Skill This
AI Skill Report Card

Mapping Human Tasks to AI Delimiters

A-83·Nov 2, 2026·Source: Web
Quick Start14 / 15

Given a human task description, produce a specification using reserved, collision-proof delimiters ⟦TAG attr="val"⟧ ... ⟦/TAG⟧:

XML
⟦TASK_SCOPE id="T1" objective="reconcile invoices" acceptance="all line items matched or flagged"⟧
  ⟦HUMAN_STEP id="H1" seq="1" actor="accountant" action="open spreadsheet, compare invoice to PO"⟧
    ⟦ARTIFACT_IO direction="in" schema="invoice.csv,po.csv"⟧
    ⟦ARTIFACT_IO direction="out" schema="reconciled.csv"⟧
  ⟦/HUMAN_STEP⟧
  ⟦MAPPING_RECORD id="M1" from="H1" to="AI_ACT1" mapping_type="direct" confidence="0.9" gap="none"⟧
  ⟦ADAPTATION_RECORD id="A1" reason="model can't access local spreadsheet" original="manual open file" adapted="file uploaded via API"⟧
  ⟦AI_MODEL_BINDING id="B1" model="gpt-x" prompt_id="recon_v1" io_schema="invoice.csv,po.csv->reconciled.csv"⟧
  ⟦VERIFICATION_ACCEPTANCE id="V1" method="row diff" expected="0 unmatched" pass_fail="TBD"⟧
  ⟦EXCEPTION_ROLLBACK id="E1" trigger="schema mismatch" fallback="route to human review"⟧
⟦/TASK_SCOPE⟧
Recommendation▾
Example 2 and 3 outputs are described rather than fully shown as concrete tag chains—render at least one more full worked output to match Example 1's rigor
The 10 Semantic Delimiters
#DelimiterPurpose
1TASK_SCOPEBounds the whole task: objective, constraints, acceptance criteria
2HUMAN_STEPOne atomic human action
3DECISION_GATEBranch, approval, or conditional point
4ARTIFACT_IOInput/output artifact for a step or binding
5MAPPING_RECORDLinks a human step/action to an AI action/function/prompt
6TRANSITION_RECORDHandoff or state change between steps, actors, or systems
7ADAPTATION_RECORDRequired workflow change due to model/tool constraints
8AI_MODEL_BINDINGExact AI binding: model, prompt, tools, params, IO schema
9VERIFICATION_ACCEPTANCETests/checks proving AI output is acceptable
10EXCEPTION_ROLLBACKFailure handling, fallbacks, escalations, undo
Canonical Chain Order
TASK_SCOPE
  └── HUMAN_STEP
        └── MAPPING_RECORD
              └── TRANSITION_RECORD
                    └── MAPPING_RECORD
                          └── ADAPTATION_RECORD
                                └── MAPPING_RECORD
                                      └── ADAPTATION_RECORD
                                            └── TRANSITION_RECORD
                                                  └── MAPPING_RECORD
                                                        └── AI_MODEL_BINDING
                                                              └── VERIFICATION_ACCEPTANCE
                                                                    └── EXCEPTION_ROLLBACK

This encodes: mapped and/or transitioned, then mapped and/or adapted (repeated as many times as needed), then bound to a model, then verified, then protected by a rollback.

Workflow14 / 15

Progress:

  • Step 1: Define TASK_SCOPE — objective, constraints, acceptance criteria
  • Step 2: Decompose the process into atomic HUMAN_STEP records (one actor, one action, one input/output each — split anything with "and")
  • Step 3: Add ARTIFACT_IO for each step's inputs/outputs
  • Step 4: Create MAPPING_RECORD linking each human step to a candidate AI action; include source_id, target_id, mapping_type, confidence, gap
  • Step 5: Insert TRANSITION_RECORD wherever an actor/system handoff occurs (do NOT merge with mapping records)
  • Step 6: If the AI can't fully replicate the step, add an ADAPTATION_RECORD (reason, original, adapted), then re-map with a new MAPPING_RECORD — repeat this map→adapt cycle as many times as needed, documenting each cycle separately
  • Step 7: Once stable, bind to infrastructure with AI_MODEL_BINDING (model, prompt_id, tools, params, io_schema)
  • Step 8: Close with VERIFICATION_ACCEPTANCE that traces back to the original acceptance criteria in TASK_SCOPE
  • Step 9: Define EXCEPTION_ROLLBACK upfront alongside the binding (never deferred to post-deployment)
  • Step 10: Validate — every tag has id, parent, seq, version; every open tag has a matching close; no orphan records

For multi-task pipelines: wrap the whole pipeline in a master TASK_SCOPE, nest each sub-task as a child TASK_SCOPE with parent pointing to the master, and use TRANSITION_RECORD at every boundary where one child scope's output triggers the next.

