Release Notes Action Gate

    1

    Every SaaS vendor announcement mixes mandatory migrations, optional features, and changes that never touch your setup — all in the same cheerful tone.

    $9

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Release Notes Action Gate

    Example session with this skill installed

    Example request (fictional vendor, reproducible in the bundled examples):

    Review this vendor note against a profile using Legacy Export and SSO. Extract the evidence, list dates, and show which source sentences still need human checking. Do not make the final applicability decision automatically.

    Release note provided

    The following is a sample vendor release note (fictional). Starting 2027-01-15, workspace admins using Legacy Export must migrate to Export v2. The Legacy Export endpoint will stop responding on 2027-03-01, and any integration still calling it after that date will fail. A new dashboard color theme is also available for all workspaces.

    • Read your context and instructions
    • Compiled the release notes action

    Actual runner report (generated with --review-date 2026-10-29; no anchor supplied):

    RELEASE NOTES ACTION GATE - EVIDENCE REPORT

    Review date: 2026-10-29
    Anchor date (for relative dates): not provided
    Coverage: 4 source sentences - 2 extracted as evidence, 2 not extracted.
    Requirement candidates detected: 2 (MATCH 2 / NO-MATCH 0)
    Detected candidates are NOT a guaranteed-complete list - the mandatory coverage check in SKILL.md accounts for every source sentence before the action list is final.

    SCOPE: this report extracts evidence and matches it against
    your setup profile only. It does not decide whether a change
    applies to you. Use SKILL.md for the judgment workflow.

    [1] MATCH (Legacy Export)
    Sentence: Starting 2027-01-15, workspace admins using Legacy Export must migrate to Export v2.
    Cue: "must" (markers: must)
    Date mentions: 2027-01-15 -> 2027-01-15 [RESOLVED] absolute date from source

    [2] MATCH (Legacy Export)
    Sentence: The Legacy Export endpoint will stop responding on 2027-03-01, and any integration still calling it after that date will fail.
    Cue: "will stop" (markers: will stop)
    Date mentions: 2027-03-01 -> 2027-03-01 [RESOLVED] absolute date from source

    DATE MENTIONS

    Every date expression found in the extracted sentences, in
    source order. A date mention is NOT necessarily a deadline -
    it can be an announcement date, a start date, or a deadline.
    Classifying each date's role is Stage 2 work (SKILL.md).

    • 2027-01-15 -> 2027-01-15 [RESOLVED] (from item 1) absolute date from source
    • 2027-03-01 -> 2027-03-01 [RESOLVED] (from item 2) absolute date from source

    UNEXTRACTED SOURCE SENTENCES (check these in Stage 2)

    Every source sentence that produced NO extracted item,
    numbered as in the source and quoted verbatim. Work
    through this list in SKILL.md Stage 2: each sentence is
    either consciously dismissed with a stated reason, or it
    is an item the extractor missed - this list is the
    checklist for that mandatory coverage step.
    [1] The following is a sample vendor release note (fictional).
    [4] A new dashboard color theme is also available for all workspaces.

    The workflow in SKILL.md then turns this evidence into a human-reviewed action list. This report is deterministic extraction output, not an automatically final compliance decision.

    list - the mandatory coverage check in SKILL.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Release notes bury the two sentences that matter. This finds them.

    Every SaaS vendor announcement mixes mandatory migrations, optional features, and changes that never touch your setup — all in the same cheerful tone. Reading them raw takes time and still leaves the only question that matters unanswered: what do we have to do, by when, and where does it say that?

    Release Notes Action Gate answers it with receipts. You save the note as a text file, add a five-minute profile of the features your team actually uses, and get back an evidence table plus a guided workflow that turns it into an action list — every item carrying the exact source cue it came from.

    How it works

    1. Extract (deterministic runner). release_notes_gate.py (Python 3, standard library only, no network) scans the note for requirement markers — must, required, will be removed, discontinued, deprecated, action required, breaking change and more — and pulls out each matching sentence with its cue phrase.
    2. Collect date mentions. Dates in those sentences are gathered as date mentions — an announcement date and a deadline deliberately look alike at this stage; the workflow classifies each date's role. Absolute dates resolve to ISO form. Relative dates ("in 30 days") resolve only against an anchor date you supply — the note's publication or effective date — and anything that cannot be pinned to a day ("next quarter", "end of the year") is flagged UNRESOLVED, never guessed.
    3. Match your setup. Each item is checked against your setup profile and marked MATCH or NO-MATCH. NO-MATCH items stay in the report — they are dismissed by a human decision, never silently dropped.
    4. Judge (skill workflow). SKILL.md walks you or your AI assistant through the applicability gate — Applies / Possibly applies / Not applicable / Unknown — then builds the prioritized checklist: confirmed-required items under Do, everything else under Verify, with owners left unassigned unless you named one.

