diff --git a/CHANGELOG.md b/CHANGELOG.md index 28c2fdc..535c1cf 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -195,6 +195,25 @@ and [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ### Changed +- `de-slop` and `compressed-deliberation` say less about themselves. A reading + of the craft tier with both skills open found each file spending about half + its words on its own provenance and its own limits. Neither loses a rule. + `de-slop` keeps one clause of the reason it ships no word list, and the + repository-governance case behind that decision moves to + `source/craft/de-slop.md`. Its `What stands behind this shape` section merges + into `Boundary`, which already stated the same no-standard proposition. + `compressed-deliberation` loses the seven bullets that described Claude Opus + 5, and keeps the sentence a reader acts on, which is that the rules aim at + the whole stack rather than at the model alone. Those bullets move to + `source/craft/compressed-deliberation.md`, beside the URLs and read dates + that already held them, so the provenance survives the move. One clause is + cut rather than moved. It said that files written to disk are the case the + vendor documents most clearly, and it attributed that to a record which never + said it. Two headings over two bullets each become labelled bullets under + `The shape of a reply`. + The `compressed-deliberation` matrix therefore carries no `G` row, and it + drops its `Source version` line, because ADR-0026 refuses that line in a + matrix that cites nothing. Both matrices renumber in document order. - `proportionate-execution` says how it differs from `de-slop`. Three of the four craft skills already related themselves to a neighbour, and `de-slop` has named this one since 0.3.0, so the closest pair was cross-referenced diff --git a/grounding/craft/compressed-deliberation.md b/grounding/craft/compressed-deliberation.md index b38926b..9b04730 100644 --- a/grounding/craft/compressed-deliberation.md +++ b/grounding/craft/compressed-deliberation.md @@ -9,38 +9,39 @@ Disposes of every unit of content in claims no authority at all. Its rule cell is empty. The `Audited` cell of a `G` row says whether a person has read that row against -the source. Every row starts at `unaudited`, and no run of the checker raises -it. A person who checks a row writes the date and the row's digest in place of -the word. Editing any other cell changes that digest, so the audit goes stale -and the check says so. So does moving the source version this file declares, -because an audit answers to one reading of the source. +the source. The `Source text` cell carries the words of the rule it cites. +There is no `G` row here, so every cell in both columns is empty, and both are +waiting for the first row that cites anything. -The `Source text` cell of a `G` row carries the words of the rule it cites, in -quotation marks. Every row below reads `unquoted`, and that is not a backlog. - -**Quotation:** forbidden. The source record states that nobody has read the -governing terms of the vendor pages these rows cite, so no sentence of one may -be carried here. Read those terms and record the check, then edit this line. +**Quotation:** forbidden. Nobody has read the governing terms of the vendor +pages behind this skill, so no sentence of one may be carried here. The source +record at `source/craft/compressed-deliberation.md` names those pages. Read +their terms, record the check there, and edit this line in the same pass. This file stays in the repository. It does not install with the skill. Checked by `stylewright ground --check --skill compressed-deliberation`. -## What the `G` rows are, and what they are not +## The vendor material left the skill, and the `G` rows went with it -The seven `G` rows below trace to Anthropic's own documentation of Claude Opus -5. They are the **problem statement** of this skill and nothing else. A vendor -saying that its model writes at length licenses a description of the model. It -licenses no rule about how anybody should write. +Seven `G` rows used to carry Anthropic's documentation of Claude Opus 5 into +the skill as its problem statement. They described a model, they licensed no +rule, and a paragraph beside them said so. Together they were most of one +section and none of the guidance. -Every prescription here is therefore an `E` row, including the ones that read as -obvious consequences of a `G` row. The citations look authoritative, and that is -why the line has to hold. +That description now sits in `source/craft/compressed-deliberation.md`, which +already held the URLs, the sections read, the read date and the expiry +conditions. The skill keeps the sentence a reader acts on, which is that the +rules aim at the whole stack rather than at the model alone. So no row here +cites anything, and this file declares no source version, because a pin no row +answers to is a line nobody maintains. -The source record at `source/craft/compressed-deliberation.md` carries the -URLs, the sections read, the read date, and the expiry conditions. +Read that as a change of scope rather than a retraction. The vendor pages still +say what they said, and the source record still cites them. A contributor who +wants one cited here again adds the `G` row, the source version and the +quotation check in one pass. -## Why no `G` row supports a word budget +## Why no rule here sets a word budget Anthropic also published a postmortem in which brevity instructions to an earlier Claude Code build reduced coding quality. That report is `A6` in the @@ -72,7 +73,7 @@ Read that as a fact about a harness, not about a model. `claude -p` is the Claude Code headless path, so "no operator instructions" is not "the stock model", and one author's rule stack is not operator instructions in general. Nothing here was measured in the agentic, tool-using, or file-writing regimes -that `G-02` and `G-03` name as the defect's worst ground, and those are the +that the source record names as the defect's worst ground, and those are the regimes the skill is meant for. **What did not survive.