A one-line prompt beats a recipe you can’t reproduce

Organized by role and task—copy, run, verify.

Recipe23 recipes6 of 23 measured

Each recipe has eight fixed sections: scenario, preparation, inputs, full instruction, run log, output, acceptance checklist, failures & boundaries.Run logs stay "to be tested" until we actually run it; no estimated numbers.

Anything marked Verified was actually run here; its log carries duration, result, and leftovers. The rest are still templates: we do not fill numbers for a run we have not done, and each page says what is blocking it. Claim one on the contribute page.

Seven sections

Status and progress are shown plainly.

Recipe / Office100 PDFs, one searchable Markdown indexVerified
Recipe / OfficeAn Excel ledger into a monthly report, with formulas you can recompute and checkVerified, with one caveat
Recipe / AutomationWatch competitor prices weekly, only speak up when something movesNot run
Recipe / OfficeTurn a meeting transcript into a properly formatted .docxVerified
Recipe / MigrationTurn exported group chat into a searchable Markdown archiveNot run
Recipe / CodeHave DSH review the changes and list the risksNot run
Recipe / DataAsk for a SQL in plain language, but verify before runningNot run
Recipe / ReadingCompress a long doc / webpage into a usable summaryNot run
Recipe / CodeProduce a refactor plan before touching codeNot run
Recipe / CodeGive a function, get runnable unit testsNot run
Recipe / TranslationTranslate technical docs between Chinese and English, with a glossaryNot run
Recipe / TriageTurn errors and stack traces into a triage orderVerified
Recipe / ProductTurn a pile of user feedback into a backlog that keeps the original wordingNot run
Recipe / ContentOne long article into multi-platform copy without altering a single factNot run
Recipe / OfficeBatch-rename images and produce a mapping tableNot run
Recipe / AutomationA weekly task: built-in schedule cannot do it, use the OS schedulerNot run
Recipe / ReadingA batch of paper abstracts into a sourced reference tableNot run
Recipe / OperationsCampaign data into a weekly report with anomalies called out separatelyNot run
Recipe / ProductA requirements doc into an acceptance checklist you can tick offVerified
Recipe / ReadingA handout into a revision outline that shows which sections it coversVerified
Recipe / AutomationWhere scheduling actually lives: an optional Web overlay, not a pluginNot run
Recipe / MigrationBringing another tool's old conversations into DSH: where the formats differNot run
Recipe / MigrationWhat migration loses: a per-field account of the metadataNot run

Find a recipe by role

One row per role, with whether it has been run marked on the right. Run ones link to a run log and what the run showed.

Role	Count	Recipes
Office	4	100 PDFs, one searchable Markdown index Verified An Excel ledger into a monthly report, with formulas you can recompute and check Verified, with one caveat Turn a meeting transcript into a properly formatted .docx Verified Batch-rename images and produce a mapping table Not run
Operations	1	Campaign data into a weekly report with anomalies called out separately Not run
Data	1	Ask for a SQL in plain language, but verify before running Not run
Code	3	Have DSH review the changes and list the risks Not run Produce a refactor plan before touching code Not run Give a function, get runnable unit tests Not run
Product	2	Turn a pile of user feedback into a backlog that keeps the original wording Not run A requirements doc into an acceptance checklist you can tick off Verified
Content	1	One long article into multi-platform copy without altering a single fact Not run
Reading	3	Compress a long doc / webpage into a usable summary Not run A batch of paper abstracts into a sourced reference table Not run A handout into a revision outline that shows which sections it covers Verified
Translation	1	Translate technical docs between Chinese and English, with a glossary Not run
Automation	3	Watch competitor prices weekly, only speak up when something moves Not run A weekly task: built-in schedule cannot do it, use the OS scheduler Not run Where scheduling actually lives: an optional Web overlay, not a plugin Not run
Migration	3	Turn exported group chat into a searchable Markdown archive Not run Bringing another tool's old conversations into DSH: where the formats differ Not run What migration loses: a per-field account of the metadata Not run
Triage	1	Turn errors and stack traces into a triage order Verified

Unclaimed: recipes never run, and what is blocking them

Only unrun recipes appear here. Each states its blocker, so claiming one means preparing that condition. Once it is run, the log goes on the recipe page and the status flips from unrun to verified.

