@hassangameryt/freeflow 0.8.0 → 0.9.0

package/skills/write-docs/scripts/plan-lint.mjs

@@ -0,0 +1,123 @@
+#!/usr/bin/env node
+// Checks that every step of a Plan names the check that will show it worked (references/spec-and-plan-shapes.md).
+// A step is a "### Slice" or "### Step" block, or an item of the list under "Ordered actions or slices", "Steps" or
+// another heading that holds the Plan's ordered work. A block names its check in a filled field such as "Tests or
+// observations:", "Check:" or "Verification claim...:"; a list item names it inline ("check: ...", "verify ...",
+// "test ...", "observe ..."). It checks presence only; whether the check could catch a wrong result is the author's
+// and the reviewer's judgment.
+// Usage: plan-lint.mjs <plan.md>... [--json]
+// Exit code 0 when every step names a check, 1 when a step does not or no step was found, 2 on a usage error.
+import { existsSync, readFileSync } from "node:fs";
+import { resolve } from "node:path";
+import { parseArgs } from "node:util";
+
+const USAGE = "Usage: plan-lint.mjs <plan.md>... [--json]";
+let parsed;
+try {
+  parsed = parseArgs({ allowPositionals: true, options: { json: { type: "boolean", default: false } } });
+} catch (error) {
+  console.error(`${error.message}\n${USAGE}`);
+  process.exit(2);
+}
+const { values, positionals } = parsed;
+if (!positionals.length) {
+  console.error(USAGE);
+  process.exit(2);
+}
+
+const STEP_HEADING = /^#{2,4}\s+(Slice|Step|Phase step|Task)\b/i;
+const LIST_HEADING = /^(#{2,4}\s+)?(Ordered actions( or slices)?|Steps|Ordered steps|Actions|Slices|Work)\s*:?\s*$/i;
+const CHECK_FIELD =
+  /^\s*(?:[-*]\s+)?(?:\*\*)?(?:[A-Za-z]+ and )?(Tests?( or observations)?|Checks?|Verification( claims?)?( and required boundar(y|ies))?|Verify|Observations?|Acceptance|Evidence|Final tests or observations)(?:\*\*)?\s*:(?:\*\*)?\s*(.*)$/i;
+const INLINE_CHECK =
+  /\b(check(ed|s)?|verif(y|ied|ies)|tests?|observe[ds]?|assert|confirm(ed|s)?|run `|passes|fails?)\b/i;
+const EMPTY = /^\s*(|-|—|none|n\/a|tbd|todo|\[.*\])\s*$/i;
+
+function lint(file) {
+  const lines = readFileSync(file, "utf8")
+    .replace(/```[\s\S]*?```/g, (block) => block.replace(/[^\n]/g, ""))
+    .split("\n");
+  const steps = [];
+  for (let index = 0; index < lines.length; index += 1) {
+    const line = lines[index];
+    if (STEP_HEADING.test(line)) {
+      const level = /^(#+)/.exec(line)[1].length;
+      let end = index + 1;
+      while (end < lines.length && !(/^#{1,4}\s/.test(lines[end]) && /^(#+)/.exec(lines[end])[1].length <= level))
+        end += 1;
+      const body = lines.slice(index + 1, end);
+      // A field is filled on its own line, or by the list under it ("**Checks:**" then bullets).
+      const fields = body
+        .map((each, at) => {
+          const field = CHECK_FIELD.exec(each);
+          if (!field || !EMPTY.test(field.at(-1))) return field;
+          for (
+            let next = at + 1;
+            next < body.length && !/^\s*(\*\*)?[A-Z][^:]{0,60}:(\*\*)?\s*/.test(body[next]);
+            next += 1
+          ) {
+            const item = /^\s*(?:[-*]|\d+[.)])\s+(\S.*)$/.exec(body[next]);
+            if (item) return [...field.slice(0, -1), item[1]];
+          }
+          return field;
+        })
+        .filter(Boolean);
+      const filled = fields.find((field) => !EMPTY.test(field.at(-1)));
+      steps.push({
+        line: index + 1,
+        step: line.replace(/^#+\s+/, "").trim(),
+        check: filled ? `${filled[1]}: ${filled.at(-1)}`.trim() : null,
+        problem: filled ? null : fields.length ? "its check field is empty" : "names no check",
+      });
+      continue;
+    }
+    if (LIST_HEADING.test(line.trim())) {
+      for (let at = index + 1; at < lines.length && !/^#{1,4}\s/.test(lines[at]); at += 1) {
+        const item = /^(\s*)(?:\d+[.)]|[-*])\s+(.*)$/.exec(lines[at]);
+        if (!item || item[1].length > 1) {
+          if (lines[at].trim() && !item && /^[A-Z][^:]{0,40}:\s*/.test(lines[at].trim())) break;
+          continue;
+        }
+        // A step's check may sit on its own nested lines.
+        const nested = [];
+        for (let next = at + 1; next < lines.length && /^\s{2,}\S/.test(lines[next]); next += 1)
+          nested.push(lines[next]);
+        const text = [item[2], ...nested].join(" ");
+        steps.push({
+          line: at + 1,
+          step: item[2].trim().slice(0, 120),
+          check: INLINE_CHECK.test(text) ? text.match(INLINE_CHECK)[0] : null,
+          problem: INLINE_CHECK.test(text) ? null : "names no check",
+        });
+      }
+    }
+  }
+  return { path: file, steps, problems: steps.filter((step) => step.problem) };
+}
+
+const results = [];
+const inputErrors = [];
+for (const input of positionals) {
+  const file = resolve(input);
+  if (!existsSync(file)) inputErrors.push(`no such file: ${input}`);
+  else results.push(lint(file));
+}
+const failed = inputErrors.length > 0 || results.some((result) => !result.steps.length || result.problems.length);
+if (values.json)
+  console.log(
+    JSON.stringify({ command: "plan-lint", status: failed ? "fail" : "pass", inputErrors, results }, null, 2),
+  );
+else {
+  for (const error of inputErrors) console.log(`error: ${error}`);
+  for (const result of results) {
+    if (!result.steps.length) {
+      console.log(
+        `${result.path}: no steps found (### Slice or ### Step blocks, or a list under "Ordered actions or slices" or "Steps")`,
+      );
+      continue;
+    }
+    console.log(`${result.path}: ${result.steps.length} step(s), ${result.problems.length} without a check`);
+    for (const step of result.problems) console.log(`  line ${step.line}: ${step.step} — ${step.problem}`);
+  }
+}
+process.exit(failed ? 1 : 0);