** Two adversarial reviews on 2026-07-27 struck more of @@ -92,7 +93,7 @@ this record than they left. - *Cost fell without cutting findings.* Weakened. Detection of the seeded bug is five of five in every arm, but the skill's samples raise one distinct issue where the unguided samples raise two or three. What it drops is secondary and - arguably out of scope under `E-09`. That judgment was made after seeing the + arguably out of scope under `E-13`. That judgment was made after seeing the divergence, so it is not evidence. - *One rule carried most of the cost.* Holds for the largest contributor only. The ordering among the remaining five rules is one draw from overlapping @@ -102,50 +103,33 @@ Five runs per cell, medians with no dispersion statistic, four prompts written by the person who wrote the rules, one seeded bug in one scenario. Nothing here is a controlled trial, and no `E` row should be read as one. -**Source version:** Anthropic's documentation of Claude Opus 5, model id -`claude-opus-5`, released 2026-07-24, as read up to the evidence cutoff of -2026-07-27. The target is a model build rather than a standard, so the build -and the cutoff are what a reading of these pages is a reading of. - | ID | Our guidance | Our anchor | Source rule | Source text | Source location | Audited | |---|---|---|---|---|---|---| -| G-01 | Claude Opus 5 writes longer visible responses than earlier Opus models. | What this corrects | Response length and verbosity | unquoted | A1, Prompting Claude Opus 5 | unaudited | -| G-02 | It writes longer reports, summaries, and files to disk. | What this corrects | Written deliverable length | unquoted | A1, Prompting Claude Opus 5 | unaudited | -| G-03 | It announces what it is about to do more often during agentic work. | What this corrects | User-facing progress updates | unquoted | A1, Prompting Claude Opus 5 | unaudited | -| G-04 | It verifies its own work without being asked. | What this corrects | Task scope and over-verification | unquoted | A1, Prompting Claude Opus 5 | unaudited | -| G-05 | It narrates its own corrections more often. | What this corrects | Self-correction | unquoted | A1, Prompting Claude Opus 5 | unaudited | -| G-06 | It can widen a task past what the reader asked for. | What this corrects | Task scope and over-verification | unquoted | A1, and A5 System Card pp. 86 to 87 | unaudited | -| G-07 | A lower effort setting does not reliably shorten the visible answer. | What this corrects | Response length and verbosity | unquoted | A1, and A3 migration guide | unaudited | -| E-01 | Lead with the result, and not with the route you took to it. | The shape of a reply | | | Our own guidance | | -| E-02 | Put the item that changes the reader's decision first, and let the rest go unsaid. | The shape of a reply | | | Our own guidance | | -| E-03 | Say each thing once, in the place where the reader can act on it. | The shape of a reply | | | Our own guidance | | -| E-04 | Write a claim as a sentence, and never as a section. | The shape of a reply | | | Our own guidance | | -| E-05 | Do not restate the request, the context, or anything the reader just supplied. | The shape of a reply | | | Our own guidance | | -| E-06 | Stop when the result and the support it rests on are on the page. | The shape of a reply | | | Our own guidance | | -| E-07 | Report a finished change as the result, and then the evidence that the result holds. | Reporting finished work | | | Our own guidance | | -| E-08 | Name what you verified in one clause, and not under a heading. | Reporting finished work | | | Our own guidance | | -| E-09 | Report an unchecked risk only when it could bite, and lead with it. | Reporting finished work | | | Our own guidance | | -| E-10 | Concede in the first sentence, and do not argue the point again. | Conceding a correction | | | Our own guidance | | -| E-11 | Say what the correction changes downstream, and then stop. | Conceding a correction | | | Our own guidance | | -| E-12 | Surface a decision only when you will stop and wait for the answer. | Surfacing a decision | | | Our own guidance | | -| E-13 | Give the options, the recommendation, and the belief the recommendation rests on. | Surfacing a decision | | | Our own guidance | | +| E-01 | Read the rules below as aimed at the whole stack you are running inside, and not at the model alone. Our own baseline found an unguided run already clean, and an operator instruction stack above it was what inflated the reply. The source record carries the vendor's documentation of the defaults this skill answers, and that documentation describes a model rather than prescribing anything. | What this corrects | | | Our own guidance, and it instructs the reader what the rules aim at | | +| E-02 | Apply the rules to an agentic session, or to a file written to disk, on their reasoning and not on our evidence. What we measured is single-turn replies, one prompt each, five runs, so those two cases are the ones we have tested least. | What this corrects | | | Our own guidance, and it states the evidence the rules rest on | | +| E-03 | Lead with the result, and not with the route you took to it. | The shape of a reply | | | Our own guidance | | +| E-04 | Put the item that changes the reader's decision first, and let the rest go unsaid. | The shape of a reply | | | Our own guidance | | +| E-05 | Say each thing once, in the place where the reader can act on it. | The shape of a reply | | | Our own guidance | | +| E-06 | Write a claim as a sentence, and never as a section. | The shape of a reply | | | Our own guidance | | +| E-07 | Do not restate the request, the context, or anything the reader just supplied. | The shape of a reply | | | Our own guidance | | +| E-08 | Stop when the result and the support it rests on are on the page. | The shape of a reply | | | Our own guidance | | +| E-09 | **Conceding a correction.