Implement token-saving skill and related evaluation files

.github/workflows/check_pr_release_notes.yml

@@ -21,3 +21,4 @@ jobs:
           github-repository: ${{ github.repository }}
           pr-number: ${{ github.event.number }}
           skip-labels: "no RN"
+          title: "## [Rr]elease [Nn]otes"

README.md

@@ -87,7 +87,7 @@ Before building a new skill, check whether one already exists:
 | [skills.sh](https://skills.sh)                                                       | Open registry — install with `npx skills add <owner/repo>`              |
 | [anthropics/skills](https://github.com/anthropics/skills)                            | Anthropic reference skills including `skill-creator`                    |
 | [absa-group/agent-skills](https://github.com/absa-group/agent-skills)                | Broader ABSA-owned skill collection                                     |
-| [absa-group/cps-agentic-toolkit](https://github.com/absa-group/cps-agentic-toolkit)  | CPS team's skill set built on top of this repo                          |
+| [absa-group/cps-agentic-toolkit](https://github.com/absa-group/cps-agentic-toolkit)  | CPS team's extended skill set (ABSA-internal)                           |
 
 ## Contributing
 

docs/getting-started.md

@@ -132,7 +132,7 @@ Project skills take precedence over global skills when both exist.
 ### Add project-specific skills
 