package/skills/write-skill/SKILL.md

@@ -1,11 +1,13 @@
 ---
 name: write-skill
-description: Use when creating or revising an agent skill.
+description: "Use when creating, revising, merging or splitting a skill or any other text an agent acts on: a SKILL.md and its description, the references and scripts in a skill, AGENTS.md, CLAUDE.md, system or runtime prompts, tool descriptions, subagent briefs or judge prompts. Also use when deciding where an instruction should live, and when a skill is read at the wrong time, not read at all, or read but not followed."
 ---
 
 # Write Skill
 
-Write the smallest agent-first skill package that changes one coherent behavioral policy.
+Write the smallest agent-first skill package that changes one coherent behavioral policy, or the smallest change to other instruction text an agent acts on.
+
+**Other instruction text** is anything an agent reads and acts on that is not a skill package: an instruction file such as AGENTS.md or CLAUDE.md, a system or runtime prompt, a tool description, a subagent brief, or a judge prompt. For it, everything below applies except what concerns the package itself: the skill's frontmatter description, resource files and their read conditions, the dependency graph, and `validate`. A tool description is a routing interface too, so Write The Description For Recall applies to it. Read "the skill" as "the text", and its first read as the point the agent receives it: every request for an always-loaded file or prompt, the tool call for a tool description, dispatch for a brief or a judge prompt.
 
 A skill is a prompt intervention, not merely a document. A host may read its body once and retain it across later messages and tool interactions until context projection or loss. Design the first useful read, retained influence, dependencies, and recovery paths; do not assume the body will be reread before every use.
 
@@ -13,16 +15,61 @@ Smallest means behaviorally minimal, not shortest. A detailed skill is justified
 
 The **Rules** below are binding; if one cannot be met, stop and say why. Everything else is **judgment**: follow its stated goal, and adapt when a different choice serves that goal better.
 