** Concede in the first sentence, do not argue the point again, say what the correction changes downstream, and then stop. | The shape of a reply | | | Our own guidance, for one situation a reply is written in | | +| E-10 | **Surfacing a decision.** Surface one only when you will stop and wait for the answer. Give the options, the recommendation, and the belief it rests on. | The shape of a reply | | | Our own guidance, for one situation a reply is written in | | +| E-11 | Report a finished change as the result, and then the evidence that the result holds. | Reporting finished work | | | Our own guidance | | +| E-12 | Name what you verified in one clause, and not under a heading. | Reporting finished work | | | Our own guidance | | +| E-13 | Report an unchecked risk only when it could bite, and lead with it. | Reporting finished work | | | Our own guidance | | | E-14 | Do not set a word budget, because a budget cuts substance before it cuts scaffolding. | What this skill does not ask for | | | Our own guidance | | | E-15 | Do not drop a finding to make a reply shorter. | What this skill does not ask for | | | Our own guidance | | | E-16 | Do not suppress the narration that a reader uses to intervene. | What this skill does not ask for | | | Our own guidance | | -| N-01 | Deliberation is not the product. Monitoring, qualifying, verifying, managing scope and correcting course are all real work, and a reader wants the edited result of that work rather than a transcript of it. | Purpose | | | Purpose framing, asserts no rule | | -| N-02 | This skill is pinned to one model build. The record of what it targets, and of when it expires, lives in the stylewright repository at `source/craft/compressed-deliberation.md`. It is not installed with this skill. | Purpose | | | Points at the source record, asserts no rule | | -| N-03 | Anthropic publishes these as differences in Claude Opus 5. They describe the model. They prescribe nothing, and no rule below inherits their authority. | What this corrects | | | States what the G rows do NOT license | | -| E-17 | Those are the documented defaults. What our own baseline found is narrower and partly different: an unguided run was already clean, and an operator instruction stack above it was what inflated the reply. So read the rules below as aimed at the whole stack you are running inside, and not at the model alone. | What this corrects | | | Records our baseline, and instructs how to read the rules | | -| E-18 | Where the rules have been measured: single-turn replies, one prompt each, five runs. Not agentic sessions, and not files written to disk. The second is the case that record names as the one the vendor documents most clearly, and the one we have tested least, so apply the rules there on their reasoning, not on our evidence. | What this corrects | | | States the evidence, and instructs where to apply the rules | | -| E-19 | `de-slop` treats prose that reads as machine-written, whatever produced it. This skill treats one model's documented defaults, and it expires when that model does. Follow both. They do not disagree. | How this differs from de-slop | | | Relates two skills, and instructs the reader to follow both | | -| N-07 | compressed-deliberation | compressed-deliberation | | | Section title, asserts no rule | | -| N-08 | Purpose | Purpose | | | Section title, asserts no rule | | -| N-09 | What this corrects | What this corrects | | | Section title, asserts no rule | | -| N-10 | The shape of a reply | The shape of a reply | | | Section title, asserts no rule | | -| N-11 | Reporting finished work | Reporting finished work | | | Section title, asserts no rule | | -| N-12 | Conceding a correction | Conceding a correction | | | Section title, asserts no rule | | -| N-13 | Surfacing a decision | Surfacing a decision | | | Section title, asserts no rule | | -| N-14 | What this skill does not ask for | What this skill does not ask for | | | Section title, asserts no rule | | -| N-15 | How this differs from de-slop | How this differs from de-slop | | | Section title, asserts no rule | | -| N-16 | Boundary | Boundary | | | Section title, asserts no rule | | -| N-17 | Only the section named "What this corrects" traces to a published source, and it traces for description alone. Every rule in this skill is our own editorial guidance. The trace lives in the stylewright repository at `grounding/craft/compressed-deliberation.md`. It is not installed with this skill. | Boundary | | | Describes the trace, asserts no rule | | +| E-17 | `de-slop` treats prose that reads as machine-written, whatever produced it. This skill treats one model's documented defaults, and it expires when that model does. Follow both. They do not disagree. | How this differs from de-slop | | | Our own guidance, and it instructs the reader to follow both | | +| N-01 | compressed-deliberation | compressed-deliberation | | | Section title, asserts no rule | | +| N-02 | Purpose | Purpose | | | Section title, asserts no rule | | +| N-03 | Deliberation is not the product. Monitoring, qualifying, verifying, managing scope and correcting course are all real work, and a reader wants the edited result of that work rather than a transcript of it. | Purpose | | | Purpose framing, asserts no rule | | +| N-04 | This skill is pinned to one model build. The record of what it targets, and of when it expires, lives in the stylewright repository at `source/craft/compressed-deliberation.md`. It is not installed with this skill. | Purpose | | | Points at the source record, asserts no rule | | +| N-05 | What this corrects | What this corrects | | | Section title, asserts no rule | | +| N-06 | The shape of a reply | The shape of a reply | | | Section title, asserts no rule | | +| N-07 | Reporting