Revises existing implementation plans with focused surgical edits, maintaining consistency. Use when a plan needs changes before or during execution.
How this skill is triggered — by the user, by Claude, or both
Slash command
/ai-research-workflows:iterating-plansThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Revise an existing `docs/rse/specs/plan-*.md` with focused, surgical edits that
Revise an existing docs/rse/specs/plan-*.md with focused, surgical edits that
preserve good content and maintain internal consistency.
This skill leans Collaborative by default. For the full Collaborative-vs-Direct protocol and override rules, see the Interaction Modes reference in the ai-research-workflows:using-research-workflows skill.
If the plan file or the requested change is missing, enter Collaborative mode
and ask for it (list recent plans with ls -lt docs/rse/specs/plan-*.md .agents/plan-*.md 2>/dev/null | head -5).
Read the entire plan completely (avoid partial reads). Identify:
Parse the requested change: adding/removing phases or tasks, updating success criteria, changing technical approach, incorporating new information, splitting or reordering phases.
Determine whether the change requires codebase research:
Only if changes require new technical understanding:
Before making any changes, present:
I've read the plan and understand you want to [summary of requested changes].
Current state:
- Phases: [list]
- Scope: [summary]
Proposed changes:
- [Change 1]
- [Change 2]
Sections affected:
- [Section 1]
- [Section 2]
[If research was done] Key findings:
- [Finding]
Does this match what you're looking for?
Get user confirmation. If the user says no, clarify and adjust; only proceed on yes.
Make focused, precise edits to the plan in place (at its existing path — docs/rse/specs/plan-*.md, or a legacy .agents/plan-*.md).
Good edits: add a new phase section between existing phases; update specific success criteria items; modify task descriptions within a phase; add file references to existing tasks.
Bad edits: replace the entire plan; rewrite sections that don't need changes; change formatting or structure unnecessarily.
Maintain existing structure unless explicitly changing it: same phase format, same task description pattern, same reference style, same level of detail.
When adding tasks or phases include specific path/to/file.ext:123-145
references — research actual code locations if needed.
When scope changes update success criteria (add/remove items) and preserve the Automated vs. Manual verification split:
### Automated Verification
[Commands and checks that run without human intervention]
### Manual Verification
[Steps requiring human testing and judgment]
Ensure consistency: new phases follow existing structure; scope changes update "What We're NOT Doing"; approach changes update "Implementation Approach"; new tasks reference actual codebase examples.
After editing, run a consistency scan: phase numbering is monotonic and contiguous; every cross-reference (to a phase, file, or success criterion) still resolves; no task references a type, function, or file removed elsewhere; and the Automated/Manual split is intact. Fix any drift before presenting.
When a change alters the technical approach or adds a results-producing phase,
re-check the plan's research criteria too — reproducibility, seeds/provenance,
and numerical-correctness success criteria (ai-research-workflows:ensuring-reproducibility,
ai-research-workflows:hardening-research-code) — so an edit doesn't silently
invalidate them.
# Plan Updated
Changes to `docs/rse/specs/plan-[slug].md`:
## Changes Made
- [Change 1 — specific section]
- [Change 2 — specific section]
## Impact
- [How changes affect implementation]
- [New/removed success criteria]
Would you like any further adjustments?
If the user wants more changes, re-read the plan first and apply the same process.
Be skeptical — push back on changes that conflict with existing phases; question vague feedback ("when you say 'more robust', do you mean...?"); verify technical feasibility with code research before adding tasks.
Be surgical — edit only what needs changing; don't reformat or rewrite for style.
Be thorough — read the entire plan before editing; never guess at file paths.
No open questions — this is a BLOCKING REQUIREMENT. If a requested change raises questions, ask or research immediately. Do NOT update the plan with unresolved questions; plans must be complete specifications.
For the five detailed patterns (adding a phase, updating success criteria,
adjusting scope, incorporating experiment results, splitting a complex phase),
see references/iteration-patterns.md.
After iterating, execute with the ai-research-workflows:implementing-plans skill. Incorporate
docs/rse/specs/experiment-*.md (or legacy .agents/experiment-*.md) results when revising the technical approach.
Before completing, verify:
npx claudepluginhub uw-ssec/rse-plugins --plugin ai-research-workflowsCreate detailed execution-ready implementation plans for complex or high-risk changes without coding. Useful for refactors, migrations, resumable phase checklists, and hand-offs to execute-plan.
Replicates Claude Code plan mode in read-only fashion: explores codebase, designs changes, gets user approval, then writes plan to .claude/plans/<slug>.md without executing. Triggers on 'plan only' phrases.
Designs multi-file implementation plans grounded in codebase research. Invoked for any feature, refactor, or change needing structured design before coding.