Mandatory confidence-scored recommendation gate for AI actions
Classifies every action as ROUTINE, NON-ROUTINE, or NON-ROUTINE-HIGH, then emits a structured 7-field recommendation block (ACTION, CLASS, CONFIDENCE, IMPACT, REVERSIBILITY, FILES, RECOMMENDATION). Non-routine actions are gated — the AI stops and waits for human approval before proceeding. Three invocation modes: GATE (blocking), LOG (non-blocking file-write summary), and ASSESS (standalone evaluation).
Confidence-Gated Recommendation Gate
Tags: Introduced in Beta, Safety & Version Control, Confidence Gate, Recommendation, Action Classification, Cowork
TL;DR
What It Does
Mandatory confidence-scored recommendation gate for AI actions. Classifies every action as ROUTINE, NON-ROUTINE, or NON-ROUTINE-HIGH, then emits a structured 7-field recommendation block (ACTION, CLASS, CONFIDENCE, IMPACT, REVERSIBILITY, FILES, RECOMMENDATION). Non-routine actions are gated — the AI stops and waits for human approval before proceeding. Three invocation modes: GATE (blocking), LOG (non-blocking file-write summary), and ASSESS (standalone evaluation).
How It Works
Five steps. (0) Determine mode — route to GATE, LOG, or ASSESS path. (1) Classify action — use provided class or apply the SP04 decision tree (escalation rules E-1 through E-4). (2) Emit recommendation block — 7-field format, FileWrite details if files affected, confidence-to-decision mapping (4-5 PROCEED, 3 CAUTION, 1-2 HOLD). NON-ROUTINE-HIGH always HOLD. Auto-proceed exceptions (AP-1 through AP-4) emit block but skip wait. (3) File-write enhanced summary — LOG mode collects routine writes for batch at handoff. (4) Execute/return — return decision to calling recipe.
What To Expect
A structured recommendation block in chat for non-routine actions. PROCEED actions log and continue. CAUTION actions emit the block and execute in the same response. HOLD actions stop execution and wait for your explicit approval. LOG mode silently collects file-write summaries for the session handoff.
Best Results When You
Let the gate run without suppressing it. The recommendation blocks are the safety net — they surface what the AI is about to do, how confident it is, and how reversible the action is. Override with E-4 only when you’re confident the action class is well-understood.
Typical Time
1–5 minutes. GATE mode with HOLD adds human review time. LOG and ASSESS modes complete instantly.
Difficulty
Intermediate — the classification taxonomy, escalation rules, and auto-proceed register require understanding the SP04 design. The gate itself is automatic once configured.
How To Start
This recipe is usually invoked by other recipes, not directly by a human. CWK-046 calls it before pushes, CWK-002 during handoff file writes, CWK-005 during checkpoints, and CWK-007 during batch operations. You can also invoke it directly via ASSESS mode to evaluate a proposed action without executing it.
STEP 1Configure the Gate
Choose the mode, describe the action, classify it, and set confidence.
Available parameters
-
action_description
· string · required
One-line description of the action to be evaluated. -
action_class
· string · required · options: ROUTINE | NON-ROUTINE | NON-ROUTINE-HIGH
Classification per the SP04 taxonomy. ROUTINE = expected session ops. NON-ROUTINE = irreversible, critical-file, or external-system actions. NON-ROUTINE-HIGH = global push, file deletion, bulk edit, cookbook/spec overwrite. -
confidence
· integer · optional · range: 1–5
AI self-assessed confidence. Required for GATE mode. 5=Certain, 4=High, 3=Moderate, 2=Low, 1=Very Low. Maps to PROCEED (4-5), CAUTION (3), or HOLD (1-2). -
files_affected
· list · optional
File paths affected by the action. Triggers the FileWrite enhanced summary when non-empty. -
file_operation
· string · optional · options: create | overwrite | append | edit
Type of file operation for the FileWrite summary. -
reversibility
· string · optional · options: easy | moderate | hard | irreversible
How difficult it is to undo this action. Included in the recommendation block. -
calling_recipe
· string · optional
ID of the recipe invoking the gate (e.g., “CWK-046”). Used for audit trail and chained non-routine detection (E-3). -
mode
· string · optional · default: “GATE” · options: GATE | LOG | ASSESS
GATE = blocking decision with human approval for HOLD. LOG = non-blocking file-write summary collected for batch. ASSESS = standalone evaluation, returns block without executing. -
auto_proceed_id
· string · optional
If this action matches an Auto-Proceed Exception Register entry (AP-1 through AP-4), provide the AP-ID. Full recommendation block still emitted for audit, but execution proceeds without waiting.