finished work | Reporting finished work | | | Section title, asserts no rule | | +| N-08 | What this skill does not ask for | What this skill does not ask for | | | Section title, asserts no rule | | +| N-09 | How this differs from de-slop | How this differs from de-slop | | | Section title, asserts no rule | | +| N-10 | Boundary | Boundary | | | Section title, asserts no rule | | +| N-11 | Every rule in this skill is our own editorial guidance, and no part of it traces to a published source. The trace lives in the stylewright repository at `grounding/craft/compressed-deliberation.md`. It is not installed with this skill. | Boundary | | | Describes the trace, asserts no rule | | diff --git a/grounding/craft/de-slop.md b/grounding/craft/de-slop.md index ba878e2..d8c60df 100644 --- a/grounding/craft/de-slop.md +++ b/grounding/craft/de-slop.md @@ -30,9 +30,9 @@ row, including the ones that read as obvious. A vendor page reporting that its own model writes at length would not change that. Such a page describes a model and prescribes nothing, which is the -distinction `grounding/craft/compressed-deliberation.md` records for the seven -`G` rows it does carry. A contributor who finds a real source for a rule here -adds the source record before adding the row. +distinction `grounding/craft/compressed-deliberation.md` records for the vendor +material its own skill used to carry. A contributor who finds a real source for +a rule here adds the source record before adding the row. `E-16` is the row where the `G` question was live, and it is the one row here that carries a fact we did not establish. Anthropic reported the Claude Code @@ -43,13 +43,18 @@ the defect the `Audited` column exists to catch, one column over. So the row is `E`, the unit states the report as a report, and the source record states where it came from and what it would take to promote the row. -## The body was rewritten, so the IDs start again +## The IDs restart whenever the body is rewritten An earlier draft of this skill was a list of prohibitions, and the owner ruled that this failure type takes a positive recipe instead. Almost no unit survived -that rewrite unchanged, so the numbering restarts in document order rather than -preserving the gaps an earlier two-row edit had left. A reviewer holding IDs -from the prohibition draft should read this table fresh. +that rewrite unchanged, so the numbering restarted in document order rather +than preserving the gaps an earlier two-row edit had left. + +A later editorial pass cut the word-list argument back to one paragraph, merged +`What stands behind this shape` into `Boundary`, and merged two paragraphs of +the comparison section into one. The numbering restarted again, on the same +rule. A reviewer holding IDs from either earlier draft should read this table +fresh. ## Where the line between an `E` row and an `N` row falls here @@ -62,11 +67,12 @@ is each departure beside them, because a departure names a defect in the reader's passage even though it is written as description rather than as a prohibition. The form changed and the authority did not. -And `No part of this matches a word` grades almost entirely `N`, because it -argues for a decision about this repository rather than telling a writer -anything. `E-19` is the exception, because "a word becomes a rule in a skill -only after a promoted study says it should" is a rule this repository follows, -and ADR-0021 is where it was decided. +And `No part of this matches a word` grades `N` throughout, because what +survives there describes this skill and argues for a decision about this +repository rather than telling a writer anything. The rule that a word becomes +a rule in a skill only after a promoted study says it should is a rule this +repository follows, and it now lives in `source/craft/de-slop.md`, beside +ADR-0021 where it was decided. ## No measurement stands behind these rows @@ -96,28 +102,22 @@ status. | E-10 | **The claim.** An invented counterposition gives an answer the shape of an argument without the substance of one. A stack of qualifications buries the one that would change what the reader does. A passive verb with no actor leaves nobody holding the claim. | What a departure looks like | | | Our own guidance, the departure from E-03 | | | E-11 | **Earned structure.** A heading over two sentences, and a sentence announcing that a point is important, each perform a thoroughness the content has not reached. | What a departure looks like | | | Our own guidance, the departure from E-04 | | | E-12 | **The end.** A closing paragraph that re-says the page gives the reader a second reading of the first. | What a departure looks like | | | Our own guidance, the departure from E-05 | | -| E-13 | One test settles most of them. Remove the sentence, read the passage again, and keep the sentence only if something changed. | What a departure looks like | | | Our own guidance, and it is the one mechanical test here | | +| E-13 | Remove the sentence, read the passage again, and keep the sentence only if something changed. | What a departure looks like | | | Our own guidance, and it is the one mechanical test here | | | E-14 | Compression has a cost, so name it | Compression has a cost, so name it | | | Our own guidance, written as a section title | | | E-15 | The shape above is not a shorter passage. It is a passage whose length is its content. | Compression has a cost, so name it | | | Our own guidance, and it refuses a reading of the shape | | -| E-16 | Anthropic reported in April 2026 that a system instruction imposing strict brevity in Claude Code reduced coding quality, and reverted it. That is a first-party report of compression paid for in correctness. So the shape above sets no length, and a finding that a shorter passage would have dropped stays in. | Compression has a cost, so name it | | | Our own guidance, on a report SOURCE.md records as unread