Recipes	Role	Blocked on
Watch competitor prices weekly, only speak up when something moves	Automation	Needs real competitor pages with a stable, long-term price source. We have no such target, and a synthetic page cannot reproduce the noise that decides this recipe: caching, A/B tests, and regional pricing.
Turn exported group chat into a searchable Markdown archive	Migration	Needs a real chat export. Every sample on hand contains real names and contact details and cannot go on a public page, while a synthetic chat cannot reproduce the real mess: line breaks, quotes, image placeholders.
Have DSH review the changes and list the risks	Code	Needs a small real changeset containing a real defect. A synthetic repo validates the flow but not whether it finds genuine problems, which is the entire value of this recipe.
Ask for a SQL in plain language, but verify before running	Data	Needs a real schema and real data distribution. Without them the generated SQL can be syntactically right and semantically wrong, making any conclusion worthless.
Compress a long doc / webpage into a usable summary	Reading	Runnable, but 'is the summary good' has no mechanically checkable criterion. Today only length and structure can be verified, not whether key conclusions were dropped. Criteria come first.
Produce a refactor plan before touching code	Code	Same as code-review: needs a real repo with real debt. A refactor plan for synthetic code always reads correctly because that code has no traps.
Give a function, get runnable unit tests	Code	Runnable on a real repo, but 'are these tests any good' must be measured by coverage change and whether they catch a known bug, which means wiring coverage tooling and planting a defect.
Translate technical docs between Chinese and English, with a glossary	Translation	Needs real domain documents plus a native reviewer who can judge the translation. Without a reviewer only format and glossary consistency are checkable, not correctness.
Turn a pile of user feedback into a backlog that keeps the original wording	Product	Needs real multi-source feedback (support chat, comments, tickets). Synthetic feedback reads too uniformly to test whether the merge rule holds when one complaint arrives in ten different phrasings, which is the hard part.
One long article into multi-platform copy without altering a single fact	Content	Runnable, but proving no fact was altered needs a real published long-form piece carrying real figures and names. Synthetic text cannot show whether a number gets quietly swapped.
Batch-rename images and produce a mapping table	Office	Needs a set of images with real content, including a corrupt one, before 'what you actually saw' can be judged. Synthetic colour blocks turn confidence into guesswork.
A weekly task: built-in schedule cannot do it, use the OS scheduler	Automation	Confirmed: the built-in schedule has no calendar rules (the docs state the protocol contains no cron expressions or calendar rules), so weekly work must go through the OS scheduler. What is missing is registering one here and waiting for a real trigger, which takes a week — hence method only.
A batch of paper abstracts into a sourced reference table	Reading	Needs real paper abstracts. Synthetic ones miss the two features that matter: how different study types describe themselves, and the genuine absence of sample sizes — exactly the errors this recipe guards against.
Campaign data into a weekly report with anomalies called out separately	Operations	Needs real consecutive-week campaign data. Without genuine variance there is no way to see whether the anomaly rule under- or over-reports, which is this recipe's only substance.
Where scheduling actually lives: an optional Web overlay, not a plugin	Automation	Runnable, but end-to-end verification needs the Schedule overlay enabled first: create a reminder, wait for it, confirm it returns to the original session. Not loaded here yet.
Bringing another tool's old conversations into DSH: where the formats differ	Migration	The DSH side is measured (path escaping, multi-frame zstd, header and event shapes); what is missing is a real sample from the other side. With no Claude Code or Codex session file on hand, the mapping can be reasoned but not measured.
What migration loses: a per-field account of the metadata	Migration	Same: the DSH session header is measured (type, version, id, createdAt, cwd, parentSession, origin, delegationDepth, agentPreset, with seq and time per event line). The other side's field list is what is missing for a per-field comparison.

Claim this recipe