design-reconcile
Skill0 starsChanged 2 months agoMITView on GitHub
Hygiene check on a design document's decision ledger. Detects status inconsistencies, leftover superseded decisions, and unconfirmed discussion that leaked into the current-design section. Proposes moves (never deletions) and requires user approval. Delegated to a Haiku subagent. Works for any project — it reasons about the ledger, not about the code. Use when: 설계 문서 정리, 설계 확정 전 점검, decision ledger 정합성, design doc hygiene, 결정 로그 점검, 설계 위생 점검
In this repository13 other files it gives its agents.
- change-walkthroughSkill
- commit-msgSkill
- convention-auditSkill
- design-workflowSkill
- explain-design-htmlSkill
- explain-diff-htmlSkill
- explain-skills-workflowSkill
- impl-pipelineSkill
- review-deskSkill
- review-gateSkill
- session-boardSkill
- track-ledger-policySkill
- track-ledgerSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 5 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 26
Run it at two moments:
- Distinctive: line 32
Document editing is delegated to a Haiku subagent — it is a mechanical cut-and-paste job, and a cheap model keeps the lead's context free.
- Very specific: line 48
If task-file is omitted, list recent task documents and ask the user which one.
- Very specific: line 69
Every SUPERSEDED by D-n in the decision log points at a D-n that actually exists.
- Very specific: line 70
Every decision marked ACTIVE in the log is present in the current-design section.
- Very specific: line 71
Every decision present in the current-design section has a corresponding ACTIVE log entry.
- Distinctive: line 98
Present a diff and get approval before any edit: Every change, without exception.
- Distinctive: line 120
Read it once, for analysis only.
- Distinctive: line 136
If any finding is of type 2 (leftover superseded content), add the CLI-bypass warning.
- Very specific: line 138
If there are no findings: 설계 문서 정합성 이상 없음.
- Very specific: line 142
Use AskUserQuestion. The user may approve all, approve a subset, or reject.
- Very specific: line 153
MUST NOT: delete content, reword decisions, touch ## 작업 큐, or edit anything the user did not approve.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Design Reconcile
Checks that a task document's decision ledger is internally consistent, and that the current design section contains only confirmed decisions.
Run it at two moments:
- Before finalizing a design — right before
/impl-pipelinestarts. - After a deviation — when implementation revealed something the design did not anticipate and a decision was superseded.
Document editing is delegated to a Haiku subagent — it is a mechanical cut-and-paste job, and a cheap model keeps the lead's context free.
MUST invoke
Skill(design-workflow)before running this skill.
This skill is project-agnostic by construction: it reasons about the ledger document, never about the codebase. Nothing here needs to know what the project's layers or conventions are.
Arguments
/design-reconcile [task-file]If task-file is omitted, list recent task documents and ask the user which one.
Ledger Shape
The task document has four sections. This skill reasons about the first three.
| Section | Content | Managed by |
## 현재 설계 | Confirmed decisions only, inside <!-- CURRENT:START/END -->. Blocks: ### D-n · {topic-slug} · {title} | ledger decide (atomic replace) |
## 작업 큐 | Implementation slices | /impl-pipeline |
## 결정 로그 | Append-only. Each entry: ACTIVE or SUPERSEDED by D-n | ledger decide |
## 진행 로그 | Progress notes | ledger task-log / task-status |
Checks
1. Status consistency
- Every
SUPERSEDED by D-nin the decision log points at aD-nthat actually exists. - Every decision marked
ACTIVEin the log is present in the current-design section. - Every decision present in the current-design section has a corresponding
ACTIVElog entry.
Any mismatch in either direction is a finding.
2. Leftover superseded content
The current-design section still describes, in prose, a decision that has been superseded — typically because a topic-block replacement did not catch narrative text written outside the block.
3. Unconfirmed discussion
The current-design section contains non-final language: "~할지 고민", "TBD", "아마도", "maybe", "probably", "고민 중", "검토 필요", open questions, or two options presented without a chosen one.
This is the actual origin of document bloat. The current section must contain confirmed decisions only.
Handling Rules (non-negotiable)
These are guards. State them explicitly in the delegation prompt.
- Never delete. A finding is resolved by moving content, never by removing it.
- Leftover superseded prose → move to the decision log (as part of the superseded entry).
- Unconfirmed discussion → move to 진행 로그.
- Never rewrite an existing decision-log body. Log entries are append-only; only the
ACTIVE/SUPERSEDED by D-nstatus may flip. - Present a diff and get approval before any edit. Every change, without exception.
- Report CLI bypass.
ledger decideperforms an atomic topic-block replacement. Therefore leftover superseded content in the current section is evidence that someone edited the document body directly instead of using the CLI. Report that fact to the user explicitly — it is a process problem, not just a text problem.
Steps
Step 1 — Read the ledger
LEDGER="${CLAUDE_PLUGIN_ROOT}/scripts/ledger/ledger"
"$LEDGER" current --file <task> --ids-only # active decision IDs
"$LEDGER" current --file <task> # active decision bodies
"$LEDGER" raw-path --file <task> # 원문을 읽어야 할 때만Unlike /impl-pipeline, this skill may read the raw document — inconsistency detection requires
comparing the current section against the decision log, and the CLI only exposes the current view.
Read it once, for analysis only.
Step 2 — Delegate the analysis (Haiku)
Delegate to a Haiku subagent with:
- TASK — run checks 1–3 above against the task document.
- CONTEXT — the document path, plus the
current --ids-onlyandcurrentoutput. - OUTPUT — a findings list. For each: check number, exact location (section + heading + line), the offending text, and the proposed destination (decision log or 진행 로그).
- MUST NOT — do not edit the document; do not delete anything; do not reword decisions. Analysis only in this step.
Step 3 — Present findings
For each finding show: what, where (section + D-n + line), and the proposed move.
If any finding is of type 2 (leftover superseded content), add the CLI-bypass warning.
If there are no findings: 설계 문서 정합성 이상 없음. and stop.
Step 4 — Get approval
Use AskUserQuestion. The user may approve all, approve a subset, or reject.
Reject → stop, change nothing.
Step 5 — Apply (Haiku)
Delegate the approved moves to a Haiku subagent.
- Status flips (
ACTIVE→SUPERSEDED by D-n) and decision replacement go throughledger decidewhere possible — never by hand-editing the current section. - Content moves that the CLI does not cover are applied with
Edit, moving text to its destination section. Cut-and-paste, not delete. - MUST NOT: delete content, reword decisions, touch
## 작업 큐, or edit anything the user did not approve.
Step 6 — Report
List what moved and where. If a CLI bypass was detected, restate it — future edits must go through
ledger decide.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