here | | -| E-17 | Counting belongs in `bench/score.mjs` in the stylewright repository, where a scorer counts and asserts nothing. A word becomes a rule in a skill only after a promoted study says it should. | No part of this matches a word | | | Our own guidance, and ADR-0021 decided it | | -| E-18 | Read the shape here as discipline that we assert. Do not read any of it as an effect that we observed, and do not let a later summary say that this skill works. | What stands behind this shape | | | Our own guidance, and it instructs the reader how to read the skill | | -| E-19 | The disagreement is about repetition. Both standards ask a writer to repeat, in a summary, in a heading, or in one term used for one thing. This skill ends a passage at its last load-bearing sentence. A procedure and a safety notice repeat on purpose, so the standard wins there and this skill yields. | How this differs from the other skills here | | | Our own guidance, and it rules on which skill wins | | -| E-20 | `compressed-deliberation` treats one model's documented defaults, and it expires when that model does. `proportionate-execution` governs a session, so it holds what an agent does as well as what it writes. This skill treats one passage of prose, whatever produced it, and it carries no model pin. Follow all three. They do not disagree. | How this differs from the other skills here | | | Our own guidance, and it instructs the reader to follow all three | | +| E-16 | Anthropic reported in April 2026 that a system instruction imposing strict brevity in Claude Code reduced coding quality, and reverted it. That is a first-party report of compression paid for in correctness. So the shape above sets no length, and a finding that a shorter passage would have dropped stays in. | Compression has a cost, so name it | | | Our own guidance, on a report the source record records as unread here | | +| E-17 | `plain-language` and `simplified-technical-english` each distil a published standard and write for a named reader. This skill does neither, and the disagreement is about repetition. Both standards ask a writer to repeat, in a summary, in a heading, or in one term used for one thing. This skill ends a passage at its last load-bearing sentence. A procedure and a safety notice repeat on purpose, so the standard wins there and this skill yields. | How this differs from the other skills here | | | Our own guidance, and it rules on which skill wins | | +| E-18 | `compressed-deliberation` treats one model's documented defaults, and it expires when that model does. `proportionate-execution` governs a session, so it holds what an agent does as well as what it writes. This skill treats one passage of prose, whatever produced it, and it carries no model pin. Follow all three. They do not disagree. | How this differs from the other skills here | | | Our own guidance, and it instructs the reader to follow all three | | +| E-19 | Read the shape here as discipline that we assert. Do not read any of it as an effect that we observed, and do not let a later summary say that this skill works. | Boundary | | | Our own guidance, and it instructs the reader how to read the skill | | | N-01 | de-slop | de-slop | | | Section title, asserts no rule | | | N-02 | Purpose | Purpose | | | Section title, asserts no rule | | | N-03 | The defect this skill treats is a rhetorical move rather than a word, so this skill gives a shape to write toward rather than a vocabulary to avoid. A move survives a model release. Whether a word does is not settled. | Purpose | | | States what this skill treats and what form it takes, asserts no rule | | | N-04 | The shape of a finished passage | The shape of a finished passage | | | Section title, asserts no rule | | | N-05 | What a departure looks like | What a departure looks like | | | Section title, asserts no rule | | | N-06 | No part of this matches a word | No part of this matches a word | | | Section title, asserts no rule | | -| N-07 | This skill ships no word list. | No part of this matches a word | | | Describes this skill, asserts no rule | | -| N-08 | A word that one setting overuses is countable, and that makes it tempting to ship. Two things argue against carrying it in a skill. A list of forbidden words teaches an agent to swap each one for its nearest neighbour, which leaves the defect and cleans the surface. A word may also recur in a setting rather than in a model, which is the live reading here rather than a finding. On that reading a shipped list dates faster than the shape above. | No part of this matches a word | | | Argues for the decision above at the strength ADR-0021 states it, asserts no rule | | -| N-09 | What a check can see | What a check can see | | | Section title, asserts no rule | | -| N-10 | Nothing here, today. `stylewright lint` carries no check for any part of the shape above, because a triad, a restatement and an invented objection are judgments about content rather than shapes a program recognises. | What a check can see | | | States what the tool does not do, asserts no rule | | -| N-11 | So this skill is the generative half on its own. | What a check can see | | | Describes how this skill reaches a reader, asserts no rule | | -| N-12 | What stands behind this shape | What stands behind this shape | | | Section title, asserts no rule | | -| N-13 | Nothing measured. A craft rule has no standard behind it, so measurement is the only evidence it can ever have, and no arm has been run for this skill. | What stands behind this shape | | | States the evidence, asserts no rule | | -| N-14 | How this differs from the other skills here | How this differs from the other skills here | | | Section title, asserts no rule | | -| N-15 | `plain-language` and `simplified-technical-english` each distil a published standard, and each writes for a named reader. This skill has no standard and no named reader. | How this differs