# Task Shaping > Judge a consequential task against a conservative Flash-class reference execution profile, then choose direct execution, guarded verification, or a transformation of oversized, branchy, long-context, weakly observable, or completeness-sensitive work into locally verifiable units that can be reconstructed without losing global constraints. Use when deciding one agent versus bounded Cells, an agent loop varies or misses obligations, a verifier skips claims despite seeing evidence, a task needs semantic decomposition before orchestration, a harness needs a dependable task primitive, or asking "can an agent reliably do this?", "how should this task be split?" / "任务怎么拆才能稳定" / "稳定能力包络". Do not use for ordinary one-step planning, proving a model capability without repeated evidence, choosing provider/concurrency/budget, executing a prepared queue, or replacing a domain skill that already owns the semantic partition. - Skill: `lidessen/task-shaping` (Agent Skill, multi-file: 7 files) - Install (CLI): `npx skillmds@latest add lidessen/task-shaping` - Raw SKILL.md: https://api.skillmd.com/api/skills/lidessen/task-shaping/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: lidessen (https://skillmd.com/u/lidessen) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lidessen/task-shaping --- # Task Shaping ## Principle expression **Primary:** P15 **Supporting:** P05, P08, P04 ## Scope Own one judgment: **relative to a conservative, economical reference execution profile, is the requested task a reliable primitive, a guarded task, a transformable unsupported form, or an escalation—and what is the smallest form that preserves the obligations needed to reconstruct and verify the whole?** The answer may be direct execution. Shaping is not synonymous with splitting. Use decomposition only when it removes a named source of instability and its results can be reconnected without making the reconciler repeat the original oversized task. The reference should normally be a Flash-class daily model rather than a frontier model: harness design should expose reliable productive structure at an economical, replaceable baseline. A stronger profile may be an escalation, but it does not silently redefine the baseline primitive set. This Skill prepares an execution shape and cultivates a conservative task- primitive map from admitted capability evidence. It does not manufacture a model claim, invent domain truth, launch child agents, choose concurrency or provider policy, accept results, or become a universal preflight. ## Principle source Use a host Sequence and matching interpretations when the host declares them. Otherwise use this package's read-only fallback in `references/sequence.md`. Read only P15, P05, P08, and P04. A live task may select a different current lead without changing this Skill's stable lineage. ## Start Ground the transformation before naming units: ```text Required whole outcome and acceptance owner: Global obligations and relations that no split may lose: Host reference profile, revision, and comparable evidence: Accepted error tolerance, failure consequence, and containment: Observed task shape, failure, variance, or overload: Available sources, tools, effects, and verification surfaces: Hard constraints, side effects, and escalation options: Observation that would show direct execution is sufficient: ``` If the task is coherent, locally verifiable, and supported by comparable runs, retain one unit. Do not manufacture packets, ledgers, or a Swarm. Apply this reference-loading gate; do not inventory or preload the reference directory: - stay in `SKILL.md` when direct versus guarded disposition is already clear from host evidence; - read [task primitives](references/task-primitives.md) only when the operation family or a possible compilation target is unclear; - read [concepts](references/concepts.md) only when capability envelope, task shape, local closure, differential analogy, or reconstruction are being conflated; and - read [transformation patterns](references/transformations.md) only after a concrete instability requires transformation detail. A direct task must not load transformation doctrine merely because it is available. ## Capability disposition Use one of four dispositions; do not collapse them into a confidence adjective: | Disposition | Meaning | Harness consequence | |---|---|---| | `reliable-primitive` | admitted repeated evidence meets an explicit task-specific error tolerance; residual failures are observable or contained | execute directly with normal result validation | | `guarded` | tolerance is unspecified, useful performance is plausible, or known material errors/variance remain | require independently evidenced checks, repetition, abstention, or human confirmation proportional to risk | | `transform` | the requested form is outside the envelope, but its obligations can be compiled into supported or guarded primitives and reconstructed | return the minimum transformation and coverage/reconstruction contracts | | `unsupported-escalate` | no evidenced direct form or truthful transformation keeps local work and reconstruction inside the envelope | gather capability evidence, use a stronger profile, or retain human execution/judgment | “Unsupported” is relative to the named reference profile and present harness; it is not a metaphysical claim that no model could perform the task. A proposed guard does not upgrade a task: it remains `guarded` or `transform` until matched runs show that the guard detects or contains the relevant error. Apply this fail-closed gate before using `reliable-primitive`: ```text Did comparable retained runs meet the stated tolerance in the same guarded form? yes -> reliable-primitive may be justified no, unknown, or the guard exists only in the proposed shape -> guarded ``` Do not subtract observed failures by imagining that a schema, abstention rule, verifier, or smaller batch would have caught them. Name the configured guarded variant separately and test it. For example, `47/50 extraction with three unsupported causal inferences` is `guarded` when accepted tolerance is zero, even if the proposed unit adds source spans and abstention. Only matched runs of that source-span-and-abstention variant can promote it. ## Core method 1. **Freeze the whole before transforming it.** State the required outcome, source revision, acceptance condition, hard constraints, authority, and load-bearing relations. A later unit may narrow its local question but may not silently narrow the original obligation. 2. **Fix the conservative reference profile and evidence.** Treat model, provider, harness, prompt and skill context, tools, permissions, loop policy, and completion contract as one profile. Use comparable retained observations when available. A token limit, model reputation, file count, or one successful run does not establish a stable envelope. Route an unevidenced reusable capability claim to `model-evaluation`. When the host has no reference, declare a provisional Flash-class baseline, mark the evidence status `discovery-needed`, and use `unsupported-escalate` as the execution disposition rather than filling the primitive map from intuition. Preserve the observation's outcome definition and denominator: `18/20 found every seeded material defect` is not `20/20 successful`, and protocol settlement is not semantic success. 3. **Classify requested operations, not only the task label.** Decompose the request just enough to identify its judgments and effects. Compare each with admitted profile evidence and `references/task-primitives.md`; distinguish a stable finite classification from an unstable open-ended completeness claim even when both are called “review.” Check observed errors against the named tolerance and consequence. A known error outside tolerance prevents `reliable-primitive`; a newly proposed containment mechanism does not erase that evidence. Classify the observed form as `guarded`, name the proposed guarded variant, and require matched evidence before promotion. Assign a capability disposition and preserve unknowns. 4. **Characterize the task shape.** Inspect semantic breadth, context volume, branching, cross-boundary coupling, horizon and external state, effect risk, local observability, output load, and completeness burden. Distinguish: protocol settlement, local semantic judgment, and system-level completeness; stability in one does not imply the others. 5. **Find the principal instability.** Identify the one mismatch between task shape and profile envelope whose removal changes the execution choice. Do not respond to every risk with more instructions or smaller chunks. 6. **Choose direct execution, a guard, or one minimum transformation.** Prefer the least machinery that addresses the named instability while retaining every global obligation. Common moves include bounding evidence, externalizing an obligation ledger, turning open completeness into an addressable queue, partitioning by semantically closed questions, separating prepare/verify/ commit, externalizing long-lived state, or escalating the executor. Select from `references/transformations.md`; do not apply the table as a pipeline. Choose an adaptive resolution: each unit must be small enough to remain inside the envelope and large enough to retain semantic closure. Smaller is not automatically more stable once boundary and coordination error dominate. Do not extrapolate a stable unit size beyond admitted evidence. An untested larger reducer, even over summaries, is a guarded discovery unit with a repartition or escalation signal—not an established primitive. 7. **Contract each execution unit—or stop at the owning boundary.