    Tested like it costs a hundred times more

    This is a $9 skill with a four-figure regression suite. 2,202 local automated tests across ten Python files passed against this v1.0.5 runner, including 163 new v1.0.5 event-time cases/CLI controls in addition to the prior 2,039. Tests exercise extraction, date mentions, negation traps, ambiguous clock origins, CLI errors and output safety. All three sample JSON outputs and all three reports were independently reproduced byte-for-byte. A passing test suite is not evidence that every possible English paraphrase is handled or that Agensi has approved this listing.

    Read this before you buy — what the runner does NOT do

    New independent-check fix (v1.0.5): The compliance cutoff depends on when you open the welcome email; SSO users must migrate within 30 days. and The due date is calibrated to the moment an invitation is accepted; SSO users must migrate within 30 days. now remain UNRESOLVED, with normalized=null and human review required, even with a supplied publication anchor. The same applies when the clock is stated immediately after the migration sentence. New and opposing controls were tested, but the recognition is a heuristic, not complete English understanding.

    Important date-origin limitation (v1.0.5): When the original note says The deadline is anchored to your first login; SSO users must migrate within 30 days, the period is UNRESOLVED — it is not calculated from the publication date. The same applies when the note states First login is day zero. immediately before or after SSO users must migrate within 30 days., including neighboring list items. Two descriptions of the same 30 days after first login obligation will not produce a conflicting publication-based deadline. However, these checks are heuristic: a clock set in a distant heading/paragraph/list entry or an unfamiliar phrase can still be missed. Before using any extracted date in an action list, Stage 2 requires reading the entire source and verifying the true start event and date role. RESOLVED means normalized, not a confirmed deadline.

    Open-source rights: The included software and documents are supplied with an MIT licence, which broadly permits reuse, modification and redistribution of copies. Agensi also imposes separate contractual limits on sharing or exporting a purchased skill through its service; this listing does not promise exclusive code rights or claim that those terms cancel the MIT permission. Buyers should consult both the bundled LICENSE.txt and Agensi's terms. The creator retains ownership; platform approval and buyer-use conditions are administered by Agensi.

    The runner performs evidence extraction and keyword matching only. It does not decide whether a change applies to your setup, and it does not understand meaning the way a person does. A MATCH is a strong hint; the verdict comes from the workflow step, where your setup knowledge is applied to the evidence. We state this plainly because a tool that claims to make that judgment from keywords alone would be overpromising — and because the evidence table is exactly what makes the human judgment fast and auditable.

    What's inside

    • SKILL.md — Stage 0 extraction plus the five-stage workflow (source boundary → applicability gate → action mapping → output → delta updates)
    • scripts/release_notes_gate.py — the deterministic extractor (stdlib Python, no network, clear error messages)
    • examples/ — three complete fictional cases with real runner outputs (a matching breaking change, a no-match removal notice, an unresolvable relative date), plus a worked end-to-end action list showing the finished deliverable for the first case
    • references/INPUT_CONTRACT.md — exact input formats with samples
    • references/SETUP_PROFILE.md — how to write your setup profile in about five minutes
    • tests/ — 2,202 self-tests (1,781 earlier checks + 258 Round 15 sale-gate tests + 163 v1.0.5 independent recheck tests) covering extraction, date handling, matching, bad-input errors, ambiguous negations, output-path safety, and simulated agent temp-folder runs of all three worked examples
    • QA.md — the pre-release check record, including a defect that was found and fixed

    Who it's for

    SaaS administrators, IT operators, and AI-agent operators who review vendor change announcements and need required work separated from optional features and non-issues — with the source quote attached to every item, so any decision can be re-checked months later.

    All examples use fictional vendors and scenarios. No productivity, compliance, or outcome results are promised. Outputs are drafts for human review.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Productivity.