Invocations
Gate a non-routine action
#H->AI::Directive: (Execute CWK-051 with mode=“GATE”, action_description=“Push project to LOCAL remote”, action_class=“NON-ROUTINE”, confidence=4, reversibility=“moderate”, calling_recipe=“CWK-046”)
Log a routine file write
#H->AI::Directive: (Execute CWK-051 with mode=“LOG”, action_description=“Append lesson learned LL-045”, action_class=“ROUTINE”, files_affected=[“project/CFT-PROJ-CP-067_LESSONS-LEARNED.txt”], file_operation=“append”)
Assess an action without executing
#H->AI::Directive: (Execute CWK-051 with mode=“ASSESS”, action_description=“Delete stale branch artifacts”, action_class=“NON-ROUTINE-HIGH”, confidence=3, files_affected=[“project/old-branch/”], reversibility=“irreversible”)
Usage Examples
Example 1: GATE mode — push with PROCEED
USER
Push to the local remote.
CAT
Evaluating action for confidence gate. ACTION: Push project to LOCAL remote. CLASS: NON-ROUTINE. CONFIDENCE: 4 — High. IMPACT: Commits pushed to remote; remote state will match local main. REVERSIBILITY: moderate. FILES: 3 files committed. RECOMMENDATION: PROCEED. Proceeding with push.
Example 2: LOG mode — routine writes batched at handoff
USER
Add this lesson learned.
CAT
CWK-051 LOG mode runs silently — FileWrite summary collected for handoff batch. At session handoff, the batch appears: ROUTINE WRITES THIS SESSION: 4. Each entry lists the file path, operation, and a brief summary.
Example 3: GATE mode — NON-ROUTINE-HIGH forces HOLD
USER
Delete all the old branch artifacts.
CAT
Evaluating action for confidence gate. ACTION: Delete old branch artifacts (12 files). CLASS: NON-ROUTINE (HIGH). CONFIDENCE: 4 — High. IMPACT: 12 files permanently deleted, no git recovery if not committed. REVERSIBILITY: irreversible. FILES: project/old-branch/file1.txt, + 11 more. RECOMMENDATION: HOLD FOR REVIEW. Awaiting your approval before proceeding. NON-ROUTINE-HIGH actions always require explicit approval regardless of confidence score.
How AI Reads This Recipe
When this recipe is triggered, the AI evaluates a proposed action through the confidence gate. The AI should:
1. Determine mode. Read the mode parameter (default GATE). If LOG, skip to step 3 for FileWrite summary only. If ASSESS, run steps 1–2 but do not execute — return the recommendation block only. If GATE, continue to step 1.
2. Classify the action. Use the provided action_class. If not provided, apply the SP04 decision tree: external push → NON-ROUTINE or NON-ROUTINE-HIGH; file deletion → NON-ROUTINE-HIGH; critical-file overwrite (CLAUDE.md, cookbook, spec, core-doc, index) → NON-ROUTINE or NON-ROUTINE-HIGH for bulk; STATE.txt or ROADMAP.md modification → NON-ROUTINE; first-time execution (E-1) → escalate; scope exceeding thresholds (E-2: >10 files or >500 lines) → escalate; chained non-routine from parent (E-3) → inherit; otherwise ROUTINE. Check for Richard override (E-4).
3. Emit recommendation block. Build the 7-field block: ACTION, CLASS, CONFIDENCE, IMPACT, REVERSIBILITY, FILES (max 5, then “+ N more”), RECOMMENDATION. If files_affected is non-empty, append FileWrite block (FILE, OPERATION, LINES, SUMMARY). Map confidence: 5 or 4 → PROCEED, 3 → PROCEED WITH CAUTION, 2 or 1 → HOLD FOR REVIEW. NON-ROUTINE-HIGH always HOLD. If auto_proceed_id matches AP register, emit block with “AUTO-PROCEED (AP-N)” and skip wait. Route: PROCEED → log and continue; CAUTION → emit and execute; HOLD → emit and stop.
