GitHub repository keeps the name analysismodel. The domain startthat.work has to be registered and added to the Vercel project, and anlysis.work should redirect to it. Both are done outside the repository. This supersedes decision 39.
58 The site header still showed the logo for the old name. The user supplied a logo for startthat.work: the words start and that and the .work ending, with a play mark. The header image text also read startthat.work.work after the rename. Decided on 22 November 2026. The header uses the supplied logo (build/assets/logo.svg) and its text description is the name followed by the word home. The change affects the website only.
59 The home page tab read "startthat.work: documentation", which described the site as a manual. The user wanted the tab to carry only the name, written as the logo writes it. Decided on 22 November 2026. Browser tab titles use STARTthat.WORK, as the logo does. The home page tab shows only that name, and every other page shows its title, a vertical bar and the name. Running text, headings and the address stay startthat.work.