+## Routes
+
+This body covers the common case: writing or revising one skill or one piece of instruction text. Open a reference only when its condition holds.
+
+| Situation | Go to |
+| --- | --- |
+| Deciding where an instruction lives (always-loaded text, a description, a body, a reference, a script, a tool's own text) or moving it between them; adding, merging, splitting or removing a reference, script or skill; changing links, shared terms, entry paths, capability cues or context-loss recovery cues between skills | [Placement And Dependencies](references/placement-and-dependencies.md) |
+| Writing a description whose missed read could cause an irreversible or costly action, or whose neighbour matches similar requests; evidence that a skill was read late, early, not at all, or on a neighbour's request, or that two descriptions match the same request; or evidence that a read skill was misread, ignored, or kept steering later work | [Activation And Wording](references/activation-and-wording.md) |
+| Measuring what a revision changes in behavior | Evaluate In One Loop, below |
+| Checking structure, copied passages, or what a revision lost | Check The Package, below |
+
 ## Rules
 
-- Before substantially revising an existing skill, preserve the complete current package (`SKILL.md`, references, scripts, assets) in a recoverable Git ref or in task evidence outside the package. A small correction needs no new snapshot when the exact previous state is already recoverable and behavior, activation, dependencies, and resources are unchanged. A baseline supports recovery and comparison; it does not show that either version behaves correctly.
-- If accepted behavior, public dependencies, safety, compatibility, or another user-owned boundary is unsettled, stop and resolve it before encoding it.
+- The baseline of a revision is the committed state before it. Before committing a substantial revision, such as a merge, a split, a move between files or a rewrite, run `skill-author compare --all-units` and settle every unit it reports unmatched: restore it, name the file and line that now owns it, or record why it is removed, in a verdicts file kept with the evidence outside the package. Then someone, you or a reviewer, reads every reworded and unmatched unit for meaning, since a rewording can keep the words and change what they ask. Outside Git only, copy the package to the task's scratch folder before editing and pass `--base <folder>`. A clean comparison shows that text survived, not that behavior did.
+- If accepted behavior, public dependencies, or another user-owned choice is unsettled, stop and resolve it before encoding it.
 - Never rely on an unread skill or reference to cause its own read. The first read must come from context the agent already has: a description, guaranteed base guidance, an already-loaded skill, a capability cue, or another declared route.
 - Rely only on context guaranteed on every valid activation path, never on an optional sibling, a common but unguaranteed route, or an accidental earlier read.
-- Link every agent-facing reference, asset, script, and package dependency from the body, with the condition that requires it. The body and the resource state the same read condition.
-- Keep authoring status, evaluation summaries, evidence gaps, changelog notes, readiness claims, and author debate out of the executing package; they belong in the evaluation artifact, task record, or delivery report.
+- In a skill package, link every agent-facing reference, asset, script, and package dependency from the body, with the condition that requires it. State the condition once, at the link; the resource opens with its content, not a restatement of its read condition.
+- When a skill is removed or merged into another, add no alias, shim or redirect under the old name: the description of the skill that now owns the job routes the work, every link and command naming the old skill is updated, and the removed name is reported as a breaking change.
+- Keep authoring status, evaluation summaries, evidence gaps, changelog notes, readiness claims, and author debate out of the executing package or text; they belong in the evaluation artifact, task record, or delivery report.
 - Claim only what the evidence observed. Structural checks prove structure. Explicit body delivery does not prove activation, one turn does not prove retained use, ambient context does not prove dependency behavior, prose review does not prove unchanged behavior, and one favorable example does not prove readiness.
-- Run `validate` after any structural or dependency change and before claiming the package is valid. Never use deprecated authoring tooling.
+- For a skill package, run `validate` after any structural or dependency change and before claiming the package is valid. Never use deprecated authoring tooling.
+
+## Choose What To Change
+
+Goal: change a skill only for behavior that will recur, and only where prose is the right fix.
+
+When revising from evidence, such as observed failures, transcripts, review findings or eval results, list each candidate change with the evidence behind it. Drop a candidate that is:
+
+- **not durable:** a one-off slip that the situation, not the guidance, explains;
+- **too specific:** it fits one task or fixture, not the class of situations the skill serves;
+- **a duplicate:** another candidate, or another skill, already owns it;
+- **already covered:** the skill says it, and the failure lies elsewhere, such as activation or placement; fix that boundary instead;
+- **structural:** a tool, script, check or runtime change prevents it more reliably than an instruction.
+
+Keep dropped candidates and their reasons with the evidence, outside the package.
+
+## Prefer A Capability To Prose
+
+Goal: the fact or check the agent needs is in front of it at the decision.
+
+Prose changes behavior only when it is in context at the decision, and a skill body often is not: in 30 of Freeflow's own evaluation runs, 3 read any skill file. A script, a tool's output or a structured artifact reaches the decision whether or not a body was read. Before writing an instruction, ask whether a capability can do its job:
+
+- **Replace mechanical prose:** a format, command list or check followed by hand becomes a script that does or checks it, such as `skill-author validate` for frontmatter and links.
+- **Enable the main work:** a script puts the decision's facts in front of the agent, such as Verify Work's `test-changes.mjs`, which lists every changed existing test so the agent and its reviewer judge each one.
+
+The prose then names the capability where it is used and keeps only the judgment a script cannot make. Add a script only for work the skill owns, never to make a package look complete.
+
+## Choose Skill Boundaries
+
+Goal: few skills, each owning one job in depth.
+
+Every skill's description is loaded on every request, and neighbours whose descriptions overlap confuse the choice between them. Prefer depth to width: when several skills do one job split by type, such as commit, push and release, merge them into one skill whose body routes to a reference per type. Keep a skill separate when its reader, trigger or job differs: another person or agent reads it, another request or situation calls for it, or it produces a different result.
+
+When the same workflow is rebuilt by hand across several tasks and no skill owns it, first check whether it is a branch of an existing skill; otherwise propose a new skill to the user, naming the tasks it recurred in.
 