    Important: This is an evidence-extraction and review workflow, not an automatic compliance decision. These limits apply before you buy:

    • No live Agensi purchase/connection test yet. Local example runs and an isolated temporary-folder simulation have been tested; a real purchase → connection → run has not been verified. A compatible agent environment must be able to execute Python 3.
    • It can miss instructions. The runner recognizes English requirement markers and selected imperatives, not every possible phrasing. You or your agent must review the complete source, including the report's unextracted sentences. Zero detected requirements does not mean zero actual obligations.
    • v1.0.5 independent check repair. A new independent test of v1.0.4 found three more false RESOLVED outcomes: compliance cutoff depends on when you open the welcome email, due date is calibrated to the moment an invitation is accepted, and adjacent deadline is pegged to the moment your invitation is accepted. The revised runner does not apply the publication anchor to a relative period when a named deadline/temporal head refers to a customer-specific when/whenever/moment/instant/point at which time clause, even if the event noun or linking verb was not previously recognized. This can conservatively over-flag unrelated declarations. 163 new checks were added (159 case fixtures and four direct CLI checks). Unrecognized and nonadjacent origin language can still be missed, so Stage 2 full-source verification is mandatory.
    • Round 15 safety repair (v1.0.4). The independent Round 13 sale review exposed named-deadline, duplicate-period and adjacent-sentence clock-origin misses. These now stay UNRESOLVED/review-candidate rather than becoming a guessed calendar day, and compact bracketed role/work assignments no longer certify NEGATED. This release adds 258 fixed checks; original examples are unchanged. Limits: nonadjacent source clocks and unknown English paraphrases still require human Stage 2 verification.
    • Historical Round 13/14 safety repair (v1.0.3). The Round 12 external check reproduced sentence-initial participial origins (Anchored/Keyed/Tied/Pegged/Linked to your first login, ...) resolving against the publication anchor, and an owner/work statement with a specialist role subject (our SSO specialists oversee certificate renewals) certified as no action. A further independent recheck of the interim v1.0.2 build then found five more explicit event-clock deadlines (calculation starts from first sign in, time zero, interval starts on first invoice paid, deadline varies by first login, depending on when you register). All are repaired in v1.0.3: such periods stay UNRESOLVED and owner/work statements route to REVIEW_REQUIRED. These safeguards are heuristic: new language can still escape detection.
    • Round 12 independent recheck repair (v1.0.1). Earlier local tests did not catch owner/work sentences that appear before no-user-action claims, or explicit first login as the reference / day zero start phrases. These are now conservatively routed to REVIEW_REQUIRED / UNRESOLVED. The historical AG-G-193 test expectation was changed from NEGATED to REVIEW_REQUIRED with a recorded justification: admins actively rotate tokens even though end users need not act. New tests include opposing no-action and bare-date controls.
    • Round 11 safety repair, not semantic omniscience. The independently reproduced misses (No action is required; our crew oversees/owns certificate renewals and Tallied from the first login, ... within 30 days) are fixed. The current 2,202-test suite includes 103 Round 11 cases (67 independent inputs and 36 added combinations/guards), 54 Round 12 independent recheck cases, 148 Round 13 external cases and 53 Round 14 recheck cases. A no-action assertion does not dismiss an explicit actor/task assignment, and a source-defined starting event blocks a made-up publication-anchored deadline. These safeguards are heuristic: new language can still escape detection. Previous Round 10 safety repair. The runner now conservatively reviews post-negation owner/work assignments such as an infrastructure crew owning certificate rotation, and leaves source-defined invoice/enrollment/clock-origin relative periods UNRESOLVED rather than applying the publication anchor. The 351 additional regression checks cover both defect families and their opposing controls. Round 9 safety closure. Added regression coverage for unfamiliar imperative commands (harden auth settings, pin dependency versions), source-defined timer origins (clock begins, timer starts), date-bearing subordinate openings (Given that ...), and Post ... / As measured from ... / With ... finished source clocks. Unknown phrasing can still escape a deterministic recognizer. A RESOLVED calculation is not proof of a vendor deadline: verify the starting event and the original note in the judgment stage. Descriptive historical text such as the Service Desk page was updated may be over-flagged as REVIEW_REQUIRED. A state sentence with by Friday remains REVIEW_REQUIRED because it carries an unresolved weekday date, even if Friday names no actor.
    • Dates need human judgment. Explicit counting-from / calculated-as-of / period-start conditions can override the publication anchor, even when their start event has no stated date; such periods stay UNRESOLVED. Date mentions are not automatically deadlines. A relative period such as "in 30 days" resolves against your supplied publication/effective anchor only when it stands at sentence end with no competing start or continuation; otherwise it is UNRESOLVED for review. An opening "From / As of / With effect from / On 2026-11-01" can block resolution even without a comma; so can a source event such as "At first login" or "As soon as an extension begins", and an opening "With the contract effective 2026-12-15".
    • "No action required" is not a guarantee. Mixed action/assignment wording, including an owner/task pair separated by |, an unfamiliar clause-leading work command such as "rekey API credentials" or "archive logs", a task noun plus a named team, and actor or direction phrases such as "on behalf of", "with help from", or "at the direction of", goes to REVIEW_REQUIRED. Conversely, genuinely harmless status or negated-work statements may sometimes be over-flagged, particularly when a state phrase has an unfamiliar trailing adjunct or a relative period opens with At ... / From ... that might not actually be a clock start. You must adjudicate every review item.
    • Negation scope remains heuristic. v1.0.1 recognizes several human work assignments before the first negation phrase, including Our admins renew certificates; no user action is required. However an unfamiliar actor, verb, or work object can still escape; the full-source Stage 2 review remains mandatory.
    • Script-size maintenance trade-off. The single-file Python runner is far larger than Agensi's advisory under-20-KB logic-per-script recommendation. This is guidance, not a hard upload limit; the self-contained runner is retained to support execution as one temporary file. The measured size and testing scope are recorded in QA.md; inspect the script if your agent has execution or review limits.
    • No live monitoring or account changes. You supply English-language release-note text and a setup profile; the skill does not fetch URLs, monitor vendors, or change settings.
    • Output files may be replaced. The runner protects the two input files, but replaces other existing output files (with a warning); choose new output names.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 12 days ago