 For skills that only apply to a specific repository, place them in `.github/skills/` within that repo. These are loaded
-automatically when Copilot CLI is launched from that directory, layered on top of your personal and CPS base skills.
+automatically when Copilot CLI is launched from that directory, layered on top of your personal and shared base skills.
 
 ```
 your-project-repo/

skills/SKILL.md

@@ -0,0 +1,66 @@
+---
+name: token-saving
+description: >
+  Always-active response formatting rules — invoke for every reply without exception: coding
+  questions, code generation, debugging, explanations, comparisons, reviews, diffs, PR updates,
+  recaps, summaries, workflow tasks, non-technical requests, and anything else. Also invoke on
+  explicit brevity signals: be concise, keep it short, save tokens, too verbose, shorter, terse,
+  brief, no fluff, summarise, can you make that shorter. Rules: no filler openers (Certainly!,
+  Great question!, Happy to help!); no closing platitudes (Let me know if you have questions!);
+  concise within line limits; skip restating prior context; prefer tables/bullets over prose;
+  append What changed / Why / How to verify footer only for code-output responses, not Q&A,
+  reviews, or planning. Boundary: when user explicitly requests full detail, deep dive, complete
+  explanation, or says "don't hold back", length rules suspend — respond fully. Another active
+  skill's more specific format requirements take precedence.
+---
+
+# Token-Saving
+
+Always-active base behaviour. Apply to every response without exception unless the user explicitly requests verbosity.
+
+## Always apply — response discipline
+
+- Default to the shortest response that fully answers the question
+- Factual or conceptual answers: aim for ≤ 5 prose lines; one minimal code block is permitted and does not count toward that limit
+- Action lists and next-step recommendations: cap at 4 bullets; no header line before the list
+- Must not repeat context already established in the conversation
+- Must not pad responses with preamble ("Great question!", "Certainly!", "As an AI...")
+- Must not add closing summaries that restate what was just said
+- Stop when the task is complete — must not append "let me know if you need anything else" filler
+- Prefer structured output when it improves clarity: bullets, tables, and short code blocks over dense prose
+- If another active skill or task requires a more specific output format, that format takes precedence
+
+## Format code output responses
+
+End every response where you output code for the user to incorporate — new functions, patches, inline diffs, config snippets, or any code block that represents a change — with exactly this structure (no more, no less):
+
+```
+**What changed:** <one line>
+**Why:** <one line>
+**How to verify:** <command or test instruction>
+```
+
+This footer does NOT apply to pure Q&A, reviews, planning, comparisons, or conceptual explanations — only when you are writing or changing code.
+
+When applying or confirming a bug fix: always show the changed line(s) or a minimal diff, then the footer. A prose description of a code change without showing the code is not sufficient.
+
+- Must not paste full file contents unless the user explicitly asks
+- Show diffs or changed sections only
+- Include enough surrounding context for the change to be unambiguous
+
+## Keep summaries and recaps concise
+
+- Aim for ≤ 10 lines in any recap
+- Prefer linking to files/lines over quoting large blocks
+- Use bullet lists over paragraphs
+- Summarise deltas — what is different — not what already existed
+
+## Update PR bodies by appending only
+
+- Treat the PR description as a changelog — append only, never rewrite
+- Append under `## Update YYYY-MM-DD` with the commit hash — use today's date from your system context (the current date, not a guessed or example date)
+- Must not delete prior update sections
+
+## Respond fully when detail is explicitly requested
+
+If the user explicitly asks for a full explanation, rationale, or deep dive — ALL rules in this skill are suspended for that response. Cover every step, concept, and detail without omitting any part of the topic. Do not apply line limits, bullet caps, or summarisation.

skills/token-saving/evals/evals.json

@@ -0,0 +1,186 @@
+{
+  "skill_name": "token-saving",
+  "evals": [
+    {
+      "id": 1,
+      "category": "happy-path",
+      "prompt": "What is the purpose of a Dockerfile ENTRYPOINT vs CMD?",
+      "files": [],
+      "expected_output": "A concise answer (2\u20135 lines) explaining the difference. No preamble, no closing filler.",
+      "expectations": [
+        "Response does not begin with filler openers: 'Great question!', 'Certainly!', 'Sure!', 'Of course!', 'As an AI...', 'Happy to help!' or equivalent",
+        "Response is \u2264 5 prose lines (a single code block, if present, does not count toward this limit)",
+        "Response does not end with closing filler such as 'Let me know if you need anything else', 'Feel free to ask', or 'Hope that helps!'",
+        "Response correctly explains ENTRYPOINT (fixed executable) vs CMD (default overridable args) without padding",
+        "Response does NOT end with a **What changed** / **Why** / **How to verify** footer — pure Q&A, not a code output response"
+      ]
+    },
+    {
+      "id": 2,
+      "category": "happy-path",
+      "prompt": "Fix the window comparison bug in the rate limiter. The diff is at evals/files/rate-limiter-fix.diff \u2014 apply the fix and confirm.",
+      "files": [
+        "evals/files/rate-limiter-fix.diff"
+      ],
+      "expected_output": "Agent confirms the fix and ends with the exact 3-line footer: **What changed**, **Why**, **How to verify**. Shows only the changed line, not the full file.",
+      "expectations": [
+        "Response ends with exactly the footer: '**What changed:** <one line>', '**Why:** <one line>', '**How to verify:** <command or test instruction>'",
+        "Response does NOT include the full contents of rate_limiter.py",
+        "Response shows only the changed line or a short diff/snippet of the fixed function",
+        "No preamble opener ('Certainly! I'll fix that for you...')",
+        "No closing filler after the footer"
+      ]
+    },
+    {
+      "id": 3,
+      "category": "regression",
+      "prompt": "Context from prior conversation is in evals/files/prior-conversation.txt. Now that we've covered the architecture \u2014 what's the concrete next step I should take?",
+      "files": [
+        "evals/files/prior-conversation.txt"
+      ],
+      "expected_output": "Agent gives a direct next-step recommendation without re-summarising the already-established context (stack, architecture decisions).",
+      "expectations": [
+        "Response does NOT restate the tech stack or architecture decisions already established in prior-conversation.txt",
+        "Response does NOT open with 'As we discussed...' or 'To summarise what we covered...' or 'So you have a FastAPI app with...'",
+        "Response jumps directly to the next-step recommendation",
+        "Response is \u2264 5 lines",
+        "No closing filler line"
+      ]
+    },
+    {
+      "id": 4,
+      "category": "regression",
+      "prompt": "Fix the validate_email bug in evals/files/large-user-service.py \u2014 it currently allows empty strings through. Fix only that function.",
+      "files": [
+        "evals/files/large-user-service.py"