 ## Establish The Behavioral Claim
 
@@ -41,13 +88,24 @@ Before writing, identify:
 
 Begin with the behavioral intervention, not with sections, wording, examples, or file layout.
 
+## Shape The Package
+
+Goal: the common case runs from `SKILL.md` alone, and the depth one situation needs sits in one place.
+
+- `SKILL.md` is enough to act correctly in the common case without opening a reference.
+- When the skill has references, a routing table comes first in the body: each row names a concrete, observable condition and the reference or section that serves it. "When unclear" is not a condition.
+- A reference holds one situation's method. Add one only when its condition is narrow, so that many activations never reach it, and its content is large, so that inlining it would bury the common route; otherwise keep it inline.
+- Merge references that are read together.
+- Give each rule one owner; everywhere else, point to it rather than restate it (Repeat Instructions Deliberately).
+- Declare only the dependencies the skill's normal and conditional routes need.
+
+Start with one `SKILL.md` and remove accidental complexity before splitting files. There is no universal body template, line limit, or reference count.
+
 ## Design The First Read
 
 Goal: the skill is read at the earliest decision it can improve, and reading it on a nearby case does no harm.
 
-Missing that decision is usually far more harmful than an unnecessary early read. A nearby case can pass by not loading the skill, or by loading it, recognizing that its method does not apply, and exiting without distorting later work. An early read is cheap only when the retained body stays safe; a skill that exits on the first turn but keeps distorting later work is not a safe read.
-
-Read [Activation Boundaries](references/activation-boundaries.md) when activation is unclear or evidence shows missed, late, early, task-hijacking, or retained-interference behavior.
+Missing that decision is usually far more harmful than an unnecessary early read. A nearby case can pass by not loading the skill, or by loading it, recognizing that its method does not apply, and exiting without distorting later work. An early read is cheap only when the retained body stays safe; a skill that exits on the first turn but keeps distorting later work is not a safe read. When a missed read is costly, give the skill an upstream activation or recovery edge, as Placement And Dependencies describes.
 
 ## Write The Description For Recall
 
@@ -55,10 +113,12 @@ The description is a routing interface, not a summary of the body.
 
