About autoplan
autoplan is an analysis-only planning skill for repository work. It grounds plans in repository evidence, adjusts depth to the request, and assigns cited risks and outcome-linked checks to plan steps.
Key features
- 01Request triage and scope selection
- 02Repository evidence gathering
- 03Risk lenses with cited findings
- 04File- and contract-grounded steps
- 05Milestones with outcome-linked checks
Use cases
- 01Planning a repository change before implementation
- 02Scoping migration or security-sensitive work
- 03Assessing a small change for a direct path
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ prathamdby/ skills --skill autoplanAfter installing, type/autoplanin Claude Code to confirm it loaded.
Bundled files
References and scripts installed along with SKILL.md.
REFERENCE.mdreferences/compatibility.mdcost.mddata.mdoperations.mdrouter-contracts.mdsecurity.mdDetails
From the project READMEAutoplan
Analysis only. Deliver one plan in thread, or at the caller's named path.
Options and triage
Default: current repo, standard depth, thread output. Quick/sketch uses Shape and Risks with always-on lenses. Thorough/risky/migration uses deep, winning over quick; a named area scopes evidence. Resolve conflicting scope narrowly and record both readings under Assumptions; never ask.
Before evidence, BLOCKED if no groundable verb/object/area, a non-repo intent
has no plannable target, or building rather than planning was requested.
Name the missing piece and one workable example; emit no lenses/map/milestones
on refusal. A build request does not authorize product edits by this skill.
Trivial means one module, no contract/data/signature change, and one-sentence
behavior. Return DIRECT plus that sentence and three lines: shape, files,
first step; no lenses. Otherwise record intent | scope | depth | terminal.
Done when triage has a terminal or a grounded target.
1. Evidence
Check whether the requested contract already exists; harden an existing match,
not rebuild it. Locate sibling recon's installed SKILL.md to resolve its
anchor and shared/legacy memory rules. Read a repository-verified snapshot
whose head resolves; do not invoke a refresh merely to read it. Without usable
memory, cold-scan at most three anchors for each of at most 30 manifest modules.
Missing recon is not an evidence dependency: use that bounded scan instead.
Scan ADR trees (docs/adr/, adr/, doc/adr/, docs/decisions/) silently
when absent. Note TODOs, recent reverts, and duplication. Dirty content is an
Assumptions overlay, not proof of existing contracts. Every later claim cites
a path/test/ADR, a scoped no-match result, or an explicit missing-evidence
Assumption; cite inherited patterns. Done when material claims have provenance.
2. Lenses
Apply all always-on probes in REFERENCE.md: contracts/API, failure/rollback,
longevity/debt. For a leaf, playbook, or matcher, also apply
references/router-contracts.md.
Score every probe for each activated lens, turning a hit into a cited risk owned by a plan step; close each lens pass with cite, risk → step, or no evidence. Deep activates all. Otherwise load only triggered lenses:
- Schema/backfill, stored artifacts, toolchain:
references/data.md. - External consumers/stored payloads:
references/compatibility.md. - Auth/secrets/PII/untrusted input:
references/security.md. - Hot paths/cost-bearing calls:
references/cost.md. - Services/deployment/paging:
references/operations.md. State one skip line per inactive lens. Done when every active probe is scored and each risk has an owner.
3. Plan
Specify decisions, changed contracts, constraints, and deviations from cited patterns, not implementation bodies or restatements. Standard/deep includes: Assumptions; Shape (bold sentence + pattern); Layer map (NEW/CHANGED roles); phase sequence; file-plus-symbol grounding (prose: file + contract step); worst-first Risks (probability × impact, or stated judgment); Milestones; non-goals; Handoff. Sketch includes only Shape and Risks.
Small work is one PR; otherwise each milestone is a reviewable vertical PR.
Begin with the thinnest end-to-end contract/check slice. Name landed symbols,
expected success and relevant failure outcomes, and assertions or observed
manual results proving them. Exit zero alone is not proof.
Handoff follows the caller's next route; default /peer-review, then ship.
When nothing will be built, such as a docs-only plan, use /peer-review or
stop. Preserve an explicit caller route.
Done with PLANNED when behavior is grounded or assumed, risks are owned,
and every milestone has outcome-linked checks.
Terminals: PLANNED, DIRECT, BLOCKED.