+      ],
+      "expected_output": "Agent shows only the fixed validate_email function or a targeted diff of that function. Does not paste the entire large-user-service.py.",
+      "expectations": [
+        "Response does NOT include the full contents of large-user-service.py",
+        "Response shows only the fixed validate_email function or a targeted diff of that function",
+        "Response ends with **What changed** / **Why** / **How to verify** footer",
+        "No preamble opener",
+        "No closing filler after the footer"
+      ]
+    },
+    {
+      "id": 5,
+      "category": "happy-path",
+      "prompt": "Give me a concise recap of what changed in the user service this sprint. Use evals/files/sprint-changelog.txt as input.",
+      "files": [
+        "evals/files/sprint-changelog.txt"
+      ],
+      "expected_output": "A recap in \u2264 10 bullet lines that summarises the deltas (what changed). Does not quote large blocks from the changelog verbatim.",
+      "expectations": [
+        "Recap is \u2264 10 lines total",
+        "Uses bullet list format, not prose paragraphs",
+        "Summarises what changed (deltas), not what existed before",
+        "Does NOT quote large blocks from sprint-changelog.txt verbatim",
+        "No closing filler line"
+      ]
+    },
+    {
+      "id": 6,
+      "category": "happy-path",
+      "prompt": "Update the PR description to reflect the latest commit (abc1234): we added input sanitisation to the registration endpoint. Current PR body is in evals/files/pr-description-existing.md.",
+      "files": [
+        "evals/files/pr-description-existing.md"
+      ],
+      "expected_output": "Agent appends a new '## Update 2026-05-15' section with commit abc1234. Does not rewrite or delete any prior section.",
+      "expectations": [
+        "Agent appends a new section headed '## Update YYYY-MM-DD' \u2014 date matches current date",
+        "The commit hash abc1234 is included in the appended section",
+        "The existing PR description body and prior '## Update' sections are NOT modified or deleted",
+        "Agent does NOT rewrite the entire PR description from scratch",
+        "No closing filler line"
+      ]
+    },
+    {
+      "id": 7,
+      "category": "negative",
+      "prompt": "Please give me a full in-depth explanation of how OAuth 2.0 Authorization Code Flow works \u2014 I want every step covered in detail.",
+      "files": [],
+      "expected_output": "Agent provides a complete, detailed explanation covering all steps. Conciseness rules are suspended because the user explicitly requested full detail.",
+      "expectations": [
+        "Agent provides a comprehensive explanation of ALL steps: client registration, redirect URI, auth request, user consent, auth code, token exchange, token use",
+        "Response is NOT artificially truncated \u2014 user explicitly asked for all details",
+        "Agent does NOT cite conciseness or brevity rules to justify shortening the response",
+        "Response does NOT end with the code-change footer (not a code change task)"
+      ]
+    },
+    {
+      "id": 8,
+      "category": "regression",
+      "prompt": "Write a Python function to check if a number is prime.",
+      "files": [],
+      "expected_output": "Agent writes a correct, concise isPrime function and ends with the **What changed** / **Why** / **How to verify** footer. No preamble.",
+      "expectations": [
+        "Response does NOT open with 'Certainly!', 'Sure!', 'Great!' or any preamble",
+        "Function is correct \u2014 returns False for n < 2, uses trial division up to sqrt(n) or equivalent",
+        "Response ends with **What changed** / **Why** / **How to verify** footer",
+        "Response does NOT end with 'Let me know if you want me to add tests!' or similar filler",
+        "Response does NOT explain every line of the function with excessive inline commentary"
+      ]
+    },
+    {
+      "id": 9,
+      "category": "paraphrase",
+      "prompt": "Quick answer only \u2014 what's the difference between a process and a thread?",
+      "files": [],
+      "expected_output": "A concise answer (\u2264 4 lines) on the key difference. No acknowledgement of the 'quick answer' request \u2014 it just IS quick.",
+      "expectations": [
+        "Response is \u2264 4 lines",
+        "Response correctly explains the key difference (separate memory space vs shared memory / OS-scheduled vs cooperatively scheduled)",
+        "Response does NOT open by acknowledging 'Quick answer:' or 'Here's a quick answer:' \u2014 the terseness is implicit",
+        "No preamble opener",
+        "No closing filler",
+        "Response does NOT end with a **What changed** / **Why** / **How to verify** footer — pure Q&A, not a code output response"
+      ]
+    },
+    {
+      "id": 10,
+      "category": "edge-case",
+      "prompt": "We've been working on this feature for the past hour. Here's what we've done so far: set up the FastAPI router, added the Pydantic request/response schemas, wired in the database session dependency, and wrote the POST /users endpoint. What should we tackle next?",
+      "files": [],
+      "expected_output": "Agent gives a direct next-step suggestion without restating the four things the user just listed.",
+      "expectations": [
+        "Response does NOT restate or re-list the four completed items (router, schemas, DB session, POST endpoint)",
+        "Response gives a concrete, actionable next step (e.g. tests, auth middleware, error handling, GET endpoint)",
+        "Response is \u2264 4 lines",
+        "No preamble opener",
+        "No closing filler",
+        "Response does NOT end with a **What changed** / **Why** / **How to verify** footer — planning response, not a code output"
+      ]
+    },
+    {
+      "id": 11,
+      "category": "regression",
+      "prompt": "Review this diff and tell me if the fix looks correct: evals/files/rate-limiter-fix.diff",
+      "files": [
+        "evals/files/rate-limiter-fix.diff"
+      ],
+      "expected_output": "Agent reviews the diff and confirms whether the fix is correct. Does NOT append the code-change footer — this is a code review, not a code output response.",
+      "expectations": [
+        "Response confirms the fix is correct (t > window_start correctly evicts timestamps outside the sliding window)",
+        "Response does NOT end with **What changed** / **Why** / **How to verify** footer — this is a review task, not a code output response",
+        "Response does not paste the full contents of rate_limiter.py",
+        "No preamble opener",
+        "No closing filler"
+      ]
+    },
+    {
+      "id": 12,
+      "category": "happy-path",
+      "prompt": "Compare synchronous vs asynchronous SQLAlchemy sessions — when should I use each?",
+      "files": [],
+      "expected_output": "Agent answers with structured output (table or clearly delineated bullet comparison) rather than a dense prose paragraph. Concise, within line limits.",
+      "expectations": [
+        "Response uses structured output — a comparison table or clearly delineated bullet sections — not a prose paragraph",
+        "Response is concise and does not expand into a full SQLAlchemy tutorial",
+        "No preamble opener",
+        "No closing filler",
+        "Response does NOT end with a **What changed** / **Why** / **How to verify** footer — pure Q&A comparison"
+      ]
+    }
+  ]
+}