 - State the broad core job in plain language, with natural user verbs or task conditions visible before activation.
 - Avoid terminology or classifications introduced only inside the skill or an optional dependency.
-- Avoid future-duration guesses, named internal phases, exhaustive trigger catalogs, marketing claims, and compressed copies of the body.
+- Name the situations each reference and script serves, in the words a request would use. A reference is reached only through its skill, so when skills merge, the owner's description keeps the triggers of every job it absorbed.
+- Make it as long as recall needs, up to the host's limit (1,024 characters in the Agent Skills format). Every description is loaded on every request, but a missed read costs the decision the skill exists to change.
+- Avoid future-duration guesses, named internal phases, marketing claims, and compressed copies of the body.
 - "Use when…" is useful wording, not mandatory grammar.
 
-Favor recall when nearby reads are safe and a missed read is materially worse. Narrow the description when retained guidance routinely overrides another owner or method, not merely to prevent harmless reads.
+Favor recall: an unneeded read is cheap when the body exits safely, and a missed read is not. Add a "not for" exclusion only when reading the skill on that request would do harm, such as retained guidance that routinely overrides another owner or method, never merely to prevent harmless reads.
 
 ## Write For The Executing Agent
 
@@ -76,8 +136,6 @@ Write for an agent seeing the skill for the first time with only guaranteed cont
 
 Use discussion, research, and author reasoning to understand the behavior, then compile that understanding into runtime instructions. The conversation, rejected drafts, fixture narration, and any explanation whose only reader is the author stay out.
 
-Read [Agent-First Instructions](references/agent-first-instructions.md) when a concrete first-read or retained-use failure involves wording, placement, undeclared context, author explanation, or competing instructions.
-
 ## Separate Rules From Judgment
 
 A skill gives two kinds of instruction, and agents need to tell them apart.
@@ -113,7 +171,7 @@ goal
 
 Examples inside judgment are illustrations, not triggers:
 
-> Delegate a unit when a cheaper profile can do it without significant quality loss and the handoff costs less than it saves. A separate record update across several entries often qualifies; a one-line update in the middle of your own work does not.
+> Delegate work when a cheaper profile can do it without significant quality loss and the handoff costs less than it saves. A separate record update across several entries often qualifies; a one-line update in the middle of your own work does not.
 
 When a skill contains both kinds, say so once near the top: which instructions are binding, and that everything else is judgment the agent may adapt to its goal. A skill may be mostly Rules, like a release protocol, or mostly judgment, like work placement; the label keeps each one from being read as the other.
 
@@ -125,19 +183,6 @@ Add a compact example when prose alone leaves a likely misclassification. Show t
 
 Keep examples transferable; the body is not a transcript, fixture library, or scenario catalog. Inline examples teach the subject and are part of the candidate intervention, not independent evidence that the skill works.
 
-## Plan Terminology And Dependencies
-
-Goal: every term and dependency is available on every path that reaches the skill.
-
-For one self-contained skill, declare only the dependencies its normal and conditional routes need. Before creating or revising several related skills, shared terminology, cross-skill links, capability activation, direct or routed entry paths, or context-loss recovery, read [Skill Dependency Graphs](references/skill-dependency-graphs.md).
-
-- Identify every valid direct and routed activation path.
-- Place shared vocabulary in the nearest upstream context guaranteed before every consumer; keep leaf-specific terms local.
-- Distinguish owner routes, method composition, environment interactions, references, capability cues, and retained context.
-- Give critical skills an upstream activation or recovery edge.
-
-A Markdown link declares a package dependency. It does not select the target owner or execute the linked skill.
-
 ## Compact Without Erasing Behavior
 
 Goal: remove what does not change behavior, and nothing that does.
@@ -150,6 +195,8 @@ If yes, remove or merge it.
 
 Keep an instruction because it protects measured behavior or prevents a specific failure you can name, not because it might help. An instruction with neither is a candidate for removal: remove it, or test whether it changes behavior, rather than keeping it by default. Skills otherwise only grow, as each incident adds instructions and none is ever shown safe to remove.
 
+Remove an instruction once the mistake it guards against can no longer happen, for example because a tool now refuses the action or the capability it guarded is gone. Keep a local guard only for a path that still bypasses the enforcement.
+
 Compaction goes wrong when it:
 
 - deletes the positive route while keeping only prohibitions;
@@ -166,37 +213,28 @@ Length is a diagnostic fact, not a quality verdict. A long skill may be minimal;
 
 Goal: each behavior has one owner, with a local guard wherever an agent could otherwise act without it.
 