from the other skills here | | | Relates three skills, asserts no rule | | -| N-16 | Boundary | Boundary | | | Section title, asserts no rule | | -| N-17 | This skill has no external standard behind it. Every rule in it is our own editorial guidance, and the trace marks the rest as narrative that asserts no rule. The trace lives in the stylewright repository at `grounding/craft/de-slop.md`. It is not installed with this skill. | Boundary | | | Describes the trace, asserts no rule | | +| N-07 | This skill ships no word list. A list of forbidden words teaches an agent to swap each one for its nearest neighbour, which leaves the defect and cleans the surface. | No part of this matches a word | | | Describes this skill and argues for the decision at the strength ADR-0021 states it, asserts no rule | | +| N-08 | What a check can see | What a check can see | | | Section title, asserts no rule | | +| N-09 | Nothing here, today. `stylewright lint` carries no check for any part of the shape above, because a triad, a restatement and an invented objection are judgments about content rather than shapes a program recognises. | What a check can see | | | States what the tool does not do, asserts no rule | | +| N-10 | How this differs from the other skills here | How this differs from the other skills here | | | Section title, asserts no rule | | +| N-11 | Boundary | Boundary | | | Section title, asserts no rule | | +| N-12 | This skill has no external standard behind it, and no arm has been run for it. Every rule in it is our own editorial guidance, and measurement is the only evidence a craft rule can ever have. The trace marks the rest as narrative that asserts no rule. It lives in the stylewright repository at `grounding/craft/de-slop.md`, and it is not installed with this skill. | Boundary | | | States the evidence and describes the trace, asserts no rule | | diff --git a/grounding/craft/subagent-returns.md b/grounding/craft/subagent-returns.md index ed127a8..247e98d 100644 --- a/grounding/craft/subagent-returns.md +++ b/grounding/craft/subagent-returns.md @@ -30,9 +30,9 @@ as obvious. A vendor page reporting how much a model writes would not change that. Such a page describes a model and prescribes nothing, which is the distinction -`grounding/craft/compressed-deliberation.md` records for the seven `G` rows it -does carry. A contributor who finds a real source for a rule here adds the -source record before adding the row. +`grounding/craft/compressed-deliberation.md` records for the vendor material +its own skill used to carry. A contributor who finds a real source for a rule +here adds the source record before adding the row. ## Where the line between an `E` row and an `N` row falls here diff --git a/skills/craft/compressed-deliberation/SKILL.md b/skills/craft/compressed-deliberation/SKILL.md index c7ea561..1101963 100644 --- a/skills/craft/compressed-deliberation/SKILL.md +++ b/skills/craft/compressed-deliberation/SKILL.md @@ -17,27 +17,16 @@ when it expires, lives in the stylewright repository at ## What this corrects -Anthropic publishes these as differences in Claude Opus 5. They describe the -model. They prescribe nothing, and no rule below inherits their authority. - -- Claude Opus 5 writes longer visible responses than earlier Opus models. -- It writes longer reports, summaries, and files to disk. -- It announces what it is about to do more often during agentic work. -- It verifies its own work without being asked. -- It narrates its own corrections more often. -- It can widen a task past what the reader asked for. -- A lower effort setting does not reliably shorten the visible answer. - -Those are the documented defaults. What our own baseline found is narrower and -partly different: an unguided run was already clean, and an operator instruction -stack above it was what inflated the reply. So read the rules below as aimed at -the whole stack you are running inside, and not at the model alone. - -Where the rules have been measured: single-turn replies, one prompt each, five -runs. Not agentic sessions, and not files written to disk. The second is the -case that record names as the one the vendor documents most clearly, and the one -we have tested least, so apply the rules there on their reasoning, not on our -evidence. +Read the rules below as aimed at the whole stack you are running inside, and +not at the model alone. Our own baseline found an unguided run already clean, +and an operator instruction stack above it was what inflated the reply. The +source record carries the vendor's documentation of the defaults this skill +answers, and that documentation describes a model rather than prescribing +anything. + +Apply the rules to an agentic session, or to a file written to disk, on their +reasoning and not on our evidence. What we measured is single-turn replies, one +prompt each, five runs, so those two cases are the ones we have tested least. ## The shape of a reply @@ -47,6 +36,8 @@ evidence. - Write a claim as a sentence, and never as a section. - Do not restate the request, the context, or anything the reader just supplied. - Stop when the result and the support it rests on are on the page. +- **Conceding a correction.** Concede in the first sentence, do not argue the point again, say what the correction changes downstream, and then stop. +- **Surfacing a decision.