4. File-write enhanced summary (LOG mode). Collect the FileWrite summary into the session batch per OQ-2. Routine summaries render as a compact FileWriteBatch block in the handoff (CWK-002) or COM footer (CWK-012). Non-routine file writes always emit inline in the recommendation block.
5. Execute or return. ASSESS mode: return the recommendation block without executing. GATE mode: return decision (PROCEED, CAUTION, HOLD, or REJECTED) to the calling recipe. LOG mode: return to calling recipe with no blocking.
When To Use This Recipe
Use this recipe before any non-routine AI action — pushes, deletions, critical-file overwrites, bulk edits, and external-system interactions. It runs automatically when called by other recipes (CWK-046, CWK-002, CWK-005, CWK-007, CWK-012). Use ASSESS mode to evaluate a proposed action without committing to it.
Do Not Use When
The action is clearly routine and no file-write summary is needed. For routine operations where you want a file-write audit trail, use LOG mode instead of skipping the gate entirely.
FAQ
What’s the difference between GATE, LOG, and ASSESS modes?
GATE is blocking — it emits a recommendation block and waits for approval if the decision is HOLD. LOG is non-blocking — it collects a file-write summary for batch rendering at handoff without interrupting the workflow. ASSESS returns a recommendation block without executing the action, letting you or the calling recipe decide next steps.
Why does NON-ROUTINE-HIGH always produce HOLD?
NON-ROUTINE-HIGH covers the most dangerous action classes — global pushes, file deletions, bulk edits, and cookbook/spec overwrites. These are irreversible or high-impact enough that even high confidence (4-5) should not bypass human review. The hardcoded HOLD ensures a human always approves these actions.
What are the Auto-Proceed exceptions?
Four registered exceptions where delay poses more risk than the action itself: AP-1 (COM emergency save at 90% token usage), AP-2 (stale .git/index.lock cleanup per LL-050), AP-3 (PAT scrub after push per LL-208), and AP-4 (auto-checkpoint at 70% token usage). Each must satisfy all three criteria: time-sensitive (AP-C1), low controversy (AP-C2), and recoverable or net-positive (AP-C3). The full recommendation block is still emitted for audit.
How do the escalation rules work?
Four escalation rules can upgrade a ROUTINE action to NON-ROUTINE. E-1: first-time execution of curated action types (first commit in a new clone, first push, first critical-file overwrite). E-2: scope exceeds thresholds (>10 files in one commit, >500 lines changed in one file). E-3: chained non-routine — if the parent recipe in the call chain already has a NON-ROUTINE gate, the child inherits. E-4: Richard override — a per-session directive that bypasses the gate entirely for a specified action type.
Where is the design document?
research/SP04-CONFIDENCE-GATE-DESIGN.txt contains the full SP04 design including the action taxonomy, confidence scale rationale, escalation rule derivations, and the auto-proceed exception criteria. The recipe implements the design; the design doc explains the “why.”
Version History
v1.00a — 2026-04-10
Initial creation at P061-H028 (SP04 Phase 2). Three modes (GATE/LOG/ASSESS), three action classes (ROUTINE/NON-ROUTINE/NON-ROUTINE-HIGH), five confidence levels with hardcoded thresholds, four auto-proceed exceptions (AP-1 through AP-4), four escalation rules (E-1 through E-4), 7-field recommendation block, FileWrite batch for routine writes (OQ-2).
Get this recipe with CRAFT for Claude Cowork
Cowork recipes ship bundled with CRAFT for Claude Cowork — there’s no separate download. Clone the framework once, and your AI runs every recipe automatically when invoked.
Pull anytime to stay on the latest version — free to clone, no login or email required.
Then start your session
Once CRAFT is in your project folder, open a new Cowork session and ask Claude to initialize. For example:
You
Please initialize my CRAFT session.
Claude
CRAFT session ready. Your project is loaded, your persona is active, and your recipes are available. What would you like to work on?
Not familiar with Git? Download as a ZIP
No command line needed. Just download, move, and unzip:
- Open the CRAFT framework repo on GitHub.
- Click the green Code button, then choose Download ZIP.
- Move the downloaded ZIP into your Claude Cowork project folder.
- Unzip it: double-click on Mac, or right-click → Extract All on Windows.