-Repetition earns its place when local presence protects an independently activated skill, a distinct decision point, a context-loss or host path, a dangerous literal reading, a previously observed failure, or a required read condition. Keep the full definition or method with one canonical owner and repeat only the compact guard at the affected choice.
+Repetition earns its place when local presence protects an independently activated skill, a distinct decision point, a context-loss or host path, a dangerous literal reading, retained use after the method ends, a previously observed failure, or a required read condition. Keep the full definition or method with one canonical owner and repeat only the compact guard at the affected choice.
 
 > Repeat the behavioral invariant, not the owning skill's entire explanation.
 
 Repetition for emphasis, author reassurance, or stylistic consistency does not earn its place. Check that repeated wording keeps the same meaning and route. Removing, merging, moving, or rewording an instruction that protects measured behavior is a behavioral change; evaluate the affected boundary.
 
-## Add Resources Only When They Change Loading Or Ownership
-
-Before adding or reorganizing references, assets, or scripts, read [Progressive Disclosure](references/progressive-disclosure.md).
+## Evaluate In One Loop
 
-Start with one `SKILL.md` and remove accidental complexity before splitting files.
+Goal: every revision answers observed evidence and is checked against it.
 
-- Keep normal first-read behavior inline.
-- Add an activity-required reference when one named normal activity needs separately owned depth but the skill may exit before that activity.
-- Add a conditional reference when an observable branch needs depth that would obscure the normal route.
-- Use an entry-required reference only when separation has independent value, such as canonical reuse, separate ownership or lifecycle, generated contracts, compatibility, or host constraints; otherwise inline it.
-- Add a script when deterministic repeated work owned by the skill is safer, clearer, or materially less wasteful than model-generated operations.
-- Keep owned resources inside the package unless canonical cross-package ownership is required.
+1. Start from the source: the evidence a change answers (Choose What To Change), or for a new skill, its behavioral claim.
+2. Name the exact claim: description activation, first-read body behavior, nearby behavior, dependency composition, retained use, artifact outcome, or cross-host behavior.
+3. Fix the cases before the draft is final: one that should use the guidance, a near miss that should not, and for judgment, one where following the default literally is the worse choice.
+4. Run the arms with [Evaluate Behavior](../evaluate-behavior/SKILL.md) under fixed declared inputs; when the claim is that the revision is better, declare the decision rule before the run.
+5. Read the transcripts and results, with the user when they review, before trusting a count.
+6. Revise one measured pressure point.
+7. Rerun the complete fixed group with every arm.
 
-There is no universal body template, line limit, reference count, or requirement that a script follow an earlier failure.
-
-## Evaluate The Behavioral Claim
-
-Goal: evidence that the skill changes behavior as intended, at the boundary being claimed.
-
-Name the exact claim: description activation, first-read body behavior, nearby behavior, dependency composition, retained use, artifact outcome, or cross-host behavior. Use [Evaluate Skill](../evaluate-skill/SKILL.md) to compare exact baseline and candidate environments under fixed declared inputs.
+Stop when the evidence supports the claim, or when the next revision would only tune against the same cases; judge it on held-out cases instead. Preserve adequate existing evidence rather than rerunning it.
 
 Evaluate the two kinds differently. A Rule passes when it is never broken, including under pressure to break it. Judgment passes on outcome, and its evidence should include a case where following the default literally is the worse choice, so a candidate that only obeys cannot pass.
 
-Preserve adequate existing evidence. Revise one measured pressure point at a time, and after a behavior-sensitive change rerun the complete fixed group with both variants.
-
 Small wording, ordering, vocabulary, repetition, and structural changes can alter model behavior. Human review establishes intended meaning; behavioral evidence establishes observed effect.
 