    Software
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    File System Read — reads only the notes file and profile file you pass on the command line. Execute Commands — running the Python script locally. The runner itself makes no network calls and accesses no accounts.

    Listed12 days ago

    What's inside

    Frequently Asked Questions

    No — and be wary of tools that claim to. The runner extracts the evidence (requirement sentences, cues, date mentions, profile matches). The applicability decision is made in the SKILL.md workflow step, where your setup knowledge meets that evidence. That split is deliberate: it keeps every decision traceable to a source quote.

    You need an Agensi account and a compatible AI agent connected to it — that connection is how the skill reaches you and how it runs. You do not need to install anything yourself: the runner is a single Python 3 file using only the standard library, with no extra Python packages to add. It does require an agent environment that can run Python 3 — that is what executes the runner.

    No. Paste or save the note's text yourself. This is deliberate: the skill works only from source text you can see and keep, so every cue in the output can be checked against it.

    Unresolvable periods are flagged UNRESOLVED in the date-mentions list, with the original wording preserved; do not treat a machine-RESOLVED date as a confirmed deadline. The runner does not intentionally guess calendar dates: however, an unrecognized source-defined origin can still make an anchor-based normalization wrong, so always verify the start event in Stage 2. "In 30 days" does resolve — but only against an anchor date you supply (the note's publication or effective date, which is what the vendor counts from) and only when it stands as the last content of its sentence; if anything follows it in its sentence, or the note counts it from a starting event of its own, it is reported UNRESOLVED for Stage 2 to confirm — exactly as stated in Known Limitations above. It is never resolved against the day you happen to review the note.

    Yes. Write the profile from what you know (the guide takes about five minutes), and treat gaps as the workflow intends: items you cannot judge stay Unknown, and the output lists precise questions instead of guesses.

    Two different things, kept separate. The Python runner itself makes no network calls and sends nothing anywhere — it reads the two files it is given and writes its output; there is no network code in it at all. Separately, the notes and profile you hand to your AI assistant are processed by that AI provider as part of your conversation with it, and the Agensi connection records skill loads and usage under Agensi's own privacy policy. So: the script sends nothing; your assistant and Agensi handle what you give them under their own policies — do not paste secrets or credentials into a release note you ask an assistant to process.

    New in Productivity