| name | feed-planning-context |
| description | Prepare relevant planning artifacts, including optional statechart and Dumplink context, for agent implementation work without overloading context. |
Feed Planning Context
Use this skill when the user wants to give an agent a planning artifact and have it act on the artifact without losing the shaped intent.
This skill prepares context only. It does not implement code.
Goal
Create a compact context packet from the provided planning artifacts so an agent can understand the current task, artifact authority, selected direction, slice boundary, relevant statechart rows, executable examples, interface contracts, Dumplink task group, execution contract, and verification target when each is present.
Inputs
Use whichever artifacts the user provides or points to:
- framing doc
- shaping doc
- appetite card, when appetite is recorded separately
- selected shape
- fit check
- reverse fit check
- breadboard
- statechart for the selected scope, when present
- executable breadboard
- interface contract sketch or plain-language interface contract
- Dumplink plan, when present
- slice plan
- kickoff doc
- breadboard reflection
- raw notes or transcripts, only when needed
Default behavior
- Identify the current task.
- Identify the source artifacts and their authority.
- Extract only the context needed for the next move.
- Preserve stable IDs for requirements, places, affordances, stores, contracts, example runs, edge cases, and slices.
- Preserve executable breadboard fixtures, example runs, expected outputs, state changes, side effects, edge cases, and acceptance tests when present.
- When a sketch-reconciliation record exists, include accepted changes only after they have been applied to the authoritative artifacts. Exclude pending, rejected, and ambiguous visual deltas from build instructions.
- Preserve field names, required/optional distinctions, enum values, nullability, and error cases when interface contracts are present.
- Preserve relevant statechart states and transitions with their source breadboard IDs when present; the breadboard remains authoritative.
- Preserve the active Dumplink task group, dependencies, risk state, cuts, and acceptance checks when present.
- Keep rejected alternatives and raw notes out of the active packet unless the user asks for discovery or reconstruction.
- Preserve the accepted appetite and cut line.
- Name explicit non-goals and exclusions.
- Add an execution contract.
- Add a verification target.
- Stop after preparing the context packet.
Output format
# Context Packet
## Task
...
## Source artifacts
- ...
## Authority order
1. ...
2. ...
3. ...
## Use these sections first
- ...
## Do not use unless needed
- ...
## Must preserve
- ...
- Accepted appetite and cut line: ...
## Selected requirements
- ...
## Relevant places / affordances / stores
- ...
## Relevant statechart
- Selected scope:
- States and transitions:
- Source breadboard IDs:
- Explicit gaps:
## Relevant executable breadboard
- Selected slice:
- Example starting data / fixtures:
- Example runs:
- Expected user-visible results:
- Expected state changes:
- Expected side effects:
- Edge cases:
- Acceptance tests:
- Open decisions:
## Relevant interface contracts
- Contract:
- Boundary:
- Input shape:
- Output shape:
- Branches / errors:
- Open decisions:
## Relevant Dumplink plan
- Active task group:
- Relevant tasks:
- Risk state:
- Dependencies:
- Cuttable scope:
- Acceptance checks:
- Stop condition:
## Current slice
- Slice: ...
- Demo: ...
- Produces: ...
- Exclusions: ...
## Execution contract
- Goal condition:
- Required checks:
- Allowed files / areas:
- Out-of-scope changes:
- Return-to-planning conditions:
- Checkpoint cadence:
- Verification caveats:
## Open questions
- ...
## Build-handoff behavior
1. Restate the relevant constraints.
2. Identify implementation implications.
3. Ask at most 3 blocking questions.
4. Propose a plan before editing code.
5. If implementation reality changes the plan, propose a planning update instead of silently drifting.
6. Flag missing field names, nullability, enum values, error cases, fixtures, expected outputs, or acceptance tests instead of inventing them.
7. Work toward the goal condition, run the required checks, and report incomplete verification directly.
## Verification target
- ...
Authority order
When artifacts disagree, use this default authority order unless the user says otherwise:
- the user's latest explicit instruction
- selected slice or kickoff doc
- executable breadboard, when present
- selected interface contract, for boundary-level input/output details
- selected Dumplink task group and sequence, for task-group scope and build order
- selected breadboard
- selected shaping direction
- framing doc
- raw notes and transcripts
- rejected alternatives and brainstorming
A statechart is derived from the selected breadboard and never outranks it.
Chunking rules
When artifacts are large:
- Start with the Context Card or top summary.
- Include only the artifact needed for the current phase.
- Prefer selected sections over whole documents.
- Keep raw notes out unless the task is discovery or reconstruction.
- Use stable IDs so the agent can request missing sections by name.
- If the artifact is long, summarize it into an implementation packet before coding.
- Keep rejected alternatives available, but clearly mark them as rejected.
- Keep non-goals close to the task so the agent does not expand scope.
Stable ID handling
For new artifacts, use the compact conventions in AGENTS.md: R0, P1, U1, N1, S1, ST1, TR1, C1, RUN1, E1, TG1, and V1.
Preserve any established legacy IDs such as REQ-01 or SLICE-01 rather than renaming them for style.
Do not rename IDs for style. If meaning changes, mark it as a planning update.
Executable breadboard handling
When an executable breadboard is present, preserve:
- selected slice ID and boundary
- relevant places, affordances, stores, and wires
- fixtures or example starting data
- example runs
- expected user-visible results
- expected state changes
- expected side effects
- edge cases
- acceptance tests
- open decisions
The executable breadboard is the build-handoff artifact for a selected slice. It is more authoritative than the ordinary breadboard for the examples and expected results of that slice.
Interface contract handling
When interface contracts are present, preserve:
- contract IDs
- boundary names
- field names
- required vs optional distinctions
- enum values
- nullability
- units, such as cents instead of dollars
- branches and error cases
- open decisions
If a field-level decision is missing, flag it. Do not invent it during context packaging.
Statechart handling
When a statechart is present, include only the states and transitions relevant to the selected slice. Preserve their source breadboard IDs and explicit gaps. If the statechart and breadboard disagree, use the breadboard and request that the statechart be regenerated.
Dumplink handling
When a Dumplink plan is present, preserve the active task group, vertical boundary, risk state, dependency order, cuttable scope, acceptance checks, and stop condition. Do not activate deferred groups or flatten the plan into a horizontal backlog.
Execution contract handling
The execution contract turns the context packet into a loop-ready handoff. It should name:
- the concrete goal condition that would make the work complete
- the checks that should be run before completion
- the files, folders, or areas the agent may touch
- the changes that are outside the selected slice
- the conditions that should return work to planning
- the checkpoint cadence for multi-step work
- the caveats to report when verification is incomplete
Drift handling
If implementation reality conflicts with a planning artifact, return:
## Planning drift found
The selected artifact says:
- ...
The implementation reality is:
- ...
Options:
1. Update the code to match the artifact.
2. Update the artifact because the original assumption was wrong.
3. Split the slice and defer the conflicting part.
Recommended move:
- ...
Do not
- implement code
- turn rejected alternatives into active requirements
- paste entire transcripts into the active context unless needed
- silently drop non-goals
- silently change the selected slice boundary
- silently change specified executable breadboard examples or interface contracts
- invent missing field names, enum values, nullability, error cases, fixtures, expected outputs, or acceptance tests
- rename stable IDs during packaging
- treat raw notes as more authoritative than selected artifacts