** Surface one only when you will stop and wait for the answer. Give the options, the recommendation, and the belief it rests on. ## Reporting finished work @@ -54,16 +45,6 @@ evidence. - Name what you verified in one clause, and not under a heading. - Report an unchecked risk only when it could bite, and lead with it. -## Conceding a correction - -- Concede in the first sentence, and do not argue the point again. -- Say what the correction changes downstream, and then stop. - -## Surfacing a decision - -- Surface a decision only when you will stop and wait for the answer. -- Give the options, the recommendation, and the belief the recommendation rests on. - ## What this skill does not ask for - Do not set a word budget, because a budget cuts substance before it cuts scaffolding. @@ -78,7 +59,6 @@ does. Follow both. They do not disagree. ## Boundary -Only the section named "What this corrects" traces to a published source, and it -traces for description alone. Every rule in this skill is our own editorial -guidance. The trace lives in the stylewright repository at +Every rule in this skill is our own editorial guidance, and no part of it +traces to a published source. The trace lives in the stylewright repository at `grounding/craft/compressed-deliberation.md`. It is not installed with this skill. diff --git a/skills/craft/de-slop/SKILL.md b/skills/craft/de-slop/SKILL.md index 16f4464..58b02ee 100644 --- a/skills/craft/de-slop/SKILL.md +++ b/skills/craft/de-slop/SKILL.md @@ -32,8 +32,8 @@ against. - **Earned structure.** A heading over two sentences, and a sentence announcing that a point is important, each perform a thoroughness the content has not reached. - **The end.** A closing paragraph that re-says the page gives the reader a second reading of the first. -One test settles most of them. Remove the sentence, read the passage again, and -keep the sentence only if something changed. +Remove the sentence, read the passage again, and keep the sentence only if +something changed. ## Compression has a cost, so name it @@ -48,18 +48,9 @@ in. ## No part of this matches a word -This skill ships no word list. - -A word that one setting overuses is countable, and that makes it tempting to -ship. Two things argue against carrying it in a skill. A list of forbidden -words teaches an agent to swap each one for its nearest neighbour, which leaves -the defect and cleans the surface. A word may also recur in a setting rather -than in a model, which is the live reading here rather than a finding. On that -reading a shipped list dates faster than the shape above. - -Counting belongs in `bench/score.mjs` in the stylewright repository, where a -scorer counts and asserts nothing. A word becomes a rule in a skill only after -a promoted study says it should. +This skill ships no word list. A list of forbidden words teaches an agent to +swap each one for its nearest neighbour, which leaves the defect and cleans the +surface. ## What a check can see @@ -67,25 +58,12 @@ Nothing here, today. `stylewright lint` carries no check for any part of the shape above, because a triad, a restatement and an invented objection are judgments about content rather than shapes a program recognises. -So this skill is the generative half on its own. - -## What stands behind this shape - -Nothing measured. A craft rule has no standard behind it, so measurement is the -only evidence it can ever have, and no arm has been run for this skill. - -Read the shape here as discipline that we assert. Do not read any of it as an -effect that we observed, and do not let a later summary say that this skill -works. - ## How this differs from the other skills here `plain-language` and `simplified-technical-english` each distil a published -standard, and each writes for a named reader. This skill has no standard and no -named reader. - -The disagreement is about repetition. Both standards ask a writer to repeat, in -a summary, in a heading, or in one term used for one thing. This skill ends a +standard and write for a named reader. This skill does neither, and the +disagreement is about repetition. Both standards ask a writer to repeat, in a +summary, in a heading, or in one term used for one thing. This skill ends a passage at its last load-bearing sentence. A procedure and a safety notice repeat on purpose, so the standard wins there and this skill yields. @@ -97,7 +75,12 @@ three. They do not disagree. ## Boundary -This skill has no external standard behind it. Every rule in it is our own -editorial guidance, and the trace marks the rest as narrative that asserts no -rule. The trace lives in the stylewright repository at -`grounding/craft/de-slop.md`. It is not installed with this skill. +This skill has no external standard behind it, and no arm has been run for it. +Every rule in it is our own editorial guidance, and measurement is the only +evidence a craft rule can ever have. The trace marks the rest as narrative that +asserts no rule. It lives in the stylewright repository at +`grounding/craft/de-slop.md`, and it is not installed with this skill. + +Read the shape here as discipline that we assert. Do not read any of it as an +effect that we observed, and do not let a later summary say that this skill +works. diff --git a/source/craft/compressed-deliberation.md b/source/craft/compressed-deliberation.md index 38cf21c..326a2f3 100644 --- a/source/craft/compressed-deliberation.md +++ b/source/craft/compressed-deliberation.md @@ -9,20 +9,59 @@ below is what makes the dating visible. - Evidence cutoff: 2026-07-27. - Rights holder of the sources below: Anthropic PBC. - Transformation: the problem statement is a digest of published vendor - documentation, in our own words. The prescriptions are ours and cite nothing. + documentation, in our own words, and it lives in this record. The + prescriptions are ours, they live in the skill, and they cite nothing. - Reproduction check: **not performed.