docs: Dispatch siblings by slash form and state directives positively

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: MinecraftFuns

AGENTS.md

@@ -143,15 +143,22 @@ unchanged in any spec-compliant agent and uploads without hard errors:
   material to sibling files the agent reads on demand.
 * When a step uses a capability another skill in this library provides (PDF
   extraction, commit drafting, prose cleanup, environment provisioning),
-  command that skill by name: "read PDFs with the read-pdf skill", never "a
-  PDF-reading tool" and never "when installed". The library is consumed as a
-  set through the submodule, so siblings are present; write the dispatch as
-  an instruction. Fallbacks exist only for conditions of the environment (an
-  unreachable file, no network), never for a sibling's possible absence.
+  command that skill in its slash form as an unconditional instruction:
+  "read PDFs with `/read-pdf`". The leading slash marks the name as a skill
+  invocation to the executing agent, and the imperative holds because the
+  library is consumed as a set through the submodule, so siblings are
+  present. A fallback chain responds to conditions of the environment (an
+  unreachable file, no network) and names the degraded path to take.
 * Write directives, not commentary about the document. A sentence that
   explains why a convention exists, restates what a table's own headings
   already say, or narrates the file's structure earns nothing and is paid
   for on every load. Keep rationale only where it changes a judgment call.
+* State each directive as the pattern to follow. A prohibition spells out
+  the unwanted pattern in full, which raises that pattern's salience for
+  the model reading it; the positive form spends the same tokens making the
+  right path more likely. Reserve negation for hard boundaries where the
+  banned form must be named to be recognized (secrets, em-dashes, spec
+  violations).
 * Bundled files are addressed by **registered name**, never by path, so a
   path cannot drift and two files cannot claim one name:
   * **One declaration site, declared before use.** Every skill that bundles

author-skill/SKILL.md

@@ -30,7 +30,7 @@ license: MIT
     <rule>Extract only verified tool calls and successful commands from the execution history.</rule>
     <rule>Parameterize all project-specific values (paths, hostnames, IDs) or provide instructions to derive them dynamically.</rule>
     <rule>Express directives conditionally and explicitly (e.g., "If X, execute Y").</rule>
-    <rule>Enforce token-economical language in the generated skill, utilizing sentence fragments and eliminating conversational filler; the caveman skill defines that register in full.</rule>
+    <rule>Enforce token-economical language in the generated skill, utilizing sentence fragments and eliminating conversational filler; `/caveman` defines that register in full.</rule>
     <rule>Never emit em-dash characters (U+2014); use a hyphen, a comma, a colon, or restructure the sentence.</rule>
   </execution_constraints>
 

caveman/references/review.md

@@ -46,5 +46,5 @@ normal paragraph, then resume terse for the rest.
 
 Reviews only: does not write the fix, approve, request changes, or run
 linters. Output comments ready to paste. One-shot: the active intensity
-level is untouched. Complements the ponytail skill's review mode, which
+level is untouched. Complements `/ponytail`'s review mode, which
 hunts over-engineering; this mode compresses any review's phrasing.

fact-check/SKILL.md

@@ -97,8 +97,8 @@ span with the corrected span using the available file-editing tool".
    the replacement introduced (report any such secondary edit as part of the
    correction, not silently).
 6. **Summarize**: claims checked, verdict counts, corrections applied,
-   rejected, abstained; branch and cost. Suggest the user commit via the
-   git-commit skill. Do not auto-invoke any other skill or tool as a
+   rejected, abstained; branch and cost. Suggest the user commit via
+   `/git-commit`. Do not auto-invoke any other skill or tool as a
    follow-up.
 
 ## Per-claim contract

lit-review/SKILL.md

@@ -124,9 +124,9 @@ Before the protocol phase, determine from actually available tools:
   supplies their own corpus (PDFs, BibTeX), skip the search phase, record
   provenance as user-supplied in the log's place, and run the remaining
   phases unchanged.
-- Full-text reading: read PDF full texts with the read-pdf skill. A paper
-  with no reachable PDF falls back to landing-page HTML, then to abstract
-  level, disclosed in the report.
+- Full-text reading: read PDF full texts with `/read-pdf`. A paper with no
+  reachable PDF falls back to landing-page HTML, then to abstract level,
+  disclosed in the report.
 - Sub-agents: optional for parallel extraction only. A worker receives one
   included paper and returns one extraction record; workers never write
   session state. Results must not depend on which branch ran.

lit-review/references/extract.md

@@ -8,10 +8,10 @@ available later without re-reading.
 
 - Start with the anchor: the included paper most cited by the others; it
   fixes the field's vocabulary and baseline claims.
-- Read each paper's `pdf_url` with the read-pdf skill; its page markers
-  become the claim locations in the extraction record. When no PDF is
-  reachable, fall back to the landing page's HTML text, then to the
-  abstract as the floor. After each
+- Read each paper's `pdf_url` with `/read-pdf`; its page markers become
+  the claim locations in the extraction record. When no PDF is reachable,
+  fall back to the landing page's HTML text, then to the abstract as the
+  floor. After each
   paper, set its `read_level` (`abstract` or `full-text`) with `update`.
   Invariant 5 makes this label the ceiling for how its claims appear in the
   report.

lit-review/references/report.md

@@ -80,5 +80,5 @@ reported numbers with units, named papers doing named things.
   deliverable; tables carry structure.
 - Recency words ("recent", "current") always bind to the method section's
   as-of date.
-- Sweep the finished report with the humanize skill's detection index
+- Sweep the finished report with `/humanize` and its detection index
   before delivery.

pl-theorist/SKILL.md

@@ -296,8 +296,8 @@ language's cost model by analogy.
 - A Bloom filter answers "possibly present." Never gate correctness-critical
   logic on a probabilistic membership test alone.
 - A validation step blocked by a missing toolchain is not a dead end:
-  provision the toolchain in userspace with the setup-env skill, or name
-  the unavailable check; never skip it silently.
+  provision the toolchain in userspace with `/setup-env`, or name the
+  unavailable check; never skip it silently.
 
 ## Completion Checks
 

ponytail/SKILL.md

@@ -143,8 +143,8 @@ test, YAGNI applies to tests too.
 
 ## Boundaries
 
-Ponytail governs what you build, not how you talk (pair with the
-caveman skill for terse prose). "stop ponytail" / "normal mode":
+Ponytail governs what you build, not how you talk (pair with
+`/caveman` for terse prose). "stop ponytail" / "normal mode":
 revert. Level persists until changed or session end.
 
 ## Completion Checks