For conditional branches: use DECISION_GATE with condition, evaluation_type (human/automated), and choices mapping each outcome to a distinct TRANSITION_RECORD → downstream TASK_SCOPE. Every choice listed must have a matching TRANSITION_RECORD destination.

Recommendation▾
Clarify how 'confidence' values should be derived/calibrated rather than left as arbitrary floats
Required Fields Checklist

Universal (every tag): id, parent, seq, version

MAPPING_RECORD specifically: source_id (or from), target_id (or to), mapping_type (direct/partial/proxy), confidence, gap

Examples15 / 20

Example 1: Input: "Human manually reconciles invoices in a spreadsheet. Give me the delimiters to map this to an AI workflow." Output: The worked TASK_SCOPE/HUMAN_STEP/MAPPING_RECORD/ADAPTATION_RECORD/AI_MODEL_BINDING/VERIFICATION_ACCEPTANCE/EXCEPTION_ROLLBACK chain shown in Quick Start.

Example 2: Input: "Data entry clerk reads PDF invoice attachments and types totals into a CRM." Output: Full chain including two mapping records, a transition record (webhook trigger), one adaptation record (no desktop/CRM access), binding to gemini-2.5-pro with pdf_extraction_helper tool, JSON schema verification, and Slack-alert rollback.

Example 3: Input: "Route support tickets: refunds ≥$500 need human compliance sign-off, <$500 auto-process." Output: Master TASK_SCOPE → DECISION_GATE with two choices → two TRANSITION_RECORDs → two child TASK_SCOPEs (TS_HUMAN_REVIEW, TS_AUTO_PROCESS), each self-contained with its own verification/rollback, with the automated branch's rollback escalating upward to the human branch on failure.

Recommendation▾
Add a minimal validation script or pseudocode snippet for the orphan/balance check referenced in Best Practices, since it's mentioned but not demonstrated
Best Practices
  • Keep HUMAN_STEP atomic — one actor, one action, one input/output; split on "and"
  • Reference real source_id/target_id everywhere — no orphan records
  • Document every map→adapt cycle separately, even if repeated; never collapse history
  • Close every chain with VERIFICATION_ACCEPTANCE that maps back to TASK_SCOPE.acceptance
  • Define EXCEPTION_ROLLBACK at spec time, alongside the model binding — not post-deployment
  • Use reserved, content-safe delimiter tokens (⟦...⟧, <<<...>>>) instead of ### or ---
  • Validate bracket/tag balance before treating a spec as complete (see validation script pattern: track an open/close stack, flag orphan closes and unclosed blocks)
  • For pipelines, isolate branch scopes — Branch A must never reference tokens/variables declared in Branch B
Common Pitfalls
  • Generic delimiters (###, ---) collide with body text/code — don't use them
  • Glossing over gaps — don't silently swap a human step for an AI one without an ADAPTATION_RECORD
  • Conflating handoffs and mappings — TRANSITION_RECORD (system/actor handoff) and MAPPING_RECORD (human↔AI functional alignment) are distinct; don't merge them
  • Delayed error handling — don't defer EXCEPTION_ROLLBACK to "after launch"; define it with the binding
  • Ambiguous nesting — always verify every opening tag has exactly one matching closing tag before finalizing
  • Missing universal fields — a tag without id/parent/seq/version fails the audit trail
  • Incomplete DECISION_GATE routing — every listed choice must resolve to a TRANSITION_RECORD and downstream scope; unmatched choices are a validation failure
0
Grade A-AI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
14/15
Workflow
14/15
Examples
15/20
Completeness
18/20
Format
14/15
Conciseness
13/15
Save to My Skills