 ## Check The Package
@@ -205,15 +243,17 @@ Use the bundled [skill author](scripts/skill-author.mjs) for deterministic packa
 
 ```text
 node <write-skill-directory>/scripts/skill-author.mjs init <directory> --name <name> --description <text>
-node <write-skill-directory>/scripts/skill-author.mjs validate <directory> [--package-root <directory>]
+node <write-skill-directory>/scripts/skill-author.mjs validate <directory> [--package-root <directory>] [--routing-map <file>]
 node <write-skill-directory>/scripts/skill-author.mjs inspect <directory> [--package-root <directory>]
 node <write-skill-directory>/scripts/skill-author.mjs similarity <directory> [--threshold <0-1>]
+node <write-skill-directory>/scripts/skill-author.mjs compare [<base-ref> [<new-ref>] | --base <folder> [<folder>]] [--all-units] [--verdicts <file>]
 ```
 
 - `init`: create a new minimal package; it never overwrites an existing one.
-- `validate`: check frontmatter, heading shape, recursive links, declared package dependencies, and containment.
+- `validate`: check frontmatter, heading shape, recursive links, declared package dependencies, and containment. It also lists link cycles among the Markdown files the skill reaches, leaving out links to any `SKILL.md`; Placement And Dependencies says what to do with one. With `--routing-map`, it checks that the map's row for the skill names the same sibling skills and resources that `SKILL.md` links.
 - `inspect`: report body sizes, resource classes, scripts, unlinked files, and validation findings, when factual inventory matters.
 - `similarity`: report Markdown files under a directory that share copied or lightly edited passages. Run it across related skills before keeping repetition; it cannot detect paraphrase, so a clean report does not prove an instruction is stated once.
+- `compare`: with no arguments, compare the model-facing Markdown changed since the branch left the default branch with the working tree, uncommitted edits included, and list each rule-like unit (under a Rules heading, a numbered step, or containing never, always, must, only, do not or stop, or a reason: because, so that, otherwise, why) that no longer appears; `--all-units`, for a substantial revision, lists every unmatched unit. It exits nonzero while one has no verdict in `--verdicts <file>`, a JSON object of `"file:line": "verdict"` that carries across runs. Read every reworded rule it lists as well: a rewording can keep the words and change the meaning. `--help` gives the other options.
 
 Commands emit JSON. Invalid structure still emits a report and exits nonzero; command errors emit structured error JSON. Package root defaults to the nearest `package.json` ancestor, then the skill directory's parent; use `--package-root` for a candidate in an overlay or another package context.
 
@@ -231,6 +271,6 @@ Stop when the requested package:
 - keeps its normal route executable on first read;
 - declares dependencies and resource conditions honestly;
 - contains no author-facing residue;
-- passes required structural checks.
+- passes required structural checks, and after a substantial revision has a verdict for every rule-like unit it no longer states.
 
-Stop and ask when accepted behavior, activation, dependencies, safety, compatibility, or another user-owned boundary cannot be settled from evidence. Do not add sections, references, examples, scripts, or procedure merely to make the package look complete.
+Stop and ask when accepted behavior, activation, dependencies, or another user-owned choice cannot be settled from evidence. Do not add sections, references, examples, scripts, or procedure merely to make the package look complete.

