Status
startthat.work is the name of the whole design: the catalogue, the process and the rules. It is also the address of the site (decision 57). It replaces "analysis model", which was confused with the language model that does the work. In this document "the model" always means the language model. The skill that runs the process is named analysis (decision 41).
The catalogue lists 236 analytical techniques in 12 families. It is packaged as a skill that a chat or agent calls. The skill is built (a zip with the process, the catalogue, the schema, the prompts and seven scripts). The user has run it once inside Claude, on an article (decision 40 came from that run). Its high-stakes path, personal information waiver and follow-up round have not been run.
Given a brief, the process chooses the techniques that fit, runs them in order, checks their findings and combines them into one answer.
Status:
- Every part has been built and tested on small cases.
- One complete run was done on 13 November 2026, on ten interview quotes, under version 1.2.4 of this design. Steps 1 to 5 ran before decision 38, which changes only step 6. Steps 7 and 8 were run again on 15 November 2026 under version 1.2.8, to show the answer form of decisions 43 and 44, and step 8 once more under version 1.2.10, to show the grouped conclusions of decision 46. Its output is shown in the simulator on the documentation site.
- The documentation site is built from this document, the schema and the catalogue. It also has a page, Use the skill, with the skill to download and the steps to install it in Claude and in GitHub Copilot (decision 45).
- Version 1.2 changes six rules (decisions 24 to 30 in section 9). Versions 1.2.1 to 1.2.23 add decisions 31 to 59. The scripts and the skill follow version 1.2.10 (versions 1.2.11 to 1.2.23 change the site and the record only, except that 1.2.21 changes the name in the skill text). The stored run follows version 1.2.4 for steps 1 to 6, version 1.2.8 for step 7 and version 1.2.10 for step 8. Section 10 lists what is untested.
- All tag values are inference unless a row or a named source says otherwise.
Open items
Changed since the run of 13 November 2026 (version 1.2.4):
-
Version 1.2.6 changed how additional sources are handled and renamed the skill (decisions 40 and 41). Version 1.2.7 added the rule for a run that does not finish (decision 42). Version 1.2.8 changed the answer form and its delivery (decisions 43 and 44). Version 1.2.10 groups the conclusions by support and shows the labels only on request (decision 46). None of these versions has been run inside Claude.
-
The stored run follows version 1.2.4 for steps 1 to 6, except that its plan table was regenerated by the current
sequence.pystep 7 was run again under version 1.2.8 and step 8 under version 1.2.10. The stored run of 13 November 2026 ran on ten interview quotes: the 5.1 findings were shown to the user, who kept the request, and the user chose the combining method. Each technique ran as a separate agent. - The personal information waiver in
select.py(decision 29) has not been tested on content that contains personal information. - Step 5.2 has not been run to the end under decision 40. The stored run needed no additional sources. In the user's run inside Claude an earlier wording searched for a pool of sources before the plan, which the new text forbids.
- The step 8 agent was not given the catalogue ceilings, so the strength of statement reached is worked out from the catalogue by the application. The stored run carried out the spot checks (by code) and the repeat runs (HE1-15 and MM-10). It is at decision-informing stakes, so the checks that need a person were left out and the answer says that no person reviewed the results.
- The high-stakes path has not been run: the pause at the end of step 6 that asks the user to review a sample of the results, and the cap of the label at medium when the user does not.
- The spot check by code covers few pairs for some techniques (0 to 3 for HE1-14, HE1-15, MM-10 and TC1-13), so it says little about them.
Not changed:
- The schema and the catalogue keep their original field names (section 8).
Design:
select.pynotes techniques that need additional sources, andsequence.pylists what to gather. Neither script finds the material. The process does that at run time. The minimum amounts field (requires_minimum) is still not applied by the script.- The deployment settings (highest permitted privacy class, code execution, human reviewer) are in the schema and not in
select.py. - Step 3 recall (22 of 32) was measured before the catalogue changes and the four criteria, and has not been re-run. AO-09 and AO-15 were missed on test job 5. Their use-when text was improved and not retested.
- Origin, the producer's interest and whether sources depend on each other are not inferred from the publisher, so appraisal techniques see unknown.
- The confidence rule, the request confirmation after 5.1 and the follow-up round are untested.
- Not decided: whether the answer includes a section on the process followed by default, or only on request.
- Steps 6 to 8 have run once, on one small brief.
- Result formats are tested on five of 236 techniques.
- The answer form of decisions 43, 44 and 46 has been tested only by running steps 7 and 8 again on the stored run, and step 8 again after decision 46. The user's request for the labels is not yet tested. Delivery of the answer as a document has not been run inside Claude.
- The skill has not been run in GitHub Copilot. Its instructions say to start separate agents and, where that is not possible, to run each technique in turn. How Copilot handles that, and whether it asks for approval before each script, is untested.
- The skill has run once inside Claude, started by the user on an article. Its technique runs, checks, combining and answer steps ran as separate subagents. The output of that run is not stored here, and the changes in versions 1.2.6 to 1.2.10 have not been run there.
- Intake does not read images, audio or video. The rules for document type, language and personal information are untested on long documents, scanned PDFs and non-English text. The rules for strength of statement and kind of finding are keyword matches.
Catalogue data (from the spot check; the first six were fixed):
- Fixed: TC1-08 (claim ceiling), TC-03 (access depth), HE-05 and CK-12 (model suitability), CM1-12 (source count), HE1-11 (consumes appraised evidence base).
- Not fixed: PE-02 (consumes and produces no result artefact), CH-03 and CH-11 (inconsistent consumes and produces), IS-06 (linkage requirement).
- Frame family: FQ-01 to FQ-06 share one identical minimum-checks string, including a second technique at high stakes.
- Input validation is listed for deterministic techniques but used by 23 non-deterministic rows. HE-06 (deterministic) uses repeat runs.
- Only 14 of 236 rows use
not_for_rules. HE1-06 and IS-05 use source dependence, a field that only changes the order, as an exclusion. - About 30 relation targets in other families were chosen from titles or short descriptions.
- No working document type for diagrams, argument maps, personas, a labelled gold set, an event table or a provenance graph. No unit of analysis for actor, case, option, text span or event.
- Review the 9 techniques whose strongest statement is the effect of an action and the 5 whose strongest statement is how or why.
- Some frameworks could not be confirmed (the PHIA technique list, Heuer and Pherson names). Most opened sources were abstracts or secondary pages. No framework specifies checks per stakes level, so that field is entirely inference.