** Give every unit one coherent contribution to the whole: ```text Whole-outcome contribution and unit question/action: Exact sources, revision, and bounded evidence: Incoming and outgoing relations: Deliberate overlap, if any: Local acceptance and disconfirming observation: Required result/status shape: Overload, repartition, or escalation signal: ``` A unit may report an external relation or `unverifiable`; it may not expand its authority merely to make the packet look complete. When a domain method owns the semantic partition, do not instantiate example file, family, claim, or subsystem units and then disclaim ownership in prose. Return only the observed envelope mismatch, partition requirements, global obligations, reconstruction requirements, and a handoff to that method. Prepared units remain absent until the domain owner forms them. 8. **Make coverage and reconstruction explicit.** Map every global obligation to one or more units. Define how results, disagreements, missing relations, and source revisions reconnect. Resolve conflicts against named sources, not votes. A local unit may report that evidence is absent from its supplied packet; it may not convert that local absence into a claim that the whole source or change lacks the evidence. The reconstruction owner must resolve that claim against the declared whole. Check integrability: common source state and boundary conditions, required execution order, cross-unit coupling, accumulated error, and global invariants must permit the local results to form one valid whole. If synthesis must reread all raw work or redo every local judgment, the proposed partition has only moved the unstable task downstream. 9. **Keep execution and acceptance external.** Return direct work or prepared units to the caller, domain method, current agent, Work Cell, Swarm, or other carrier. The carrier chooses release policy; a verifier and human or host owner settle semantic acceptance. This Skill never launches or approves its own shape. 10. **Observe and cultivate.** Compare completion, material errors, omissions, cross-unit repair, retries, usage, latency, and human repair with the unchanged form when claiming improvement. Reuse valid units and reshape only the failed region. Send reusable capability evidence to `model-evaluation`; update the task-primitive map only after admission and evidence from more than the development case. Retain failure boundaries and profile revision. Treat the observation as profile/task-shape evidence, not a universal decomposition law. ## Task Shape result Return the smallest form needed for execution or handoff: ```text Original outcome, source, hard constraints, and acceptance owner: Reference execution profile, revision, and evidence status: Capability disposition for each requested operation: Task-shape diagnosis and principal instability: Disposition: reliable-primitive | guarded | transform | unsupported-escalate Prepared units or direct execution boundary: Coverage and cross-unit relation map: Reconstruction, verification, and commit owners: Expected benefit, added coordination cost, and disconfirming observation: ``` Do not attach concurrency, provider, retry, or token allocations unless the caller has separately supplied those decisions. A Task Shape is semantic work preparation, not an orchestration manifest or budget. ## Boundaries and routing | Need | Owner | |---|---| | Establish or revise a reusable reference-profile capability claim | `model-evaluation` | | Select source information and delivery timing for one unit | `context-engineering` | | Define semantic packets already owned by code review, refactoring, cognition, or another domain | that domain Skill; use this Skill only for the envelope mismatch it does not own | | Choose the next practice after observing a result | `practice-cycle` | | Estimate necessary work or convert it into resources | `work-estimation` and the authorized resource owner | | Release prepared units, choose concurrency, retry, queue, or provider | orchestration runtime or caller | | Verify and accept semantic results | designated verifier and human or host authority | Do not force a domain method through generic vocabulary when it already produces stable, reconstructible units. Do not split by equal file counts, directory names, or token quotas when meaning crosses those cuts. Do not claim that more Cells, smaller contexts, structured output, or terminal success improves correctness without comparable behavioral evidence. Declaring the domain owner in a routing note does not cure an artifact that has already chosen that owner's semantic boundaries. ## Verification Read [evaluation](references/evaluation.md) when testing a new transformation, claiming improved stability, or promoting this Skill in a new task family. ## Completion standard A Task Shape is ready when it preserves the original outcome and authority, names the reference profile and evidence status, assigns a capability disposition, identifies the principal instability, justifies direct, guarded, transformed, or escalated work, keeps every unit locally coherent and verifiable, maps global obligations and cross-unit relations, makes reconstruction fit an evidenced envelope, and states an observation capable of defeating the shape. Without comparable execution evidence, report a proposed shape or discovery probe rather than a stability claim.