** No sentence from any source below is - carried into the skill — the seven `G` rows are one-line paraphrases of - behaviour the documentation reports, written here from scratch — so nothing - in this skill depends on a reproduction right. The governing terms of those - pages have not been read, and no claim is made about what they permit. Anyone - adding a quotation must read them first and replace this line with the check. - Recorded 2026-07-27. The grounding matrix gained a `Source text` column on - 2026-08-06 under ADR-0020, and every row reads `unquoted`. This unperformed - check is the reason. + carried into the skill or into this record — the seven statements under "What + the vendor documents" are one-line paraphrases of behaviour the documentation + reports, written here from scratch — so nothing in this skill depends on a + reproduction right. The governing terms of those pages have not been read, and + no claim is made about what they permit. Anyone adding a quotation must read + them first and replace this line with the check. Recorded 2026-07-27. The + grounding matrix gained a `Source text` column on 2026-08-06 under ADR-0020. + It carries no `G` row now, so no cell there quotes anything, and this + unperformed check is why the first new one would read `unquoted`. + +## What the vendor documents + +These seven statements are what `A1`, `A3` and `A5` below report about the +target build. They describe the model. They prescribe nothing, and no rule in +the skill inherits their authority. + +- Claude Opus 5 writes longer visible responses than earlier Opus models. + `A1`, Response length and verbosity. +- It writes longer reports, summaries, and files to disk. `A1`, Written + deliverable length. +- It announces what it is about to do more often during agentic work. `A1`, + User-facing progress updates. +- It verifies its own work without being asked. `A1`, Task scope and + over-verification. +- It narrates its own corrections more often. `A1`, Self-correction. +- It can widen a task past what the reader asked for. `A1`, Task scope and + over-verification, and `A5` pages 86 to 87. +- A lower effort setting does not reliably shorten the visible answer. `A1`, + Response length and verbosity, and `A3`. + +The skill carried these seven as `G` rows until the 2026-08-16 editorial pass, +which found them to be half of one section and none of the guidance. They moved +here whole, so the provenance survives the move. The skill keeps the sentence a +reader acts on, and its grounding matrix now carries no `G` row and names no +source version. Restoring a citation to the skill takes a `G` row, a source +version and the reproduction check above, in one pass. + +One claim did not come here whole. The skill used to say that files written to +disk are the case the vendor documents most clearly, and it attributed that to +this record. This record never said it, and nobody here has re-read `A1` to +settle it, so it stands here unverified and it is out of the skill. The claim +is also a judgment about coverage across pages rather than a rule, so no `G` +row could carry it: the `Source rule` cell would have to name an identifier +that says no such thing. What this repository does know is that it has tested +the file case and the agentic case least, and the skill now says that in its +own words. ## What the sources are, and what each one licenses -A `G` row in the grounding matrix may trace only to the first group. A vendor +A `G` row in a grounding matrix may trace only to the first group. A vendor saying that its model is verbose licenses a **description**, never a rule. ### Vendor documentation of the defaults diff --git a/source/craft/de-slop.md b/source/craft/de-slop.md index fa9d3ac..fe23fc8 100644 --- a/source/craft/de-slop.md +++ b/source/craft/de-slop.md @@ -98,13 +98,23 @@ recipe was followed. The departures are what keep a named failure in view, and ## Why no word list ships with this skill A word that one setting overuses is the most countable part of the defect this -skill treats, and ADR-0021 keeps it out of every skill directory. The signature -layer lives in `bench/score.mjs` instead, as a `signatures` metric that starts -empty. - -A word becomes a rule in a skill only after it clears a promoted study under -the measurement design. Until then a scorer counts it and the product asserts -nothing about it. +skill treats, and that is what makes a list tempting to ship. ADR-0021 keeps it +out of every skill directory. The signature layer lives in `bench/score.mjs` +instead, as a `signatures` metric that starts empty. + +Two arguments carry the decision. A list of forbidden words teaches an agent to +swap each one for its nearest neighbour, which leaves the defect and cleans the +surface. A word may also recur in a setting rather than in a model, which is the +live reading here rather than a finding, and on that reading a shipped list +dates faster than the shape the skill states. + +Counting therefore belongs in `bench/score.mjs`, where a scorer counts and the +product asserts nothing. A word becomes a rule in a skill only after it clears +a promoted study under the measurement design. + +The skill states the first argument in one clause and leaves the rest here. The +2026-08-16 editorial pass moved it, because the argument is about what this +repository ships rather than about how a reader writes. ## Where a measurement would live diff --git a/source/craft/proportionate-execution.md b/source/craft/proportionate-execution.md index cec9e37..f8a5ca0 100644 --- a/source/craft/proportionate-execution.md +++ b/source/craft/proportionate-execution.md @@ -47,10 +47,10 @@ land near one in substance. None is copied, and none claims authority from one. A rule that reads like vendor guidance is still an `E` row, because the vendor wrote guidance for its own product and not a standard for anybody to cite. -`compressed-deliberation` cites the same two pages, and its own record at -`source/craft/compressed-deliberation.md` carries a wider register including the -system card and the counterevidence. Read that record with this one. Neither skill supersedes the -other. +The record for `compressed-deliberation` cites the same two pages, and that +record at `source/craft/compressed-deliberation.md` carries a wider register +including the system card and the counterevidence. Read that record with this +one. Neither skill supersedes the other. ## What this record does not carry