package/skills/write-skill/references/activation-and-wording.md

@@ -0,0 +1,143 @@
+# Activation And Wording
+
+This reference is judgment: adapt it to the skill's goal. Its binding statements restate [Write Skill](../SKILL.md) Rules.
+
+Two failures look alike from outside: the skill is not read at the decision it should change, or it is read and its text does not steer the agent as written. Find which one the evidence shows before rewriting anything; a better body cannot fix a missed read, and a sharper description cannot fix a misleading body.
+
+## Activation
+
+A description should make the skill available at the earliest useful decision without requiring language introduced by the skill itself.
+
+### Diagnose The First Read
+
+Beyond the behavioral claim in Write Skill, identify:
+
+- the earliest natural prompt or task state where the skill can change behavior;
+- the consequence of missing or delaying the read;
+- the consequence if a nearby prompt also loads the skill;
+- how retained guidance should behave after the immediate method ends.
+
+### Compare Activation Costs
+
+False-positive and false-negative activation have different costs.
+
+A false-positive read may be acceptable when:
+
+- loading cost is proportionate;
+- the body identifies that its method does not apply;
+- it exits without changing ownership or route;
+- retained instructions do not hijack later work.
+
+A false negative is material when missing the first useful read causes:
+
+- an irreversible or expensive action;
+- a wrong owner or method;
+- repeated low-value environment interactions;
+- missing safety or evidence boundaries;
+- later work that the skill cannot reliably undo.
+
+Prefer recall when nearby reads are demonstrably safe and missed activation is materially worse. Tighten activation when early or retained guidance can distort ordinary work.
+
+### Check The Cues
+
+Beyond Write The Description For Recall, a cue names an observable distinction from neighboring jobs, and never rests on generic "help with" or "improve" language. Guaranteed base context may define a term; otherwise use plain observable language.
+
+### Test Nearby And Retained Behavior
+
+A nearby case can pass through:
+
+- **Non-trigger:** the skill is not read.
+- **Safe read:** the body loads, exits or defers correctly, and does not interfere later.
+
+Test both the immediate response and later declared turns when retained influence matters.
+
+Broad useful trigger:
+
+> Use when committing work.
+
+Over-specified trigger that hides the job:
+
+> Use after final verification when a planned checkpoint requires a locally reviewable commit.
+
+Broad safe-read trigger:
+
+> Use when working with software design—especially boundaries, interfaces, ownership, state, and failure behavior.
+
+The body must exit when the local interface remains sound and avoid turning later local choices into architecture work.
+
+### Preserve The Evidence Boundary
+
+Natural activation proves the host read the exact skill snapshot from a natural prompt. Explicit body delivery tests first-read instructions, not activation. Dependency composition tests exact declared skill and context order. Multi-turn evidence tests retained use. Artifact and transcript evidence show whether activation helped, exited safely, or hijacked behavior.
+
+Preserve:
+
+- the earliest useful positive prompt;
+- at least one close nearby case;
+- later-turn evidence when retained influence is part of the claim.
+
+Do not infer one evidence class from another.
+
+## Wording
+
+### Translate Author Context
+
+Keep a sentence only when it:
+
+- changes an action or judgment;
+- defines required language;
+- protects a likely or observed failure boundary;
+- states an evidence, return, or stop condition;
+- makes a dependency executable;
+- provides a compact example resolving likely misclassification.
+
+Translate rationale into a direct instruction plus only the why needed to generalize. Translate debate into the selected boundary and stop condition. Translate a measured failure into the smallest transferable guard that prevents it. Omit rejected drafts, conversation history, fixture narration, evaluation status, and explanation whose only audience is the author.
+
+### Make Instructions Concrete
+
+A Rule gets concrete by naming its trigger, its action, and its stop. Prefer:
+
+> If saved evidence already answers the fixed question, inspect it before proposing a rerun.
+
+Over:
+
+> Keep evaluation efficient and aligned with best practices.
+
+Concrete Rules reduce the judgment the agent must reconstruct.
+
+Judgment gets concrete differently: a sharper goal, clearer signals for deviating, and a realistic example on each side of the line. Turning judgment into an absolute makes it easier to follow and worse to follow. Prefer:
+
+> Start a Slice for work that would leave partial state to reconcile if interrupted. An investigation across many sources can qualify; a discussion whose results land directly in the record usually does not.
+
+Over:
+
+> Start a Slice for every activity.
+
+### Pair Negative Instructions With The Route
+
+Bad:
+
+> Do not recreate the old worker/verifier architecture.
+
+Better:
+
+> The producing activity handles implementation and verification. The surrounding workflow may select independent review.
+
+Bad:
+
+> Do not implement the recommendation.
+
+Better:
+
+> Return the recommendation to the owning workflow. Implementation requires separate authorization.
+
+A negative instruction may preserve an invariant directly when the positive route is already clear:
+
+> Preserve the accepted scope. Do not split it into MVP or roadmap phases without approval.
+
+### Fix Placement Before Adding Prose
+
+A late caveat rarely defeats an early command. Move or sharpen the controlling sentence before adding another section. When an early instruction and a later exception conflict, restructure the route rather than adding another disclaimer.
+
+### End The Skill's Influence
+
+A skill body may remain in context after its immediate method ends. State stop, return, or exit conditions clearly enough that retained instructions do not continue claiming ownership or broadening later work. A first-turn exit is insufficient when later turns still show method hijacking. After editing, rerun the same first-read and retained-use pressure while keeping unrelated instructions stable.