Unicorn estimation

Skill andrey-learning-machines/swe-harness/plugins/swe-harness/skills/unicorn-estimation

Portable SWE harness plugin for Codex and Claude Code

Install

npx -y skills add andrey-learning-machines/swe-harness --skill unicorn-estimation

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Guides the user through risk-based task estimation using decomposition, three-point estimates, and PERT formula. ALWAYS trigger on "estimate", "how long will this take", "time to complete", "sizing", "scope this", "effort estimate", "how many hours", "story points", "project timeline", "when will this be done", "cost estimate", "level of effort". Use when sizing any work item, feature, or project.

SKILL.md

4.7 KB, as published. Nobody here has run it

Estimation

Process

1. Decompose Exhaustively

Break tasks until each piece is < 8 hours and estimatable with confidence.

Task: "Add user authentication"
  ├─ Database schema (2h)
  ├─ API endpoints (16h) → decompose further
  ├─ Password handling (8h)
  ├─ JWT tokens (19h) → decompose further
  ├─ Middleware (14h) → decompose further
  ├─ Testing (23h) → decompose further
  └─ Integration (22h) → decompose further

Rule: If you can't estimate confidently, decompose further.

2. Identify Unknowns

unknowns:
  technical:  ["Never used bcrypt", "Don't know JWT refresh pattern"]
  domain:     ["Password rules unspecified", "SSO needed later?"]
  external:   ["DBA review has 2-day SLA"]
  resource:   ["Frontend dev availability uncertain"]

Every unknown adds risk. Quantify it.

3. Three-Point Estimate

For each atomic task:

Estimate	Meaning
Optimistic (O)	Everything goes perfectly
Realistic (R)	Normal conditions, typical hiccups
Pessimistic (P)	Murphy's Law applies

"Implement password hashing"
O: 1h  (copy existing pattern)
R: 3h  (read docs, test, handle edge cases)
P: 8h  (version issues, need alternatives)

Rule: If P > 3x R, decompose further.

4. Calculate PERT

Expected = (O + 4*R + P) / 6

Example: (1 + 4*3 + 8) / 6 = 3.5h

5. Apply Risk Buffers

Risk Level	Multiplier	When
Low	1.0-1.2x	Well-understood, clear requirements
Medium	1.2-1.5x	Some unknowns, moderate complexity
High	1.5-2.0x	New tech, unclear requirements
Critical	2.0-3.0x	Multiple unknowns, bleeding-edge

Risk categories:

Category	Multiplier	Triggers
Technical	1.3-2.0x	New lang/framework, complex algorithms, perf requirements
Domain	1.4-2.5x	Unclear requirements, regulatory, stakeholder disagreement
External	1.5-3.0x	Third-party deps, vendor timelines, cross-team
Resource	1.2-2.0x	Availability uncertain, skills gap, access issues

6. Add Integration Buffer

Integration is where 50% of bugs live. Always add 20-30%.

Subtasks total: 40h
Integration buffer (25%): 10h
Final: 50h

Anti-Patterns

Anti-Pattern	Problem	Fix
"2 Hours"	"I have no idea"	Use three-point + PERT
Secret Padding	Destroys trust	State explicit buffer: "5h + 5h buffer"
Ignoring Integration	Underestimate 30-50%	Always add integration buffer
Forgetting Testing	"Done" != done	Include test time (often 1:1 with dev)
No Confidence Level	False precision	Use ranges: "45h +/- 5h"

Communication Template

[ESTIMATE] (+/-[UNCERTAINTY]) assuming [ASSUMPTIONS]

Breakdown:
- Component 1: Xh
- Component 2: Yh
- Integration: Zh

Confidence: [High/Medium/Low] (XX%)

Assumptions:
1. [Critical assumption]

Risks:
- [Risk]: [impact] / [mitigation]

Dependencies:
- [External dependency]

Unknowns:
- [Unknown] - [how to resolve]

Validation Checklist

Before finalizing any estimate:

When to Re-Estimate

Requirements change -- re-decompose affected components
Unknown resolved -- update with actual findings
Overrun > 20% -- stop and reassess remaining work
Dependency changes -- adjust timeline
Team changes -- different skills = different estimates

Script

Run the interactive PERT estimation helper:

skills/estimation/scripts/estimate.sh
# or
./scripts/estimate.sh [--output filename]

References

references/decomposition-examples.md -- full worked examples
references/risk-analysis.md -- risk assessment framework
references/communication-guide.md -- stakeholder communication patterns
scripts/estimate.sh -- interactive PERT estimation script

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.